Every charity uses AI now and almost none are ready

Ninety-two percent of nonprofits now use artificial intelligence in some form, but only 7% say it has produced a major improvement in what their organization can accomplish. Those two figures, from a December 2025 benchmark survey of 346 organizations published by Virtuous and Fundraising.AI in February 2026, define the current moment better than any slogan about “AI for good.” Adoption is no longer the story. The story is the gap between having the tools and getting real results from them, and that gap is where most of the sector currently sits.

The state of AI adoption across the nonprofit world in 2026

The pattern repeats across regions. In the United Kingdom, the Charity Digital Skills Report 2025, based on responses from 672 charities, found that 76% of charities were using AI tools, up from 61% in 2024 and roughly 35% in 2023. Adoption more than doubled in two years. Yet the same report shows the number of charities with a digital strategy actually fell, from 50% to 44%, and around a third of respondents rated their chief executive’s AI knowledge as poor. Usage is racing ahead of understanding, and understanding is racing ahead of governance.

The Virtuous data gives the sharpest picture of how AI is actually used day to day. Sixty-five percent of surveyed organizations describe their AI use as reactive and individual: a single staff member pasting a draft appeal into a chatbot, a program officer summarizing a report, one-off prompts with no shared method behind them. Only 18% report operational use across team workflows, and just 7% say AI is embedded into goals, budgets, and performance indicators. Eighty-one percent of organizations use AI individually and ad hoc, while only 4% have documented, repeatable workflows. Nearly half, 47%, have no AI governance policy at all.

The report’s authors call the result an “efficiency plateau.” Nearly four in five organizations report small to moderate improvements: faster first drafts, quicker research, better-polished content. Those gains are real and, for chronically understaffed teams, welcome. But they are incremental, and they rarely show up in the numbers that matter most to a charity, such as funds raised, people served, or outcomes achieved. Gabe Cooper, chief executive of Virtuous, summed up the dominant pattern as one person using a chatbot to draft an appeal while the rest of the team stays buried in manual processes, which he described as a workaround rather than a strategy.

Money explains part of the gap. In the UK survey, 69% of charities named squeezed organizational finances as the single biggest barrier to digital progress, and 64% pointed to the difficulty of funding infrastructure and tools. Larger nonprofits with budgets above $1 million adopt AI at nearly twice the rate of smaller ones, 66% versus 34% by one sector count, a divide that reflects access to staff time and money rather than any difference in ambition. Small organizations are not skeptical of AI; they are stretched too thin to build systems around it.

The timing of this plateau matters because the external environment has turned harsher. Government funding for humanitarian and development work shrank sharply through 2025 and 2026, demand for charitable services keeps rising, and donor bases in many countries are aging and narrowing. Leaders inside large humanitarian organizations have started framing AI bluntly as one of the few tailwinds available to a sector facing mostly headwinds. The International Rescue Committee, with roughly 12,000 staff across 40 countries, has said publicly that it is trying to use AI as much as possible precisely because budgets are falling while needs are not.

That is the honest context for everything that follows in this analysis. AI in charity is neither a miracle nor a fad. It is a general-purpose technology arriving in a sector that is under-resourced, data-rich in some places and data-poor in others, bound by exceptional duties of care toward vulnerable people, and dependent on public trust in a way no commercial industry is. The organizations getting real value from it, the 7%, share identifiable habits: they redesigned workflows instead of layering a chatbot on top of old ones, they wrote rules before scaling usage, and they measured results against mission outcomes rather than vibes. The organizations getting hurt by it, quietly and often invisibly, share habits too: unreviewed outputs sent to funders, sensitive beneficiary data pasted into consumer tools, and automated messages that erode the human relationships giving depends on.

This article maps both sides in detail: the strongest evidence-backed uses of AI across fundraising, programs, humanitarian aid, conservation, health, and education; the money and programs flowing from big technology companies and foundations; the honest costs and risks; the regulation now landing, especially in Europe; and a practical adoption path for organizations that want to move from experiments to results.

Key terms explained, from generative models to agents and predictive scoring

Precision about terms matters, because “AI” now covers technologies with very different risk profiles and very different uses in charitable work.

Artificial intelligence is the umbrella term for software systems that perform tasks which previously required human judgment: recognizing patterns, generating text, classifying images, predicting outcomes. In the nonprofit context, almost everything currently labeled AI falls into one of four families: generative models, predictive models, computer vision, and agents.

Generative AI produces new content, including text, images, audio, and code, from a prompt. The large language models behind Claude, ChatGPT, and Gemini were trained on enormous volumes of text to predict likely word sequences, which is why they draft grant narratives, donor emails, and board summaries fluently. Their core limitation follows from the same mechanism: they produce plausible text, not verified truth, and they will sometimes state false information with full confidence. That failure mode, usually called hallucination, is the single most consequential technical fact a nonprofit leader needs to internalize, because a fabricated statistic in a funder report damages credibility in a way a slow report never could.

Predictive AI, often just called machine learning, learns statistical relationships from historical data to estimate something about new cases: which lapsed donor is most likely to give again, which household is most likely to be living below a poverty line, which patch of forest is most likely to see poaching next month. Predictive systems have run inside fundraising software and humanitarian research for over a decade, long before chatbots made AI a household topic. Their key limitation is inherited bias: a model trained on skewed history reproduces that history.

Computer vision applies machine learning to images and video. In the charitable world it reads satellite photos to map building damage after earthquakes, identifies species in millions of camera-trap images, and screens medical photographs collected by community health workers.

AI agents are the newest layer: systems that do not just answer a question but carry out multi-step tasks, calling software tools, querying databases, and chaining actions toward a goal. The plumbing that makes this practical for nonprofits is the Model Context Protocol (MCP), an open standard that lets an AI assistant connect directly to platforms like a Blackbaud donor database or Candid’s grants data and work with live records rather than pasted text. Agentic systems promise the largest productivity gains and carry the largest governance demands, because an agent acting on real systems can make real mistakes at machine speed.

A few adjacent terms round out the vocabulary. Retrieval-augmented generation (RAG) grounds a language model’s answers in an organization’s own documents, sharply reducing fabrication for internal knowledge tasks. Fine-tuning adapts a general model to a specialized domain, such as maternal health questions in Swahili. Propensity scoring is the fundraising industry’s term for predictive models that rank donors by likelihood to give, upgrade, or lapse. Natural language processing (NLP) is the older umbrella term for machines working with human language, still common in humanitarian and health projects. And AI governance covers the policies, review steps, and accountability structures that determine who may use which tools, on what data, with what human oversight.

One distinction deserves special emphasis for charities: the difference between assistive and autonomous use. Assistive use keeps a human decision-maker in the loop; the model drafts, the person judges and sends. Autonomous use lets the system act or decide on its own: publishing content, answering a beneficiary, scoring an application. Nearly all of the sector’s success stories to date are assistive or narrowly autonomous with heavy guardrails, and nearly all of its public failures happened when systems were allowed to act autonomously in sensitive contexts without adequate review. The rest of this article uses that lens repeatedly, because it predicts outcomes better than any vendor claim.

A short history of technology in charitable work

Charities have always adopted the tools of their era a step behind business, then bent them toward purposes business never imagined. Direct mail built the modern fundraising industry in the mid-twentieth century, and the databases that managed those mailing lists were, for many charities, their first serious software. By the 1990s, dedicated donor management systems such as Blackbaud’s Raiser’s Edge turned fundraising from card files into structured data, which quietly laid the foundation for everything machine learning would later do, because predictive models are only as good as the records beneath them.

The 2000s brought online giving and, with it, the first behavioral data at scale: open rates, click paths, abandoned donation forms. The 2010 Haiti earthquake proved text-message giving could move tens of millions of dollars in days and taught the sector that channel innovation changes donor behavior, not just processing costs. Around the same time, a first wave of data science for good took institutional form. DataKind, founded in 2011, began pairing volunteer data scientists with nonprofits. Academic groups started publishing on machine learning for poverty measurement, patrol planning against poachers, and disease surveillance. The United Nations’ ITU launched its AI for Good summit in 2017, giving the field a recurring global stage.

Predictive fundraising arrived commercially in the same decade. Wealth screening and propensity scoring, once the preserve of the largest institutions, spread through mid-market donor platforms. By the late 2010s a mid-sized charity could buy models that ranked its file by likelihood to give, lapse, or leave a bequest, even if most bought the scores and never changed their workflows around them, a pattern that would repeat.

The rupture came in November 2022 with the public release of ChatGPT. For the first time, frontier AI required no data team, no budget line, and no procurement process; any staff member with a browser could use it. Adoption in the charity sector was correspondingly bottom-up and largely invisible to leadership. UK survey data captures the speed: roughly 35% of charities used AI tools in 2023, 61% in 2024, 76% in 2025. In the US, the figure reached 92% by late 2025. No prior workplace technology, not email, not the web, not smartphones, moved through the sector this fast.

Institutions then scrambled to catch up with their own staff. Google.org launched a Generative AI Accelerator in 2024, backing 21 nonprofits with more than $20 million, and followed with a second $30 million cohort in 2025. OpenAI introduced nonprofit discounts in 2024, convened over a thousand nonprofit leaders at a national “Nonprofit Jam” in 2025, and committed $50 million to a People-First AI Fund that paid out $40.5 million across 208 US nonprofits by December 2025. Anthropic launched Claude for Nonprofits on Giving Tuesday 2025 with discounts of up to 75% and direct connectors into nonprofit data platforms. Foundations moved too, led by the Patrick J. McGovern Foundation, which has now directed roughly half a billion dollars toward public-purpose AI and data work.

Seen against this arc, 2026 is not the beginning of AI in charity. It is the moment three older threads, decades of donor data, a decade of applied machine learning research, and three years of grassroots generative AI use, are being braided into something organizations must now actually manage.

The mechanics behind the tools nonprofit teams actually use

Nonprofit leaders do not need to build models, but the ones making good decisions understand, at least in outline, what is happening inside the tools they buy. The mechanics explain both the gains and the failure modes.

A large language model is trained in two broad phases. First, it learns from a vast corpus of text to predict the next token, a word fragment, given everything before it. This is where fluency, world knowledge, and reasoning-like behavior come from, and also where the knowledge cutoff comes from: the model knows what its training data contained, frozen at a point in time. Second, it is refined with human feedback to follow instructions, decline harmful requests, and write in a helpful register. Nothing in either phase involves checking claims against a database of facts. The model generates the most statistically plausible continuation, which is usually accurate for well-documented topics and unreliable for specific figures, niche entities, and anything after its training cutoff. Modern deployments compensate by bolting on retrieval: web search, document search across an organization’s files, or structured queries into live systems. When a charity’s AI assistant reads the actual donor record before drafting the stewardship note, fabrication risk drops from a central problem to an occasional one.

Predictive models work differently. A team assembles historical examples, features describing each case and a label recording what happened, and an algorithm learns the mapping. A donor propensity model might use recency, frequency, gift sizes, event attendance, and email behavior as features, with “gave again within twelve months” as the label. A poverty-targeting model in a humanitarian program might use mobile phone metadata, call patterns, top-up amounts, data usage, with survey-measured consumption as the label. The output is a score, and the honest way to read a score is as a probability estimate with error bars, not a verdict. Every serious deployment therefore reports exclusion and inclusion errors: eligible people the model missed, and ineligible people it selected. Those error rates, not accuracy headlines, are what determine whether an allocation system is defensible.

Computer vision models learn from labeled images the way predictive models learn from labeled rows. The xBD research dataset that underpins much disaster-mapping work, for example, contains over 850,000 annotated buildings across more than 45,000 square kilometers of pre- and post-disaster satellite imagery, letting models learn what “destroyed,” “major damage,” and “intact” look like from above across many disaster types.

Agents combine a language model with tool access and a loop: read the goal, plan, call a tool, observe the result, continue. Connected through MCP to a CRM, a spreadsheet, and email, an agent can execute a task like “pull all lapsed donors above $500, draft personalized re-engagement notes referencing their last gift, and queue them for my review.” Every step it takes is an action on real systems, which is why the practical discipline around agents is scoping: least-privilege access, read-only where possible, human sign-off before anything leaves the building.

Understanding these mechanics collapses most vendor mystique into three plain questions a charity can ask of any AI product. What data was this trained or grounded on, and does it include ours? Where does human review sit in the loop? And what are the measured error rates on cases like ours, not on a benchmark? Organizations that ask those questions consistently end up in the successful minority; organizations that do not tend to discover the answers later, in public.

Fundraising as the first frontier for machine intelligence

Fundraising was always going to be the entry point. It is the part of a charity that most resembles a commercial function, it runs on structured data, and its output is measured in currency, which makes gains visible. The 2026 Virtuous benchmark confirms it: content generation, donor communications, and donor research are the most common AI uses in the sector, and fundraisers were among the earliest and heaviest adopters of generative tools.

The applications sort into three tiers of ambition. The first tier is production work: drafting appeals, thank-you letters, event copy, social posts, and the endless variants each channel demands. This is where nearly everyone starts and where the “faster drafts” gains of the efficiency plateau live. The craft here is not prompting so much as editing; strong teams treat model output as a first draft from a talented intern who has never met a single donor, and they rewrite for voice, accuracy, and specificity before anything ships. Weak teams ship the draft, and donors notice, because generic gratitude reads as no gratitude at all.

The second tier is intelligence work: research and analysis that previously did not happen because nobody had time. Prospect research on a foundation before a meeting. Summarizing five years of a donor’s correspondence before a stewardship call. Mining event feedback for themes. Analyzing which appeal segments actually performed. Fundraisers report far more comfort here than with outbound communication; in one 2025 sector survey, 82% of fundraisers were comfortable using AI for donor research while 63% remained unsure about using generative AI for donor communications, an instinct that maps neatly onto the assistive-versus-autonomous distinction.

The third tier is predictive and structural: propensity models that rank a file by likelihood to give, upgrade, or lapse; churn alerts that trigger human outreach before a monthly donor cancels; smart donation forms that adjust suggested amounts. Adoption here remains thin, around 13% of nonprofits use predictive AI for donor prospecting by one 2025 count, but the results where it is used well are the most concrete in the whole fundraising conversation. Payment platform data shows AI-assisted donation forms producing an average one-time gift of $161 against an industry average of $115, and average monthly recurring gifts of $32 against $24. First-time donor retention across the sector has been stuck between 20% and 30% for years; predictive prioritization exists precisely to concentrate scarce human attention on the relationships most at risk and most promising.

Selected adoption and results figures for AI in nonprofit fundraising

The table condenses the paradox this article keeps returning to: near-universal usage, thin structural adoption, and strong results clustered exactly where organizations moved beyond drafting text into data-grounded, workflow-level applications.

What separates fundraising teams that convert AI into revenue from those that convert it into slightly faster busywork? The benchmark evidence points to three habits. They connect the AI to their actual donor data instead of working from memory and pasted snippets, which is what the new generation of CRM-native tools and MCP connectors enables. They redesign a whole workflow, for example, lapsed-donor recovery from scoring through drafting through human call, rather than automating one step of an unchanged process. And they measure against fundraising outcomes, retention, average gift, portfolio coverage, not hours saved. Hours saved is a fine start, but hours saved evaporates unless it is deliberately reinvested in the work only humans do in fundraising, which is building relationships.

One caution belongs in any honest account of AI fundraising: the technology raises the volume of solicitation everywhere at once. When every organization can produce ten times the content, donors do not read ten times more email; they raise their filters. The arms race dynamic means production-tier gains decay over time, while intelligence-tier and relationship-tier gains compound. Charities planning multi-year strategy should weight their AI investment accordingly.

Grant writing and grant seeking in the age of language models

No fundraising task has been reshaped faster than grant work, because no task combined so much writing with so much repetition. A development officer assembling twenty grant reports a quarter spends most of those hours reformatting the same program facts for different funder templates. That is exactly the shape of work language models absorb well, and the sector noticed immediately: the State of AI in Nonprofits research from TechSoup and Tapp Network found 60% of nonprofit professionals strongly interested in AI for grant writing and fundraising, with about a quarter already using it specifically for grant writing as early as 2025.

Used well, the workflow looks like this. The organization maintains a source-of-truth library: mission language, program descriptions, outcome data, budgets, staff bios, prior successful proposals. The model drafts against that library, converting a program report into a letter of inquiry, adapting a core case statement to a specific funder’s priorities and word limits, generating first-pass answers to standard application questions. Humans then do what only humans can: verify every figure, inject the specific relationship context, and decide what this funder actually needs to hear. Teams working this way consistently report cutting proposal drafting time by half or more while raising baseline quality, because the model never forgets to address a required criterion and never submits the wrong funder’s name from a stale template, two errors human grant writers know intimately.

AI is also changing the search side of grant seeking. Matching tools scan funding databases against a charity’s profile, activities, and outcomes, surfacing well-aligned opportunities a stretched team would never have found. In a UK grants market where average application success rates hover around 35% and some funders report application volumes surging by 50% or more, anything that improves targeting, applying to fewer, better-fit funders rather than more, has direct income effects and reduces wasted effort on both sides of the relationship.

The funder side of the table is where the situation gets genuinely contested. Foundations now receive visibly AI-drafted applications at scale, and their reactions span the full spectrum: one sector survey found 23% of foundations saying they will not accept applications with generative AI content, 10% explicitly accepting it, and 67% undecided. In research funding, the UK’s NIHR, Wellcome, and fellow funders drew a clean line in a 2024 joint statement: generative AI must not be used in peer review of applications, while applicants may use it in preparation with acknowledgment. The Institute for Voluntary Action Research and others have documented a deeper worry beneath the policies, which is that AI-polished applications can flatten the signal funders rely on. When every application reads fluently, fluency stops indicating capacity, and funders start looking harder at track records, references, and site visits instead.

That dynamic points to the strategic reading for charities. The near-term advantage of AI grant writing is real but self-eroding, because it is available to every applicant. The durable advantage goes to organizations that use the reclaimed hours to strengthen what AI cannot fake: measured outcomes, financial transparency, and genuine funder relationships. It also argues for disclosure. A short line acknowledging that drafting was AI-assisted with human verification costs nothing with the majority of funders today and protects trust with the significant minority for whom silent AI use, once discovered, reads as deception.

Two operational rules keep grant AI on the safe side of the line. First, the model never invents evidence; every statistic, beneficiary count, and outcome claim must trace to a verifiable internal source, because a hallucinated number in a funded proposal is not embarrassment, it is misreporting to a funder. Second, confidential funder documents and unpublished program data belong only in tools whose terms guarantee the data is not used for model training, a distinction between consumer and enterprise AI plans that grant teams learn quickly or regret slowly.

Donor communication, personalization, and the authenticity problem

Personalized communication has always outperformed generic appeals; the constraint was never knowledge, it was labor. A major-gifts officer with a portfolio of 150 relationships can personalize. An email program serving 80,000 supporters cannot, or could not, until models made it cheap to reference a donor’s giving history, the program they funded, and the campaign that first brought them in, at unlimited scale. This is the most seductive promise AI makes to fundraising, and the one that requires the most care, because it sits directly on top of the sector’s most fragile asset.

The evidence on donor sentiment is more nuanced than either enthusiasts or skeptics claim. The Donor Perceptions of AI study, which surveyed 1,031 recent donors in August 2025, found conditional acceptance rather than rejection: 43% of donors said nonprofit AI use would have a positive or neutral effect on their giving, 32% said it would make them less likely to give, 14% more likely, and a quarter said it depends entirely on how AI is implemented. Donors drew sharp, consistent lines about where AI belongs. Back-office applications, impact measurement, financial controls, data analysis, cleared majorities as appropriate. Front-line applications that touch the donor relationship itself evoked ambivalence, and anything suggesting a machine was performing the gratitude or the relationship provoked the strongest resistance. The earlier 2024 wave of the same research found 39.8% of donors uncomfortable with how personalization might use their data, softening to acceptance for 28.1% when organizations are transparent about their practices.

Transparency, in other words, has shifted from a nice gesture to the operative variable. Yet disclosure remains rare: by one 2025 estimate, only about 15% of charitable organizations disclose their use of generative AI. That gap between practice and disclosure is a liability accumulating quietly across the sector. Donors who discover undisclosed AI use tend to reframe past communications retroactively; the warm note they valued becomes, in memory, a machine output, and the organization’s other claims inherit the doubt. The Give.org Donor Trust Report found a majority of donors, 54.5%, would be discouraged from giving if they knew an AI-generated image in an appeal had not been verified for accuracy by a staff member, a finding that generalizes: what donors are rejecting is not technology, it is the absence of human accountability behind it.

The practical synthesis working teams have converged on has three parts. Use AI heavily on the inside of communication work: research, segmentation, first drafts, subject-line variants, accessibility adaptations, translations. Keep humans visibly on the outside: a named person signs, a human reviews anything referencing a donor’s personal circumstances, and gratitude for major moments, first gifts, memorial gifts, bequest conversations, is composed by a person, full stop. And say what you do: a short AI use statement on the website, mirrored in privacy policies, describing which tasks are AI-assisted and how human review works. Organizations that adopted this posture early report no measurable donor backlash and, more interesting, a modest trust dividend among the roughly one-seventh of donors who actively favor organizations that innovate openly. High-capacity donors skew friendlier still: 30% of high-value donors support nonprofits using AI against 13% of small donors in one 2025 dataset, which means the audiences fundraisers most fear alienating are, on average, the least alarmed.

The authenticity problem, then, is real but tractable. What is not tractable is faking it. Voice cloning a founder for thank-you calls, synthetic beneficiary testimonials, fabricated urgency, these are not communication tactics with disclosure problems, they are frauds with production values, and the section of this article on synthetic scams returns to what they are doing to giving at large.

Service delivery reshaped by chatbots, triage, and always-on support

Fundraising gains save a charity money. Service delivery gains change what a charity is, because they alter how many people it reaches and when. This is where the most consequential nonprofit AI deployments now live, and where both the best and the worst outcomes in the sector’s short AI history have occurred.

The strongest pattern is extension of reach beyond staffed hours. The Epilepsy Foundation uses Claude to provide round-the-clock informational support for the 3.4 million Americans living with epilepsy, a population whose questions do not arrive between nine and five. The model answers from vetted foundation content, escalation paths route people to humans, and the design goal is explicit: extend the helpline, never replace the clinician or the crisis counselor. The same shape recurs across the Google.org accelerator cohorts: IDinsight built AI-powered responses to health questions from new and expectant mothers in South Africa; Jacaranda Health scaled digital maternal health services using language models trained in African languages; Partnership to End Addiction used AI to build training simulations and quality assurance for its family support services, applying the technology to make its human counselors better rather than to substitute for them.

The second pattern is navigation of hostile bureaucracy on behalf of poor people. Benefits systems are, functionally, adversarial information mazes, and AI turns out to be very good at mazes. mRelief built an assistant helping US families apply for SNAP food assistance. Justicia Lab created an AI assistant that walks immigrants through their options for legal status. MyFriendBen, working with Claude-powered agents that track more than forty benefit programs per state, reports connecting families to unclaimed benefits and tax credits four times faster than before, identifying over $1.2 billion in value for more than 70,000 households. These numbers deserve a moment of attention, because they answer the question that hangs over all nonprofit AI: the value here is not staff hours saved, it is money moved into poor households that would otherwise have gone unclaimed. That is mission impact, denominated in the mission’s own units.

The third pattern is triage: using models to sort incoming demand so scarce human expertise lands where it matters most. Crisis lines pioneered this years before generative AI, using machine learning on incoming messages to prioritize the highest-risk contacts in the queue, on the straightforward logic that in a crisis service, waiting time is not an inconvenience metric but a safety one. The same triage logic now runs in legal aid intake, housing assistance, and case management, where models pre-sort, summarize, and flag while humans decide.

The cautionary tale the whole sector cites is Tessa. In 2023, the US National Eating Disorders Association took down its wellness chatbot after users reported it giving weight-loss and dieting advice to people seeking help for eating disorders, advice of exactly the kind that feeds the illness the organization exists to fight. The failure was not exotic: a system addressing vulnerable people in an autonomous configuration drifted outside its guardrails, and the humans found out from the harmed users rather than from monitoring. Every subsequent responsible deployment in health-adjacent services has been shaped by that event, visibly or not, toward constrained knowledge bases, aggressive escalation to humans, continuous transcript auditing, and the design principle that the more vulnerable the user, the narrower the machine’s autonomy must be.

Two structural observations complete the picture. First, service-delivery AI shifts a charity’s cost curve rather than merely lowering it: the marginal cost of serving the next person falls toward the cost of compute, which is why organizations like Rising Academies can put a tutoring chatbot in front of more than 150,000 students, a number no staffing model reaches. Second, it creates a new inequality inside the served population, between people who can and will engage with a digital channel and people who cannot. UK data puts foundation-level digital exclusion at roughly 7.9 million adults, concentrated among older, disabled, and low-income groups, which overlap heavily with charity beneficiaries. A service redesign that quietly moves resources from phone and in-person channels to AI channels can improve average outcomes while abandoning the hardest cases, and no dashboard will flag it unless someone builds the flag. The organizations doing this well treat AI channels as additive capacity and protect analog routes as a matter of policy, not leftover budget.

Humanitarian cash aid and the machine learning targeting breakthrough

If one body of work proves AI can change charitable outcomes at the level of rigorous evidence rather than anecdote, it is the machine learning targeting of humanitarian cash transfers, and its origin story runs through Togo during the pandemic.

The problem it solves is old and brutal. Cash transfers are among the best-evidenced tools in poverty relief, but in low-income countries governments cannot see household income; most economic life is informal. Traditional targeting relies on surveys and social registries that take months and money to build, or on community nomination, which is slow and politically fraught. In a fast crisis, the choice was between speed and accuracy, and the poorest usually paid for whichever was sacrificed.

In 2020, Togo’s government, the nonprofit GiveDirectly, and researchers at the University of California, Berkeley’s Center for Effective Global Action built a third option for the rural expansion of Togo’s Novissi emergency cash program. The insight, later published in Nature in 2022, is that poverty leaves a signature in mobile phone metadata: call patterns, top-up behavior, data usage. The team trained machine learning models on traditional survey data linked to phone records, then used the models to screen hundreds of thousands of applicants and prioritize payments to the poorest mobile subscribers, delivered as mobile money with no field enrollment at all. Within weeks the program was paying people; it ultimately delivered roughly $10 million to around 137,000 of the country’s poorest citizens.

The evaluation is what makes this a landmark rather than a stunt. Compared with the geographic targeting the government would otherwise have used, the machine learning approach reduced exclusion errors, eligible poor people wrongly left out, by 8 to 14%, which the researchers estimate translated into 4,000 to 8,000 additional genuinely poor people receiving aid. Where blanket targeting of the poorest districts would have reached 33% of people living under $1.25 a day, the model-based approach reached 47%. The researchers were equally clear about limits: phone-based targeting works best as a rapid, supplemental tool in crises, not a wholesale replacement for surveys, because it can only see people who use phones.

GiveDirectly has since stress-tested the approach, now branded MobileAid, across contexts. In Bangladesh, a pilot with the government trained a poverty-prediction model on phone indicators, verified selected households through the national 333 helpline, and paid each 15,000 taka, about $136, in two installments between November 2023 and March 2024; the evaluation found mobile-based targeting outperformed community selection and matched survey-based methods at lower cost and higher speed. In Malawi, a hybrid design identified about 35% of recipients through phone metadata and 65% through in-person outreach, cutting the cost of identifying and enrolling each additional recipient roughly in half while still reaching households without phones. In Kenya, the same machinery is being wired to climate early-warning systems so that cash can move to flood- and drought-prone households before the shock lands, the model of anticipatory humanitarian action that agencies have talked about for a decade and can now increasingly execute.

The honest ledger has entries on both sides. On the benefit side: speed measured in weeks instead of quarters, cost per enrollment cut sharply, exclusion errors measurably reduced, and a template governments can adopt, which is the ultimate scale mechanism. On the risk side: systematic exclusion of the phoneless, who skew toward the very poorest, older people, and women in some contexts; privacy questions around repurposing telecom metadata, even anonymized, for eligibility decisions; and the sobering fact that an eligibility model is infrastructure of state power, usable next time for purposes its designers did not choose. The research community around these programs has been unusually candid about all three, which is itself a model: the fairness analyses were published alongside the accuracy results, not extracted later by critics.

For the broader question of how AI is best used for charity, Togo and its successors established the template that the most serious deployments follow: a costly bottleneck with quantified stakes, a model attacking that bottleneck specifically, rigorous evaluation against the incumbent method including error and fairness metrics, hybrid designs that keep human and analog channels for the people the data cannot see, and publication of results either way.

Disaster response, satellite imagery, and damage maps in hours instead of weeks

When an earthquake or cyclone hits, the first operational question is always the same: where is the damage, and how bad? Every downstream decision, where to send search teams, how much shelter to stage, which roads to assume gone, waits on that answer. For decades the answer came from expert analysts at the UN Satellite Centre, UNOSAT, manually comparing before-and-after satellite images building by building, work that is respected as the industry’s best and that, for large disasters, takes days to weeks the response cannot afford. Computer vision has now compressed that timeline to hours, and the compression is operational, not experimental.

The foundations were laid by open research infrastructure. The xBD dataset, built with disaster response agencies, provided over 850,000 annotated buildings across 45,000 square kilometers of paired pre- and post-event imagery spanning earthquakes, storms, floods, and fires, and the associated xView2 challenge produced a generation of damage-classification models. Microsoft researchers subsequently showed deployable models running three times faster than the fastest challenge winners while remaining accurate enough for operational use, built explicitly for a humanitarian stakeholder to run in the field.

The flagship operational system is SKAI, co-developed by the World Food Programme’s Innovation Accelerator and Google Research and released as open source. SKAI analyzes satellite imagery to assess building damage at scale across disaster types, combining Google’s Open Buildings footprint data with zero-shot and fine-tuned damage models, meaning it produces useful first assessments even for places and disaster types it was never specifically trained on. Its 2025 record shows what operational maturity looks like: after the March 2025 Myanmar earthquake, SKAI assessed nearly 500,000 buildings to support the early response, and it analyzed more than 300,000 following the Sri Lanka cyclone, with a radar-based variant, able to see through clouds that blind optical satellites, tested after an earthquake in Afghanistan. A follow-on model for detecting temporary settlements, the tarps and shelters that mark displaced populations, performed strongly in trials and moved toward pilot deployment with WFP country offices in early 2026. In parallel, the UN’s DISHA coalition paired UNOSAT with Google Research to build AI-assisted damage assessment directly into the UN’s own workflow, positioning the models as force multipliers for the human analysts whose judgment remains the quality bar, exactly the assistive configuration this article keeps flagging as the one that works.

The same modeling families now run before disasters, not just after. Flood forecasting models publish river-level predictions days ahead for basins covering large parts of Africa and Asia where gauge networks are sparse. Famine early warning combines satellite vegetation data, prices, and conflict signals. WFP’s HungerMap-style monitoring fuses survey and remote data into near-real-time food security estimates. The strategic significance is the same as in the cash-transfer story: prediction converts response into anticipation, and money spent before a shock reliably buys more protection than money spent after.

The limits are worth stating plainly, because disaster AI is often oversold. Optical satellites cannot see through cloud, which is precisely the weather that accompanies many disasters; radar helps but classifies damage less finely. Models trained mostly on one region’s building styles degrade on another’s. Damage visible from directly above understates damage to walls and interiors, so ground truth still matters, and every serious deployment keeps human analysts in the loop for exactly that reason. None of this dulls the headline: assessment that took weeks now takes hours, at a coverage no analyst team could match, and in disaster response, time is the currency that converts directly into lives.

Conservation charities as unexpected AI pioneers

Long before chatbots reached charity offices, conservation organizations were running some of the most sophisticated applied AI in the entire social sector, driven by a simple asymmetry: nature generates data at a volume no human workforce can process, and the threats to it move faster than manual monitoring can track.

Camera traps illustrate the volume problem. A network of motion-triggered cameras across a protected area produces millions of images a year, most of them empty frames or swaying grass. Platforms like Wildlife Insights apply species-recognition models that classify images automatically, collapsing months of manual sorting into hours and turning camera traps from an archival tool into a near-real-time monitoring system. iNaturalist, whose community has amassed one of the largest biodiversity datasets on Earth, joined the 2025 Google.org accelerator cohort to use generative models to convert thousands of expert identification remarks into natural-language explanations, teaching the crowd what the experts see. Acoustic monitoring extends the same pattern to sound: models trained on bird calls, whale song, and chainsaw noise listen to forests and oceans continuously, flagging both species presence and illegal activity from audio streams no team of humans could audit.

Anti-poaching is where conservation AI became genuinely strategic. PAWS, the Protection Assistant for Wildlife Security, developed by Milind Tambe’s research group, models poaching as a security game: rangers with limited patrol hours defending a vast area against adaptive adversaries. Trained on historical poaching incidents, patrol records, and terrain features, PAWS predicts where snares and poaching activity are most likely and generates randomized patrol routes that concentrate effort on high-risk zones while staying unpredictable. Field tests in Uganda and Cambodia found patrols guided by the system uncovering substantially more snares and illegal activity in predicted hotspots, and through integration with SMART, the patrol-management platform used across hundreds of protected areas by a consortium including WWF, the Wildlife Conservation Society, and Panthera, the approach became available to protected-area managers worldwide rather than remaining a research demonstration. Complementary systems attack other links in the chain: connected cameras with on-device detection that alert rangers to vehicles and humans in restricted zones, drone-mounted models that one network estimates let rangers detect poachers many times faster than manual sweeps, and radar techniques for mass-detecting snares beneath canopy.

The pattern extends across environmental work. Global Fishing Watch applies machine learning to vessel transponder and satellite data to map industrial fishing worldwide and expose illegal fleets, evidence that has directly supported enforcement actions and marine protection designations. Earth Genome, another accelerator participant, translates satellite imagery into plain-language environmental intelligence so that a frontline organization without a remote-sensing team can ask where illegal mining expanded this quarter and get an answer. Reforestation groups use vision models to audit planting survival from drone imagery instead of walking transects. Climate-focused funders, meanwhile, are pushing the frontier upstream: Google.org’s $30 million AI for Science challenge, opened in 2026, funds work from AI-guided crop disease resistance to foundation models for reducing livestock methane.

Conservation’s decade of experience yields two lessons the rest of the sector is only now learning. First, the binding constraint was never model quality; it was the pipeline around it, getting data off cameras in places without connectivity, getting predictions into a ranger’s hands in a usable form, getting institutions to change patrol planning because of them. Second, adversaries adapt. Poachers learn patrol patterns, which is why PAWS randomizes; illegal fleets turn off transponders, which is why Global Fishing Watch cross-references radar. Any charity deploying AI against a human adversary, fraud, trafficking, disinformation, inherits the same arms-race logic, and should budget for it.

Health and mental health organizations on the AI frontier

Health charities operate under the most demanding conditions in the sector: the informational stakes are clinical, the users are often at their most vulnerable, and the regulatory and ethical scrutiny is correspondingly intense. That makes their AI record unusually instructive, because it contains the clearest successes and the most instructive failure.

The successes cluster around information access and workforce augmentation. The Epilepsy Foundation’s always-available support service, discussed earlier, extends vetted medical information to 3.4 million Americans on their own schedule. Jacaranda Health scales digital health services to underserved mothers using natural language processing trained in African languages, meeting patients in Kiswahili rather than requiring them to meet the system in English. IDinsight’s maternal health assistant answers questions from new and expectant mothers in South Africa promptly and accurately enough to matter in the window where maternal outcomes are decided. THINK, a South African public health organization, found that health professionals initially wary of generative tools came out of structured training using them to sharpen analysis and program design, and is building a toolkit to spread that capability across public health programs. In Guatemala, a global health team used Claude to build an interactive geospatial tool mapping dengue-risk populations in three days instead of weeks, giving the health ministry a concrete basis for deciding where prevention money buys the most protection, a small story that captures a large truth: in shrinking funding environments, targeting quality is a form of capacity, and AI has made targeting radically cheaper.

Mental health and crisis services adopted machine learning earlier than almost anyone, in the narrow, high-stakes form of triage. Crisis Text Line built models on millions of anonymized conversations to identify which incoming texters faced the highest risk, moving them to the front of the queue; the organization reported that its models surfaced risk signals human intuition missed, and the design kept every conversation with a trained human counselor, using AI only to decide who could not wait. That allocation-only configuration has aged well precisely because of what happened when others went further. The Tessa failure at the National Eating Disorders Association, where an autonomous wellness chatbot gave dieting advice to people with eating disorders before being taken down in 2023, remains the sector’s defining cautionary case, and its lesson has hardened into something close to consensus: in mental health contexts, generative AI may support counselors, drafting, training simulation, post-conversation summarization, quality review, and may handle bounded informational tasks with tight guardrails, but the therapeutic relationship itself is not a delegation target.

Beyond services, AI is reshaping the research charities that fund cures. Protein-structure prediction of the AlphaFold generation compressed years of structural biology into days and is now routine infrastructure in disease research, including the neglected diseases where charitable funding dominates because markets fail. Medical research charities increasingly fund AI-driven drug discovery for exactly those gaps, and diagnostic models, screening retinal images, skin lesions, tuberculosis X-rays, are moving through nonprofit and public health channels to populations commercial medicine underserves, with Visilant’s AI-assisted eye screening across India, backed by the Google.org accelerator, a representative example.

The synthesis for health-adjacent charities is a gradient of autonomy calibrated to vulnerability. Information retrieval from vetted content: broad autonomy with monitoring. Navigation and logistics: moderate autonomy with escalation. Anything touching diagnosis, treatment, or a person in crisis: human decision, machine support, no exceptions. Organizations that internalized that gradient are compounding gains; the ones that did not supplied the sector’s warning stories.

Education nonprofits and AI tutors for underserved learners

Education is where the theoretical case for charitable AI is strongest, because the intervention with the best evidence in the field, one-to-one tutoring, is also the least affordable one. Benjamin Bloom documented in the 1980s that tutored students dramatically outperform classroom peers; the problem was always that tutors do not scale. Language models are the first technology that plausibly attacks that constraint directly, and education nonprofits have moved on it faster than school systems.

Khan Academy, a nonprofit throughout, built Khanmigo as an AI tutor and teaching assistant designed around pedagogy rather than answers: it questions, hints, and scaffolds instead of solving, and it gives teachers lesson planning and progress-analysis support. Its significance for the sector is partly the product and partly the posture, a nonprofit demonstrating that mission organizations can build frontier-adjacent tools with guardrails commercial products lacked, then publish what they learn about where the tutor works and where it stumbles.

The more radical deployments are in low- and middle-income countries, where the counterfactual is not a human tutor but nothing. Rising Academies runs chatbot tutors reaching more than 150,000 students across Sub-Saharan Africa, delivering structured practice over low-bandwidth channels including basic phones. EIDU, backed by the Google.org accelerator, provides personalized tutoring content for teachers and customized digital exercises for students in low-income countries. Darsel delivers a math tutor adapting to individual student needs across K-12 globally. Quill.org gives students immediate feedback and coaching on writing and reading comprehension, an application with a specific economic logic: writing feedback is the most labor-intensive thing a teacher does, so automating the first pass returns hours to instruction while students get responses in seconds instead of days.

Adjacent to schooling, workforce nonprofits apply the same machinery to the transition into employment. Tabiya, addressing global youth unemployment, built Compass, an open-source conversational agent that helps young jobseekers translate informal experience into recognized skills; after its accelerator cohort, it reported reaching more than 8,000 jobseekers in half the time and at a quarter of the previous cost, one of the cleaner unit-economics results in the whole nonprofit AI record. Opportunity@Work uses AI to help employers see career paths that do not run through degrees, widening the door for workers skilled through alternative routes. And a fast-growing category of nonprofit work is AI literacy itself, teaching students, workers, and community organizations to use and judge these tools, an area OpenAI’s People-First AI Fund and groups like Day of AI have made a funding priority on the theory that distributional fairness in the AI era starts with who knows how to use it.

The evidence discipline matters more in education than anywhere else, because edtech has a long history of tools that demo well and teach nothing. The credible deployments share three properties: they are grounded in curriculum rather than open-ended chat, they measure learning outcomes against comparison groups rather than engagement minutes, and they position the teacher as the client, not the competitor. Early results from the grounded deployments are genuinely encouraging; early results from ungrounded chatbot-in-a-classroom experiments are not, and the difference is design, not model quality. For education charities, the strategic opening is unusually clear: the marginal cost of tutoring has collapsed, the pedagogy that makes it work is public knowledge, and the populations that benefit most are exactly the ones commercial edtech ignores.

Refugee support, translation, and the language access breakthrough

Language is the quiet tax on humanitarian work. Every intake interview, asylum form, medical consultation, and information leaflet that crosses a language boundary costs interpreter time, and interpreter time is scarce in exactly the languages crises speak. Machine translation was long too unreliable for high-stakes use in precisely those languages; that has changed enough, in the last three years, to reorganize how refugee-serving organizations work.

Tarjimly shows the hybrid model. The nonprofit connects refugees and aid workers to volunteer translators across more than a hundred languages, and it now uses AI to expand the accessibility, quality, and scale of that service, machine translation handling routine exchanges and drafting, humans handling the medical, legal, and protection conversations where an error changes an outcome. Supported by a Google.org Fellowship in which Google employees worked full-time alongside its team, and by the Patrick J. McGovern Foundation for extending language access to refugee, immigrant, and indigenous communities, Tarjimly’s design accepts the technology’s real limits and allocates human attention exactly where those limits bite. CLEAR Global, which grew out of Translators without Borders, works the deeper layer of the same problem: building language technology and datasets for marginalized languages that commercial providers skip because the markets are small, on the blunt observation that a language absent from training data is a population absent from every AI benefit this article describes.

Large humanitarian agencies apply translation and generation across their internal machinery. The International Rescue Committee, operating in 40 countries with 12,000 staff across dozens of working languages, uses Claude to communicate with local partners and analyze field data faster in time-sensitive settings, and its Signpost program, information services for displaced people, has been an early adopter of AI-assisted content localization, keeping guidance current across languages and rapidly shifting legal contexts, work no human editorial team could sustain at that cadence. UNHCR and partner agencies apply speech recognition and translation to make feedback mechanisms actually multilingual, so that a complaint or protection report in a minority language does not wait weeks for a translator before anyone acts.

The risk profile here is specific and serious. Translation errors in asylum contexts have consequences measured in refoulement, and documented cases of machine-translation errors in asylum applications predate the current model generation. Dialect coverage is uneven; a system fluent in standard Arabic may mangle the Sudanese or Levantine speech of the person actually in front of the caseworker. And refugee data is among the most sensitive any charity holds, since disclosure can endanger family members still in the country of origin, which rules out consumer-grade tools for anything containing protection information. The operating rules mature organizations follow mirror the health sector’s: machine translation for information out and routine dialogue, qualified humans for legal and protection decisions, and clear signaling to the person which one they are getting. Within those rules, the gain is transformative in the plainest sense: a caseworker who once served the clients who matched her languages now serves the clients who need her, and the leaflet that existed in five languages exists in forty.

Big tech’s AI-for-good programs compared across Google, OpenAI, Anthropic, and Microsoft

Between 2024 and 2026, every major AI provider built a dedicated nonprofit program, and the differences between them reveal competing theories of what the sector needs: money, discounts, integrations, or training. Charities choosing where to standardize should understand all four, because the offers are not interchangeable.

Google runs the oldest and broadest machine. Google for Nonprofits provides deep product discounts, roughly 70 to 75% off Workspace tiers whose paid versions now bundle Gemini, alongside the long-running Ad Grants program, whose AI-driven campaign types quietly matter to charity visibility. The distinctive layer is Google.org’s grant capital. The Generative AI Accelerator’s first cohort in 2024 put more than $20 million behind 21 nonprofits building AI products, from Full Fact’s health-misinformation summarizer to mRelief’s benefits assistant; the 2025 open call added $30 million for 20 more, each receiving $500,000 to over $2 million plus six months of pro bono support from Google engineers and Cloud credits. Two further $30 million challenges followed, one for AI in government services and one, closing applications in May 2026, for AI in science across health and climate. Google’s theory is explicit: fund a comparatively small set of organizations to build durable AI products, then let the sector adopt what works. Graduates like Tabiya, reaching jobseekers at a quarter of prior cost, are the proof cases the theory rests on.

OpenAI’s program is the widest funnel. Verified nonprofits get discounted ChatGPT Business, reported by implementation partners at around $8 per user per month on annual terms in 2026, with negotiated discounts of up to 75% on Enterprise for larger organizations, alongside a free OpenAI Academy training track. The distinctive layer is cash and convening: the $50 million People-First AI Fund distributed $40.5 million in unrestricted grants across 208 US nonprofits in December 2025, deliberately targeting community organizations with budgets between $500,000 and $10 million rather than technology builders, and a 2026 round followed. The Nonprofit Jam gathered more than a thousand nonprofit leaders across ten cities for hands-on building, each leaving with a year of free ChatGPT Plus. A joint fund with the GitLab Foundation backs nonprofits applying AI to economic opportunity, with API credits even for unsuccessful applicants. OpenAI’s theory: put the tools and modest money in as many mission-driven hands as possible and let use cases emerge bottom-up.

Anthropic entered last and most surgically. Claude for Nonprofits, launched with GivingTuesday on December 2, 2025, offers up to 75% off Team and Enterprise plans, Team seats from $8 per user per month with Claude Code and Cowork included, but its distinctive layer is plumbing: open-source MCP connectors that let Claude work directly inside Blackbaud’s donor systems, Benevity’s database of over 2.4 million validated nonprofits, and Candid’s foundation and grants data, so a fundraiser can query giving patterns or research funders in natural language against live records. The program was shaped by months of piloting with Robin Hood, Tipping Point Community, and the Constellation Fund across more than 60 grantee organizations, and it ships with a free AI Fluency for Nonprofits course built with GivingTuesday. Anthropic’s theory: the sector’s bottleneck is not tool access but workflow integration, so build the integrations. Reference customers give the theory teeth, IDinsight reporting work up to 16 times faster, the IRC using Claude across a 40-country operation.

Microsoft, the incumbent inside most charity offices, offers granted and discounted Microsoft 365 with about 15% off Copilot seats, plus the AI for Good Lab, which partners its own researchers with humanitarian, health, and conservation organizations on applied problems like the post-disaster damage assessment work described earlier. Its theory is embedding: AI arrives inside the tools staff already use, which minimizes adoption friction and, critics note, maximizes lock-in.

Read together, the programs also serve corporate interests, seeding future markets, generating goodwill during a period of regulatory attention, and enlisting sympathetic use cases, which does not make them less useful but does mean charities should negotiate like customers, not supplicants. Practical guidance follows from the differences: organizations wanting funded product development look to Google.org’s open calls; small teams wanting cheap general capability compare the $8-per-seat tiers from OpenAI and Anthropic; fundraising-heavy shops on Blackbaud weigh Anthropic’s connectors; Microsoft-standardized organizations test Copilot before adding anything. And every organization should note the pattern beneath the offers: discounts of this depth are promotional-era pricing, and dependence built at 75% off should be stress-tested against what the tools cost at list price.

Philanthropy’s new money and the foundations funding AI for humanity

Corporate programs get the headlines, but a parallel philanthropic infrastructure has quietly become the sector’s most important source of patient capital for AI, and its priorities differ from big tech’s in ways that matter.

The anchor institution is the Patrick J. McGovern Foundation, the legacy of the IDG founder, which has made AI and data science for social benefit its entire grantmaking identity. Its recent pace is striking by foundation standards: $73.5 million across 144 organizations in 11 countries in 2024, followed by $75.8 million across 149 grants, bringing its cumulative public-purpose AI and data commitments to roughly half a billion dollars. The portfolio spans the full stack this article has covered, crisis response and AI-enabled emergency dispatch, climate resilience, health equity, human rights documentation, journalism, refugee language access through Tarjimly, and, distinctively, the governance layer: support for the OECD’s AI policy work, UNESCO’s AI ethics implementation across dozens of countries, the AI Safety Fund for responsible frontier development, and AI literacy organizations like EqualAI. Beyond checks, its Data Practice Accelerator embeds data and AI capacity directly into nonprofits, treating capability as the grant, an approach that addresses the sector’s real constraint more directly than tool discounts do.

Around that anchor, a wider funding ecosystem has formed. data.org, backed by the Rockefeller Foundation and Mastercard’s Center for Inclusive Growth, funds data and AI capacity across the social sector globally. Schmidt-family philanthropy pushes AI into science and, more recently, the humanities. Mozilla funds democratic and open-source AI tools. Fast Forward accelerates tech nonprofits as a category. Community foundations and regional funders, Robin Hood and Tipping Point among them, now underwrite AI adoption among their own grantees, having piloted the tools themselves. And the corporate foundations blur into this space: Google.org’s challenges are structured like philanthropy, and OpenAI’s fund was designed with an external commission of nonprofit leaders precisely to import philanthropic legitimacy.

Two tensions run through this money, and honest observers on all sides acknowledge them. The first is direction-setting: when the largest funders of nonprofit AI are AI companies and tech-fortune foundations, the sector’s technology agenda risks being written by its vendors. Unrestricted grants, like the People-First Fund’s, and governance-focused portfolios, like McGovern’s, are partly answers to that critique, and the critique itself has been productive, pushing funders toward community-defined priorities. The second is the crowd-in question: whether AI funding adds to philanthropic budgets or cannibalizes program funding. The evidence in 2026 is mixed but leans additive; the sharper problem is distribution, because AI grants flow disproportionately to organizations already strong enough to absorb them, the same pattern the digital divide section of this article examines.

For charities seeking this money, the field has legible rules. Foundation AI funders in 2026 reward institutional readiness over technical novelty: a defined problem with quantified stakes, existing data or a plan to build it, governance in place, cross-sector partnerships, and measurable public impact. Grant seekers should also mind the layer beneath project funding, because several funders, McGovern prominent among them, will fund the unglamorous foundations, data infrastructure, staff capability, evaluation, that every glamorous application depends on. In a sector where 47% of organizations lack even an AI policy, money for foundations may be the highest-return philanthropy available.

Pricing, discounts, and the real cost of AI for a charity

Budget conversations about AI in charities tend to fixate on subscription prices, which are now the smallest and best-behaved line in the true cost. A realistic accounting has four layers, and organizations that price only the first layer are the ones that later report AI as an expense without a return.

The first layer, licenses, has become genuinely cheap for verified nonprofits. Claude for Nonprofits prices Team seats from $8 per user per month at up to 75% off, verified in a few minutes through Goodstack, with the discount persisting as long as charitable status does; Claude Code and Cowork ride along on every seat, and eligibility extends beyond classic 501(c)(3) equivalents to K-12 schools and mission-based rural health providers. OpenAI’s nonprofit pricing has deepened over time from the original 20% off announced in 2024 to Business-tier pricing that implementation partners in 2026 report at around $8 per user per month annually, with Enterprise discounts up to 75% negotiated through sales. Google’s nonprofit tiers discount Workspace, where Gemini now lives, by roughly 70 to 75% depending on edition. Microsoft grants basic Microsoft 365 to small nonprofits and discounts Copilot around 15%. For a ten-person charity, frontier AI capability across a whole team now costs roughly what one staff mobile phone plan did, on the order of $1,000 a year, a price at which the old affordability objection has simply expired. International eligibility matters for readers outside the US: all the major programs verify equivalents of charitable status across most countries, typically through Goodstack or similar validators, so a Slovak, Kenyan, or Brazilian nonprofit faces the same short form a Californian one does.

The second layer is data readiness, and it is where real money hides. AI grounded in an organization’s information is only as good as that information: a donor database full of duplicates, a shared drive with no structure, program records living in seven spreadsheets. Most organizations discover that their first serious AI project is secretly a data-cleanup project, weeks of staff time or a consultant engagement that dwarfs the license fee. This cost is front-loaded and nonrecurring, and it pays back across every future use, but it must be budgeted or the project stalls exactly where most sector projects stall.

The third layer is people: training, workflow redesign, and governance time. The free curricula, Anthropic’s AI Fluency for Nonprofits, OpenAI Academy, Google’s offerings, remove the course-fee excuse, but course hours are staff hours, and the workflow redesign that separates the 7% from the 92% is senior staff time by definition. A sensible planning figure for a small organization’s first year is two to four staff days per employee across training and process work, plus a governance rhythm, an owner, a policy, a quarterly review, that costs a few leadership hours a month. Organizations comparing this to the 15 to 20 hours per week of administrative time that sector estimates suggest automation returns to finance and executive teams can see the shape of the payback; organizations that skip the investment collect neither.

The fourth layer is variable and strategic: API usage for anything custom-built, which scales with volume and needs a ceiling in the budget; integration or consulting for connecting AI to CRMs and case systems, mitigated but not eliminated by the new standard connectors; and the dependence risk noted earlier, that promotional-era discounts define today’s prices, not tomorrow’s. Prudent treasurers model continuity at list price and prefer architectures, open standards like MCP, exportable data, documented prompts and workflows, that keep switching costs low.

Grants can offset every layer. Google.org’s accelerator checks run $500,000 to over $2 million for organizations building AI products; OpenAI’s fund pays unrestricted five- and six-figure grants to community nonprofits; McGovern’s Data Practice Accelerator funds up to $125,000 of capability building; the GitLab Foundation partnership hands API credits even to rejected applicants; and UK and EU digital funders increasingly accept AI readiness as a legitimate core cost. The honest bottom line for 2026: for organizations that budget all four layers, AI is among the highest-return spending available to them; for organizations that budget only the subscription, it is a small recurring cost that buys a plateau.

Inside the efficiency plateau, the gap between adoption and impact

The defining statistic of nonprofit AI in 2026, 92% adoption against 7% major impact, deserves closer mechanical inspection, because the gap is not mysterious once decomposed, and each cause has a known remedy.

Start with what the plateau feels like from inside, because most readers work there. Individual staff use AI daily and genuinely faster: appeals drafted in an hour instead of an afternoon, reports summarized on the way into meetings. Nearly four in five organizations report exactly these small-to-moderate gains. But the gains stay personal. When the person who knows the prompts leaves, the capability leaves with them. Different staff produce different quality with the same tools. Time saved diffuses into the general overload rather than being reinvested deliberately. And nothing about the organization’s core numbers, retention, cost per outcome, people served, moves. The Virtuous benchmark quantifies each link in that chain: 65% of organizations describe their use as reactive and individual, 81% use AI without shared workflows, only 4% have documented, repeatable processes, and 47% operate with no governance policy that would make scaling safe even if someone tried.

The report’s readiness framing sorts the sector into rough fifths: about one fifth has foundations in place, governance, documentation, measurement, and is compounding; another fifth is experimenting energetically and at risk of plateauing; the majority uses AI actively with no structure at all. The barriers migrate as organizations mature, which is diagnostic in itself. Early-stage organizations cite tools and knowledge; daily-use organizations cite time, privacy concerns, and staff skepticism, the frictions of integration rather than access. Notice what almost nobody cites anymore: model capability. The technology stopped being the constraint somewhere around 2024; the organization became the constraint, which is uncomfortable and also hopeful, because organizations are fixable in ways one cannot fix a model from a charity office.

What do the 7% actually do differently? Cross-referencing the benchmark’s findings with the case evidence assembled across this article yields a consistent list. They pick whole workflows, not tasks: lapsed-donor recovery end to end, grant reporting end to end, intake-to-case-note end to end, redesigned around the tool rather than decorated with it. They connect AI to their systems of record, CRM, case management, document repositories, so outputs are grounded in real data instead of pasted fragments, which is what the connector infrastructure of 2025-2026 exists to enable. They standardize and document, shared prompt libraries, templates, review checklists, so quality stops depending on which staff member is typing. They govern proportionately, a short policy, named ownership, escalation rules for sensitive data, which paradoxically accelerates adoption because staff stop guessing what is allowed. They measure against mission metrics, and they deliberately reinvest saved hours into named activities, more donor calls, more casework, rather than letting the surplus evaporate. And their leadership treats this as operating-model change, budgeted and sponsored, not as a software rollout delegated downward.

None of these moves is technical, and the benchmark’s authors emphasize that the foundations take weeks, not years, to build. That is the genuinely encouraging reading of the plateau: the sector’s problem is not talent or money in the amounts usually assumed, it is sequencing. Organizations that stop asking staff to use AI more and start choosing one workflow to rebuild properly tend to cross from the 92% toward the 7% within quarters. The gap, meanwhile, is widening in both directions, compounding gains on one side, accumulating unmanaged risk on the other, which is why the plateau is better understood as a fork.

Governance gaps, board literacy, and the policy vacuum

The most dangerous number in the sector’s AI data is not an adoption figure; it is the governance figure sitting beside it. Forty-seven percent of nonprofits report no AI policy of any kind, per the 2026 US benchmark; UK data tells the same story in motion, with only 6% of charities holding an AI policy as of mid-2024 and about 48% developing one by 2025, meaning policy is chasing practice with roughly a two-year lag. During that lag, the sector’s AI use has been governed, in the literal sense, by nobody.

Shadow use fills the vacuum. Staff across the sector quietly integrate AI into drafting, summarizing, analysis, and funding applications without telling anyone, and the silence is rational: Microsoft and LinkedIn’s Work Trend Index found 52% of people using AI at work reluctant to admit it on important tasks, fearing it makes them look replaceable or lazy. In a charity context, shadow use has sharper edges than in a company, because the pasted text may be a case note about a vulnerable client, the funding application may carry an implicit representation of authorship, and the reputational blast radius of a public error includes donor trust, which does not regenerate on a product cycle. The organizations handling this well have drawn the obvious conclusion from the psychology: amnesty plus clarity beats prohibition. They surface existing use without blame, then channel it through approved tools and rules, converting their most enterprising staff from hidden risk into internal champions.

Leadership literacy is the second gap, and it is measured, not anecdotal. In the UK survey, over a third of respondents rated their chief executive’s AI skills, knowledge, and confidence as poor, and four in ten said the same of their board. This matters because AI governance is now unambiguously a board matter. The UK Charity Commission has signaled that trustees’ existing duties, prudence, care, acting in the charity’s interest, extend to AI use, and the Fundraising Regulator’s December 2025 guidance made trustee accountability for AI in fundraising explicit. Ethical AI oversight is settling in beside cybersecurity and financial controls as a standing governance responsibility, and boards that cannot ask a single informed question about how their organization uses AI are, in 2026, failing a duty they may not know they have.

What proportionate governance looks like is, fortunately, well established and small. A one- to two-page policy answering five questions: which tools are approved and which are banned; what data may never enter an AI system, with beneficiary personal data and special-category information named explicitly; where human review is mandatory, funder-facing documents, anything published, anything touching an individual’s circumstances; who owns the policy and answers questions; and how incidents get reported without punishment. Around the policy, a light rhythm: AI on the risk register, a named senior owner, a quarterly review of what is being used for what, and board-level exposure at least annually, ideally including a hands-on session, because trustees who have used the tools govern them better than trustees who have only read about them. Vendor governance belongs in the same page: before any tool touches organizational data, someone checks whether inputs train the vendor’s models, where data is stored, and what the terms say about confidentiality, the questions that separate enterprise-grade deployments from consumer accounts holding a charity’s crown jewels.

The governance gap is the cheapest major risk in the sector to close, which makes its persistence a puzzle with an unflattering answer: policy work is unglamorous, nobody funds it, and until something goes wrong it appears optional. The regulator sections later in this article explain why “optional” has a shrinking shelf life, in Europe especially.

Privacy, beneficiary data, and the duty of care

Charities hold data that would make a bank compliance officer sweat: mental health disclosures, immigration status, addiction histories, domestic violence records, children’s information, HIV status, criminal records, debt details. They hold it about people with the least power to object to its misuse and the most to lose from its exposure. AI multiplies both the usefulness of that data and the routes by which it can leak, which makes privacy the domain where nonprofit AI ethics stops being abstract.

The primary leak route is mundane: staff pasting sensitive text into consumer AI tools. A case summary pasted into a free chatbot may, depending on the service and settings, be retained, reviewed, or used to train models; either way it has left the organization’s control, likely without the legal basis GDPR and equivalent regimes require and certainly without the client’s meaningful consent. The fix is equally mundane and non-negotiable: sensitive work happens only in enterprise-grade deployments whose terms exclude training on customer data and specify retention, region, and security standards, the configuration all major nonprofit programs now offer, and the policy names which data classes may go where. European charities carry the extra weight of GDPR mechanics: processing personal data through an AI vendor makes that vendor a processor requiring a data processing agreement; novel or large-scale processing of vulnerable people’s data triggers a data protection impact assessment; and purpose limitation means data collected to deliver a service cannot be silently repurposed to train or feed a model, however good the intention.

Beyond leaks, three subtler duties define the frontier. Minimization: the question is not whether AI could use all the data but whether the task needs it; a triage model may perform nearly as well on far less intrusive features, and the extra performance rarely justifies the extra exposure. Anonymization and pseudonymization before analysis should be default practice, with the humility to know both can fail against re-identification when datasets combine. Transparency toward beneficiaries: people have a right to know when a machine is in their loop, whether drafting the letter about their benefits or ranking their place in a queue, and privacy notices written for donors rarely cover clients. The organizations taking this seriously write plain-language explanations of their AI use for the people they serve, not just for regulators. Security of the new surface: prompt libraries, AI-generated case summaries, connector credentials, and agent permissions are all new assets to protect; an agent with write access to a case management system is exactly as dangerous as its credentials are exposed. The humanitarian sector adds the extreme case that clarifies the principle: refugee and protection data can endanger lives in origin countries if breached, which is why agencies treat it under data protection frameworks stricter than commercial law requires, and why any charity serving at-risk populations should map which of its data would be catastrophic, not merely embarrassing, to expose, and wall that data off from all but the most controlled AI use.

None of this argues against using AI on sensitive work; it argues for earning the right to. Handled through proper deployments with minimization and human review, AI applied to case notes, safeguarding patterns, and service data has produced some of the genuine impact stories in this article. The duty of care is not a fence around the technology. It is the price of admission, and it is affordable to any organization willing to treat its beneficiaries’ data with the seriousness it treats its donors’ money.

Bias and the risk of automating the inequalities charities exist to fight

Every predictive system a charity deploys encodes a theory of who deserves attention, and it learns that theory from history. History, in the populations charities serve, is a record of inequality. This is the deepest ethical problem in nonprofit AI, deeper than privacy and harder than accuracy, because it produces harm precisely when the system works as designed.

The mechanics are unglamorous. A donor propensity model trained on past giving learns that the profile of a good prospect looks like the organization’s historical donor base, and quietly deprioritizes cultivating anyone who does not, freezing yesterday’s demographics into tomorrow’s pipeline. A benefits-navigation assistant tested mainly in English performs worse for the immigrant families who need it most. A triage model trained on completed case files underweights populations the organization historically failed to retain, reading its own past service gaps as evidence of lower need. A translation system covering major languages fluently and minority dialects poorly delivers its errors selectively to the least powerful users. None of these systems is malicious; each is a mirror, and the sector’s data reflects the world the sector is trying to change.

The humanitarian targeting literature, to its credit, confronted this openly rather than defensively. The Togo research published fairness analyses alongside accuracy results and named the structural bias plainly: phone-based targeting cannot see people without phones, who skew poorer, older, more rural, and more often women. The mitigations that emerged are the field’s best current playbook. Hybrid designs preserve analog channels, as in Malawi, where two-thirds of recipients were enrolled through in-person outreach precisely so the model’s blind spots would not become exclusion. Error audits are disaggregated, asking not just how often the system is wrong but for whom. And the choice of comparison is kept honest: the relevant question is never whether the model is biased, all methods are, but whether it is more or less biased than the incumbent method at the same cost, a test machine targeting in Togo passed against geographic targeting and would fail against a well-funded census that did not exist.

For charities buying rather than building, the discipline translates into procurement questions and operating rules. Ask vendors what populations their models were trained and validated on, and what performance looks like across the demographic lines that matter to your mission; a vendor without an answer has answered. Keep humans deciding wherever stakes touch individual lives, using scores to allocate attention rather than to allocate outcomes, the configuration in which a wrong score costs a phone call instead of a benefit. Monitor outputs by group over time, because bias drifts as populations and data change. And write down, in the AI policy, the decisions the organization will never automate: safeguarding judgments, eligibility denials, anything with an appeal right attached.

There is also a mission-level version of this risk that deserves naming, sometimes called automation of the status quo. Charities exist because markets and states misallocate; AI trained on how resources currently flow is a machine for perpetuating current flows more cheaply. The organizations using AI most powerfully invert this: they aim the technology at the misallocation itself. The benefits-navigation deployments that recovered over a billion dollars for low-income households are anti-bias engines in the most practical sense, using machine reading of hostile bureaucracy to deliver entitlements the status quo was structured to withhold. The same tools that can automate inequality can automate its correction; the difference is entirely in what problem the organization points them at, which is a leadership choice, not a technical one.

Deepfakes, synthetic scams, and the new fraud threat to giving

Charitable giving runs on a fast emotional circuit: see suffering, feel moved, act. Generative AI has industrialized the counterfeiting of exactly that circuit, and the resulting fraud wave is no longer hypothetical. It is measured in the FBI’s fraud statistics and visible after every major disaster, and it attacks something charities collectively own, which is public willingness to believe an appeal.

The pattern is now consistent enough to describe as a playbook. Within hours of a disaster, fabricated images and videos of the devastation circulate, often interleaved with real ones, a sequencing researchers studying the phenomenon call a truth sandwich because it borrows credibility from the authentic material around the fake. Synthetic celebrity content amplifies reach; after the July 2025 Texas Hill Country floods, viral AI images showed famous athletes rescuing animals and distributing supplies, none of it real, all of it captioned to route viewers toward fraudulent donation links. During Hurricane Helene in 2024, an AI-generated image of storm victims spread so widely that public officials shared it before deleting. The FBI attributes over $200 million in losses to deepfake-enabled scams in the first quarter of 2025 alone, with disaster charity appeals a recurring category, and the pattern is global; Cambodian authorities in 2025 warned publicly about AI-generated images of orphaned children and grieving families attached to QR codes soliciting donations. Security researchers add a forward-looking warning with experimental backing: language-model agents can already run persuasive fundraising conversations autonomously, which in criminal hands means charity scams that scale like software and converse like people.

The damage lands twice. Defrauded donors lose money, and the intended beneficiaries lose the help that money represented. But the second-order harm is the one that should concentrate the sector’s attention: every exposed fake trains the public to discount emotional appeals, authentic ones included. When donors learn that a crying child in a flood photo may be synthetic, the marginal real appeal converts less. Trust is a commons, AI-enabled fraud is grazing it, and legitimate charities are the parties with the most to lose from its depletion.

The defensive agenda is taking shape on three levels. Regulators and platforms: the US FTC finalized an impersonation rule enabling direct federal action against scammers posing as organizations, and has weighed extending liability toward AI tools knowingly used for impersonation; the EU’s amended AI framework bans certain synthetic-content abuses outright and phases in machine-readable marking of AI-generated content from late 2026, imperfect measures that at least raise the cost of fakery. Donors: the durable guidance from the FTC, FBI, and giving watchdogs is to route around the emotional channel entirely, giving through a charity’s own verified website rather than links in posts, checking registrations through validators, and treating urgency plus payment-by-gift-card-or-crypto as a diagnostic of fraud. Verified-organization databases, such as the 2.4 million validated nonprofits Benevity exposes and Candid’s registry data, are quietly becoming trust infrastructure for exactly this reason. Charities themselves: monitor for impersonation of your brand and your leadership, because your logo on a fake appeal is your problem regardless of authorship; publish and promote your one authentic donation route; pre-agree a rapid public response for impersonation incidents; and, at the level of practice, hold your own content to the standard the fraudsters violate. An organization that quietly uses synthetic imagery of suffering it did not witness has no principled ground to stand on when criminals do the same at scale, which is the practical argument, alongside the ethical one, for the human-verified content standards a majority of donors already say they expect.

Regulation arrives, from the EU AI Act to fundraising codes

For three years, charities used AI in a regulatory vacuum filled only by general data protection law. That period is ending, fastest in Europe, and organizations that treat compliance as a large-company problem are misreading how the new rules are written, because the central obligations attach to anyone deploying AI, nonprofit status included.

The EU AI Act, Regulation 2024/1689, entered into force on August 1, 2024 and applies in phases. Since February 2, 2025, its prohibitions on unacceptable practices apply, alongside Article 4’s AI literacy duty, which requires every organization deploying AI, explicitly including nonprofits operating in the EU, to ensure staff using these systems have adequate training and understanding. Rules for general-purpose AI models took effect August 2, 2025. The most consequential tier for charities, high-risk systems under Annex III, covers uses several nonprofits are closer to than they assume: AI evaluating eligibility for public assistance benefits, systems used in education and vocational training decisions, employment and recruitment screening, and emergency service dispatch. A charity that merely uses a chatbot to draft newsletters is nowhere near this tier; a charity whose model scores applicants for a housing program or screens job-training candidates may be squarely inside it, carrying duties around risk management, data governance, human oversight, and logging.

The timeline for that high-risk tier moved in 2026, and the movement is worth understanding precisely because half-remembered headlines are producing bad decisions. Implementation ran behind schedule, so the European Commission proposed a Digital Omnibus on AI in November 2025; after tense negotiations, the Parliament and Council reached provisional agreement on May 7, 2026, the Parliament formally endorsed it on June 16, and the Council gave final approval on June 29, 2026, with publication in the Official Journal imminent as of this writing. The substance: obligations for stand-alone Annex III high-risk systems are deferred from August 2, 2026 to December 2, 2027, and for AI embedded in regulated products to August 2028. What did not move matters just as much. The Article 50 transparency duties still apply from August 2, 2026, including the deployer-side obligations most relevant to charities: informing people when they are interacting with an AI system and disclosing AI-generated content in defined circumstances. Provider-side watermarking of synthetic content was pushed only to December 2, 2026, the same date a new prohibition on AI systems generating non-consensual intimate imagery and child sexual abuse material takes effect.

EU AI Act milestones most relevant to charities operating in Europe

The table compresses a genuinely confusing legislative sequence into the dates a charity’s board actually needs; the working summary is that disclosure duties are already arriving in 2026, while the heavy high-risk regime now lands at the end of 2027 for the eligibility-style systems some nonprofits run.

Outside the EU AI Act, the regulatory picture is a lattice rather than a single law. GDPR continues to do most of the day-to-day governing of charity AI in Europe, since nearly every interesting use touches personal data. The UK has stayed principles-based, but its sector regulators moved: the Charity Commission framed AI use as falling under existing trustee duties, and the Fundraising Regulator issued dedicated guidance in December 2025 making trustees explicitly accountable for AI in fundraising, with expectations around transparency and human oversight. In the US, no federal AI statute governs charities, but the FTC’s impersonation rule, state privacy laws, and emerging state AI acts apply, and funders are becoming de facto regulators through grant conditions about AI use and disclosure. The strategic read for charity leadership is straightforward: build the inventory of what AI you use for what, satisfy the literacy and transparency duties that are already live, and classify honestly whether anything you run touches eligibility for benefits, education, or employment, because that classification, not organization size, determines whether the December 2027 regime is your problem.

Donor trust as the decisive currency of the AI era

Strip away the technology and the charitable transaction is an act of faith: a donor hands over money and receives, in exchange, a claim about the world, that suffering was reduced, a forest protected, a child taught. Everything a charity is worth commercially, its brand, its file, its ability to raise next year, is the accumulated stock of believed claims. AI touches that stock at every point, which is why donor trust deserves treatment as a strategic variable rather than a communications afterthought.

The measured state of that trust in the AI era is conditional, not hostile, and the conditions are consistent across studies. Donors accept AI where it makes organizations demonstrably better at the mission, and majorities rate analytical and back-office uses as appropriate; the 2025 donor research found enhancing fundraising results, at 61%, edging out operational efficiency as the top perceived benefit, evidence that donors increasingly connect AI with outcomes rather than merely cost-cutting. They resist AI where it simulates the human relationship, and they punish concealment more than use. Roughly two-thirds of online donors in one 2025 survey agreed nonprofits should use AI for marketing, fundraising, and administration, while a solid third of the general donor population still says AI use would make them less likely to give, a split that reads less as contradiction than as segmentation: comfort rises with familiarity, with giving capacity, and above all with how the organization frames what it is doing.

Framing, the evidence suggests, has three load-bearing elements. The first is disclosure as default: a standing, findable statement of which tasks are AI-assisted, which are human-only, and how review works, treated like a privacy policy rather than a campaign. The second is human accountability made visible: named people signing communications, a staff member verifying anything a donor might reasonably assume a human witnessed, the assurance the Give.org research found donors explicitly seeking. The third is a story that connects AI to mission arithmetic, because donors respond differently to “we use AI” than to “AI cut our grant-reporting time in half, and those hours became forty more home visits.” The organizations that tell the second version are converting a perceived risk into a differentiator, and the roughly one in seven donors who actively reward innovation are, for them, an acquisition audience.

The trust account also has a systemic ledger this article has already touched: synthetic scams draining public belief, AI-polished appeals inflating the sector’s baseline polish, funders discounting fluent writing. Individual charities cannot fix the commons alone, but they choose, appeal by appeal, whether to replenish it or free-ride on it. Sector-level moves are beginning, responsible-AI frameworks for fundraising, validator databases, disclosure norms, and the self-interested case for joining them is simple compounding: in an environment where anyone can fake sincerity cheaply, verified integrity is the scarce asset, and scarcity is where value concentrates. The charities that will raise most easily in 2030 are the ones building that verification into their operations now, while it is still a choice rather than a requirement.

Small charities and the widening digital divide

Averages flatter this story. Sector-wide adoption figures above ninety percent conceal a distribution in which the largest organizations are pulling away from the smallest at a pace the charitable world has not honestly reckoned with. The Charity Digital Skills Report found larger nonprofits adopting AI at roughly twice the rate of small ones, 66% against 34% in its comparative analysis, and 68% of small charities describing themselves as being at an early stage with digital generally. The AI Equity Project’s data sharpens the picture into something close to a warning: around 80% of the small and midsize organizations it studied were using AI in some form, but only about 9% assessed themselves as genuinely ready for it, with policies, training, and data practices in place. Widespread use combined with narrow readiness is not democratization; it is exposure.

The constraints binding small charities are mundane and interlocking. Money comes first, with 69% of charities in the UK research naming finances as the barrier to progressing digitally, and even discounted software is a new line in a budget that has no slack. Time comes second and cuts deeper, because a three-person organization has no one to send to a course, no one to write a policy, and no evenings free to experiment; the founder who is also the fundraiser and the safeguarding lead cannot become the AI lead as well. Data readiness comes third: the predictive tools that produce the largest fundraising gains presuppose a clean, structured donor database, and many small charities run on spreadsheets, three generations of inherited CRMs, and institutional memory. Vendor discounts address the first constraint only, which is why generous pricing from Anthropic, Google, and OpenAI, welcome as it is, has not closed the gap and will not by itself.

There is also a quieter dynamic at work in the funding market. As AI-assisted grant applications drive submission volumes up, the organizations best equipped to produce polished applications at scale are the larger ones with dedicated development staff, prompting libraries, and review processes. Small charities, which the grant system already disadvantages, now compete in a pool inflated by machine-drafted bids, and funders responding with AI screening tools introduce a second machine layer that favors applicants who understand how the screening works. A technology sold as a leveler is, on the current trajectory, compounding the sector’s existing concentration, with the intermediary funding bodies best placed to notice it and, so far, only beginning to respond.

None of this argues that small organizations should abstain. The counterintuitive finding across the adoption research is that small charities often see the largest proportional returns from the least sophisticated uses, because the baseline is so constrained. An executive director who reclaims six hours a week from correspondence, minutes, and report drafting has recovered a larger share of organizational capacity than any enterprise deployment achieves. The realistic path for a small charity in 2026 is narrow and deliberately unambitious: one paid assistant subscription at nonprofit pricing, two or three named recurring tasks, a one-page policy covering data and disclosure, and a refusal to feed beneficiary information into anything until that policy exists. What closes the divide at sector scale, though, is not individual discipline but infrastructure: funders paying for readiness rather than tools, umbrella bodies producing shared templates and pooled training, and the recognition that 7.9 million digitally excluded adults in the UK alone sit on the far side of every AI-first service design decision the sector makes.

Measuring impact with AI and measuring the impact of AI

Two measurement questions run through every serious conversation about this technology, and conflating them is one of the sector’s most reliable sources of muddle. The first is whether AI improves how charities measure their impact on the world. The second is whether charities are measuring what AI itself contributes. The answers, at the moment, are cautiously yes and mostly no.

On the first question, the tooling has moved faster than practice. Impact measurement has always been the part of charitable work most starved of capacity: qualitative data piles up unread, beneficiary feedback sits in survey tools no one has time to code, and evaluation gets compressed into whatever the funder’s reporting template demands. Language models are unusually well suited to exactly this backlog. They summarize hundreds of open-text survey responses into themes in minutes, code interview transcripts against a framework, translate feedback gathered in multiple languages into a single analysis, and draft the evaluation narrative that once consumed a program officer’s week. The Charity Digital Skills data shows this frontier is still thinly populated, with 15% of charities using AI to analyze qualitative data, 13% for numeric analysis, and just 4% for predictive analytics, but the organizations working there report a distinctive benefit: they hear more of what beneficiaries actually said, because the cost of listening at scale collapsed. IDinsight’s work with language models on development data, which produced some of the largest documented productivity multiples in the sector, points at what the mature version looks like, analysis pipelines where human evaluators design the questions and audit the outputs while machines do the reading.

The discipline this requires is the same discipline the bias section of this article demanded: sampling model outputs against human coding of the same material, reporting uncertainty rather than laundering it into confident percentages, and never letting a summary of beneficiary voices replace direct contact with beneficiaries. An AI-written impact report that no beneficiary would recognize is not measurement; it is content.

On the second question, the sector’s homework is largely not done. The same 2026 adoption research that found 92% of organizations using AI found only 4% with documented workflows, and without documented workflows there is rarely a baseline, and without a baseline the claimed benefits are testimony rather than evidence. Most organizations asserting that AI saves staff hours have never timed the task before and after; most claiming better fundraising results have not separated the effect of the tool from the effect of the extra attention that accompanies any new initiative. This matters beyond internal honesty, because boards approving AI budgets, funders underwriting AI programs, and donors reading AI disclosures are all pricing claims the sector cannot currently substantiate. The fix is not elaborate. Pick the three tasks where AI is used most, record how long they took before, measure quarterly, and count errors as carefully as hours, because a tool that saves four hours and introduces one fabricated statistic into a funder report has a negative return. Organizations that run even this minimal accounting discover two useful things: which uses to expand, and which enthusiasms to quietly retire. The 7% of nonprofits reporting major impact from AI are, overwhelmingly, organizations that could show their working, and that correlation is the closest thing the adoption data offers to an instruction.

Jobs, volunteers, and the human core of charitable work

Every technology wave arrives in the nonprofit sector carrying the same anxious question, phrased in staff meetings more often than in strategy documents: which of us does this replace? The honest answer emerging from the first years of generative AI is more textured than either the doomsayers or the vendors suggest, and it turns on a distinction charities are unusually well placed to understand, the difference between tasks and relationships.

What the tools absorb first are tasks, and the absorption is real. Drafting, summarizing, translating, scheduling, data entry, first-pass research, and report formatting are precisely the activities that fill the administrative fifteen to twenty hours a week the sector’s own analyses estimate AI recovers for a typical knowledge worker. In a commercial firm, recovered hours convert to headcount decisions; in a charity, the arithmetic runs differently, because almost every organization operates far below the staffing its mission demands. The caseworker relieved of writing up notes does not become redundant, she sees another family. The development officer freed from reformatting budgets writes fewer, better applications. Across the adoption studies, redeployment rather than reduction is the dominant reported pattern, and the organizations describing layoffs attributed to AI remain rare enough to be case studies rather than a trend. The displacement risk concentrates instead in specific roles, junior communications positions, transcription and translation contractors, entry-level grant researchers, and in the pipeline effect this creates: if the tasks that once trained juniors are automated, the sector must deliberately rebuild the ladder it just removed, through apprenticeship-style review work, rotation into beneficiary-facing roles, and supervision structures that teach judgment rather than production.

Staff behavior meanwhile tells its own story about organizational culture. The Microsoft and LinkedIn workplace research finding that 52% of employees using AI are reluctant to admit it maps directly onto the nonprofit adoption data showing individual, unofficial use running far ahead of sanctioned programs. People hide tools when they fear the disclosure costs, judgment about their competence, or the reassignment of their saved hours. Charities that respond with surveillance get concealment; those that respond with amnesty, shared prompt libraries, and public celebration of time redirected to mission get the documented workflows that separate the 7% seeing major impact from everyone else.

Volunteers complicate the picture in a direction the sector has barely mapped. Volunteering is not labor supplied to fill task gaps; it is participation, the mechanism by which supporters convert sympathy into identity. An organization that automates its volunteer newsletter, thank-you notes, and shift coordination may run more smoothly while quietly dissolving the texture of belonging that made people show up. The same logic applies to the sector’s distinctive asset, moral witness. A human who visits, listens, and testifies carries a credibility no synthesis can, and donors, courts, journalists, and parliaments all price that credibility. The strategic conclusion is not to slow the automation of tasks but to be ruthless about protecting the relationships: home visits, bedside presence, community meetings, and the unhurried phone call are not inefficiencies awaiting a tool, they are the product.

Environmental cost and the ethics of compute

For charities whose missions include a habitable planet, and for the wider sector that asks donors to trust its judgment, AI arrives with an uncomfortable externality: the systems run on electricity, water, and minerals, at a scale growing faster than almost any other industrial demand. Data center electricity consumption has become a first-order concern for grid planners, with the International Energy Agency projecting global data center demand roughly doubling toward the end of the decade, AI workloads the fastest-growing component. Training frontier models consumes energy at industrial scale, and inference, the everyday querying that charities participate in, now constitutes the larger share of total AI energy use. Cooling draws water, often in water-stressed regions where hyperscale facilities cluster, and the hardware refresh cycle feeds demand for mined materials with their own supply-chain ethics.

Honesty requires proportion in both directions. An individual charity’s chatbot queries are a rounding error beside video streaming, and paralysis over prompt-level emissions is a poor use of a sustainability officer’s attention. The sector’s real exposure is threefold. First, reputational consistency: an environmental organization publicly campaigning on energy demand while silently building its operations on compute-intensive tools invites the charge of selective accounting, and the answer is disclosure rather than abstinence, a line in the annual report treating AI compute like flights and office energy. Second, procurement power: charities collectively purchase enough software that vendor selection is a lever, and asking providers for renewable-energy commitments, published efficiency figures, and water stewardship policies costs nothing and signals demand. Third, model choice as a practical discipline, because the difference between routing a routine task to a small, cheap model and a frontier one is measured in orders of magnitude of compute; the same habit that controls an API budget controls a footprint.

There is also a genuinely positive ledger that environmental charities themselves are writing. The conservation systems described earlier in this article, poaching prediction, fishing-fleet monitoring, deforestation alerts, deliver ecological returns that plausibly dwarf their compute costs, and AI-driven optimization of energy grids, logistics, and materials research sits among the technology’s strongest claims to net benefit. The mature position for the sector is neither guilt nor dismissal but bookkeeping: count the costs, count the gains, publish both, and let the mission arithmetic, not the marketing of either the boosters or the catastrophists, decide where the tools belong.

A practical adoption roadmap for organizations starting this quarter

Everything this article has surveyed, the adoption statistics, the program discounts, the humanitarian deployments, the failures, the regulation, converges on a practical question the leadership of a charity can act on this quarter: in what order should a sensibly cautious organization proceed? The sequence below is assembled from the documented practice of organizations that reached measurable results, and its defining feature is that the technology arrives in the middle, not at the beginning.

The first month belongs to inventory and policy, and skipping it is the single most common sequencing error in the sector. Inventory means finding out what is already happening, because the adoption data guarantees something is: 81% of nonprofit AI use is individual, most of it undisclosed, some of it involving data that should never have left the building. A short, blame-free staff survey, what tools, which tasks, what data goes in, surfaces the real baseline in a week. Policy then codifies the floor rather than the ceiling. A serviceable first policy fits on one page and answers five questions: which categories of data may never be entered into external AI tools, with beneficiary personal information at the top of the prohibited list; which tools are approved and paid for; who reviews AI-assisted outputs before they leave the organization; how AI use is disclosed to donors, funders, and service users; and who owns the policy. The 47% of organizations operating with no policy at all are not moving faster than everyone else, they are accumulating liabilities that surface later as breaches, funder disputes, and staff grievances.

The second month is for procurement and training, in that order of thought but the reverse order of spend. Procurement in 2026 is genuinely favorable to nonprofits, and the earlier sections of this article detailed the terms: Claude for Nonprofits at up to 75% off with Team seats landing around eight dollars per user per month, comparable OpenAI nonprofit pricing, Google Workspace nonprofit discounts of roughly 70 to 75 percent, and Microsoft’s Copilot reduction. The selection criteria that matter for a charity are less about model benchmarks than about administrative controls, whether the organization can turn off training on its data, manage users centrally, and connect the tools it already runs, and the sector-specific integrations now emerging, such as validated connectors into donor databases and grants data. Training is where the budget should actually concentrate, because the persistent finding across every study cited here is that the binding constraint is skill, not access. A third of charity CEOs and about 40% of boards rate their own AI understanding as poor, and Article 4 of the EU AI Act has, since February 2025, made staff AI literacy a legal duty for organizations deploying these systems in Europe rather than a nice-to-have. Free curricula exist, including the fluency courses bundled with the nonprofit programs, and two hours of structured practice per staff member beats any tool upgrade available at any price.

The third month is for two or three pilots, chosen by a rule that sounds unambitious and is not: pick tasks that are frequent, low-risk, internally consumed, and currently hated. Meeting minutes, first drafts of routine correspondence, summarizing long documents, reformatting reports for different audiences, translating internal materials, and coding open-text survey feedback all qualify. Beneficiary-facing systems, eligibility decisions, and crisis communication do not, and the earlier sections on Tessa, bias, and the EU’s high-risk tier explain at length why they must wait. Each pilot needs three numbers written down before it starts, time the task currently takes, error rate the current process produces, and the quality bar the output must clear, because these are the baselines that separate evidence from testimony when the board asks, three months later, whether any of this worked.

Months four through six are for measurement, pruning, and the first careful expansion. Quarterly review against the recorded baselines will typically sort the pilots into three piles: clear wins to document as standard workflows, marginal cases to adjust or extend, and disappointments to retire without sentiment. Documentation is the step that converts individual cleverness into organizational capability, and it is where the sector’s 4% figure for documented workflows shows how much low-hanging fruit remains. Expansion then follows the data rather than the demo: the organizations reporting major impact moved from internal drafting toward donor analytics, qualitative data analysis, and program-adjacent uses only after the internal layer was stable, and they moved with the governance already built, disclosure statements published, human review named, escalation paths tested.

Woven through the whole sequence are the three disciplines that the failure cases in this article keep teaching. Never let AI fabricate the organization’s knowledge, which in practice means every statistic, name, and citation in an AI-assisted draft gets verified by a human who would stake their signature on it. Never let AI touch beneficiary data before the policy, the contracts, and the security review exist, however inconvenient that ordering feels in the enthusiasm of a pilot. And never conceal, because every strand of donor research converges on the same asymmetry: use is largely forgiven, concealment is not, and disclosure done confidently reads as competence rather than confession. An organization that follows this sequence spends its first quarter looking slower than its peers and its second year looking inexplicably further ahead, which is a fair one-sentence summary of how the 7% got there.

Strategic outlook, agentic AI, and the open questions ahead

The tools this article has described are already yesterday’s frontier. The direction of travel for 2026 and beyond is agentic: systems that do not merely draft when asked but plan, execute multi-step work, operate software, and act with delegated authority. The nonprofit editions of the major platforms now bundle agentic coding and workflow tools, and the first charity deployments, agents that monitor grant portals and assemble application drafts, reconcile donations across systems overnight, or triage inbound casework email into structured queues, are moving from demonstrations into production at the better-resourced organizations. The efficiency ceiling of assistance is the productivity of one helped human; the ceiling of delegation is set by how much work an organization can safely hand over, which is why the agentic turn will widen, not narrow, the gap between charities with governance and charities without it. An agent with access to a donor database and an email system is an operational risk category the sector’s one-page policies were not written for, and the second generation of nonprofit AI policy, covering delegated authority, spending limits, action logs, and kill switches, will need writing years before most boards expect to need it.

Several open questions will decide whether this technology’s charitable ledger ends net positive, and none of them are engineering questions. The first is whether funders modernize. Grantmaking currently sits on both sides of the AI arms race, receiving machine-drafted applications in rising volume while deploying machine screening in response, and the exit from that spiral is structural: shorter applications, shared data standards, funding for readiness and data infrastructure rather than tools alone, and honest published positions on AI-assisted bids to replace the current ambiguity in which two-thirds of foundations remain undecided. The second is whether the sector builds shared infrastructure or fragments into thousands of duplicate experiments. The precedents are encouraging where they exist, open damage-assessment models in humanitarian response, shared conservation platforms, pooled validator databases for fundraising, and the economics are obvious: no midsize charity should be solving prompt injection or model evaluation alone. The third is whether beneficiaries get a voice in systems built about them, because the sector’s legitimacy rests on representing people, and an AI layer designed entirely between charities, vendors, and funders reproduces the paternalism the last two decades of development practice tried to unlearn.

The fourth question is the deepest, and it returns this article to where it began. Charitable work exists because markets and states leave human needs unmet, and its distinctive resources have never been operational, they are trust, witness, and the willingness of strangers to fund care for people they will never meet. Artificial intelligence, honestly deployed, extends those resources further than any technology the sector has adopted: it reads every survey response, answers at three in the morning, finds the eligible family the spreadsheet missed, watches every hectare of forest at once. Dishonestly or carelessly deployed, it counterfeits the sector’s currency, synthetic sincerity, fabricated evidence, automated exclusion wearing the language of care. The 92% adoption figure says the sector has chosen the tools; the 7% impact figure says it has barely begun choosing how to use them well. That second choice is still open, it is being made now, one policy, one pilot, and one disclosure at a time, and the organizations making it deliberately are quietly building the version of this future in which the machines do the reading and the humans, finally, have time to do the visiting.

AI in advocacy, campaigning, and public persuasion

One arena of charitable work has gone conspicuously underdiscussed in the sector’s AI conversation, and it is the one where the stakes are most political: advocacy. A large share of nonprofits exist not to deliver services but to change rules, budgets, and public minds, and generative AI has quietly rewritten the economics of that work in both directions at once.

On the capability side, the gains are concrete and already in use. Policy teams that once assigned a week to reading a three-hundred-page consultation document now extract the clauses touching their beneficiaries in an afternoon, with the caveat, by now familiar, that every extracted claim gets checked against the source before it reaches a press release. Legislative monitoring, once a subscription service only the largest organizations afforded, is reproducible by a small charity with a language model and a list of keywords across parliamentary records and regulatory registers. Campaign teams analyze thousands of petition comments or supporter messages to understand which arguments actually move their base, draft testimony tailored to individual committee members’ stated concerns, and localize a national campaign into a dozen community variants in the time one version used to take. For under-resourced groups facing well-funded opposition, this is the rare technology that narrows rather than widens the asymmetry of policy combat.

The same machinery, pointed the other way, threatens the currency advocacy runs on, which is the belief that expressed public opinion is real. Mass-generated consultation responses, synthetic grassroots letters, and coordinated AI-written comment campaigns are cheap enough that regulators and legislatures are already discounting high-volume input, a rational response with a brutal side effect: the more astroturf machines produce, the less any petition signature or public comment is worth, including the genuine ones charities spent decades learning to mobilize. The episodes of fabricated mass comments in regulatory proceedings that predate modern language models look, in retrospect, like a rehearsal. Advocacy charities therefore face a sharper version of the sector’s general disclosure question. A campaign that uses AI to draft template letters its supporters knowingly personalize sits on the defensible side of a line; one that generates volume designed to impersonate spontaneous public sentiment sits on the other, and the reputational and, increasingly, legal consequences of crossing it land hardest on organizations whose entire asset is credibility. The practical stance emerging among the more careful campaigning groups is to use AI heavily for analysis and drafting, sparingly and transparently for supporter-facing mobilization, and never for manufacturing the appearance of humans who do not exist.

Building internal capability and the nonprofit talent question

Beneath every roadmap and policy this article has proposed sits a resourcing question the sector answers badly by default: whose job is this? The prevailing arrangement, visible in the adoption statistics, is that AI belongs to whoever happens to be enthusiastic, which is how organizations end up with 81% individual use, 4% documented workflows, and a single point of failure who leaves for a better-paid job. Building durable capability is a talent problem before it is a technology problem, and the sector has more options than its budget anxieties suggest.

The first option is naming ownership without creating an empire. Most charities below a few hundred staff do not need a head of AI; they need an existing senior operator, often the director of operations or finance, formally assigned the policy, the vendor relationships, and the quarterly measurement, with a few hours a month protected for it. What the role needs is authority and curiosity, not an engineering background, and treating it as an add-on to a junior communications post, a common pattern, guarantees the governance stays decorative. Boards carry a parallel duty: with roughly two in five rating their own AI understanding as poor, trustee recruitment and training are governance interventions as surely as any policy, and regulators in the UK have now said in terms that accountability for AI use sits with trustees, not with whoever set up the account.

The second option is borrowing expertise the sector already organizes. Intermediaries such as NTEN, TechSoup, and NetHope run training, peer groups, and procurement guidance built specifically for nonprofit constraints, and skilled-volunteering channels connect data scientists and engineers to charitable projects that would never afford them commercially, an arrangement that works precisely when the charity arrives with a defined problem and an owner, and fails when it arrives hoping the volunteers will find one. Fractional and shared roles, one data lead serving a consortium of small organizations, are spreading for the same reason shared finance functions did a generation ago. And the vendor programs themselves now bundle structured curricula, the fluency courses attached to the nonprofit editions, which cost nothing except the discipline to schedule them.

The third option is growing capability from the inside, which the hidden-use statistics suggest is less about teaching than about legitimizing. Every organization already employs people who quietly became competent with these tools on their own time; the capability program that works fastest is often an amnesty, a shared prompt library, an hour a month where staff show each other what they have automated, and public credit for time redirected to mission. The scarce resource is not talent but permission, and organizations that grant it convert their shadow adopters into the trainers, policy reviewers, and workflow documenters the roadmap requires, at a price no consultancy will match.

A donor’s toolkit for judging charitable AI claims

Most of this article has faced the organizations. Its final analytical section faces the people who fund them, because donors are not spectators to the sector’s AI transition, they are its regulators of last resort, and the questions they ask shape behavior faster than any statute. A donor does not need technical literacy to evaluate a charity’s AI posture; they need perhaps six questions and a feel for what good answers sound like.

The first question is simply whether the organization will say what it uses AI for. A charity with a findable disclosure statement, even two paragraphs on a policies page, has done more governance than roughly half the sector, and the tone of the answer is diagnostic: specific and unembarrassed is the mark of an organization in control of its tools, while evasive or defensive suggests the use is real but unmanaged. The second question concerns beneficiary data, and it has a single acceptable shape of answer: personal information about the people served does not go into external AI tools, or goes only under contracts that prohibit training on it, with named exceptions and named accountability. Any answer that wanders is a warning. The third is about human review, who checks AI-assisted material before it reaches donors, funders, or service users, and the strong answer includes a person’s role, not a process abstraction.

The fourth question tests measurement: what has AI actually changed here? Organizations in the sector’s productive minority answer with before-and-after specifics, hours redirected, response times cut, more feedback analyzed, while organizations riding the hype answer with adjectives. A donor who hears that machine-recovered time became more home visits, faster benefit applications, or extra caseload capacity is hearing the mission arithmetic this article has argued is the entire point. The fifth question is about images and stories, whether what the donor is shown depicts real people and real events, and after the documented rise of synthetic disaster appeals and fabricated beneficiary photographs, no legitimate organization should bristle at being asked. The sixth is the quiet one, best judged rather than asked: does the charity’s AI enthusiasm sit in service of a strategy, or in place of one? The research finding that digital strategy declined even as AI adoption soared describes a sector at some risk of mistaking tools for direction, and donors, reading annual reports and appeals, are well placed to notice which organizations kept their bearings.

None of this asks donors to become auditors, and none of it should tip into reflexive suspicion of any charity that mentions AI, since the evidence throughout this article is that the technology, governed well, is among the better efficiency stories the sector has had in decades. The point is narrower and more powerful: charities respond to what their supporters visibly value. A funding public that calmly, routinely asks about disclosure, data, and human oversight will produce a sector that has answers, and it will do so years before regulation reaches the same destination. Donors built the nonprofit world’s culture of financial accountability by asking about overheads and audits until answering became standard; the same patient pressure, applied to algorithms, is how the sector’s AI era gets its guardrails.

Questions charities and donors keep asking about AI in philanthropy

Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

The 2026 Nonprofit AI Adoption Report Survey of 346 nonprofit organizations by Virtuous and Fundraising.AI, published December 2025, documenting 92% AI adoption, the 7% major-impact minority, and the gap between individual use and organizational workflows.

Nonprofit AI adoption hits 92%, but only 7% see major impact NonProfit PRO’s analysis of the 2026 adoption report, detailing the reactive, individual-level pattern of AI use across the sector.

Charity Digital Skills Report Annual UK benchmark of 672 charities tracking AI adoption rising from 35% in 2023 to 76% in 2025, alongside data on skills gaps, policy development, and the small-charity divide.

Substantial growth in AI adoption as three quarters of charities now use it Civil Society Media’s coverage of the 2025 Charity Digital Skills findings, including the decline in charities holding a digital strategy.

Surprising insights from the Charity Digital Skills Report 2025 ICAEW’s analysis of the report’s governance findings, including board and CEO self-assessed AI skill levels.

The hidden use of AI in charities Charity Digital’s examination of undisclosed staff AI use and its implications for policy and trust.

Claude for Nonprofits Anthropic’s December 2025 launch announcement covering sector discounts of up to 75%, data connectors to Blackbaud, Benevity, and Candid, and the AI Fluency for Nonprofits curriculum.

Claude for Nonprofits solutions page Program details, eligibility, and pricing for Anthropic’s nonprofit offering, including Team and Enterprise tiers.

Getting started with Claude for Nonprofits Anthropic’s implementation documentation covering verification through Goodstack and onboarding steps for charitable organizations.

Giving Tuesday: Anthropic offers nonprofits discounts NBC News reporting on the launch, with pilot results from Robin Hood, Tipping Point, IDinsight, and the International Rescue Committee.

People-First AI Fund OpenAI’s philanthropic fund, which distributed $40.5 million in unrestricted grants to 208 US nonprofits in December 2025.

Introducing OpenAI for Nonprofits OpenAI’s nonprofit program page covering discounted ChatGPT business tiers and eligibility.

OpenAI Nonprofit Jam Documentation of OpenAI’s ten-city training event for more than a thousand nonprofit leaders, with twelve months of free ChatGPT Plus for participants.

Google.org Generative AI Accelerator 2025 cohort Announcement of the $30 million cohort of twenty nonprofits receiving funding and six months of technical support.

Google.org Generative AI Accelerator Program page describing grant ranges, technical assistance, and selection criteria for nonprofit AI projects.

Google.org AI for Science funding Details of the $30 million science-focused funding opportunity that closed to applications in May 2026.

Machine learning and phone data can improve targeting of humanitarian aid The peer-reviewed Nature study documenting Togo’s Novissi program, including the 8 to 14 percentage point reduction in exclusion errors.

GiveDirectly MobileAid GiveDirectly’s program page on machine-learning-targeted cash transfers, covering deployments in Togo, Bangladesh, Malawi, and Kenya.

AI targeting in Bangladesh Results of the 2023-2024 Bangladesh deployment, where AI targeting matched survey-based methods at lower cost.

SKAI, the WFP Innovation Accelerator The World Food Programme’s page on its open-source satellite damage-assessment system built with Google Research.

AI from Google Research and UN boosts humanitarian disaster response ReliefWeb reporting on SKAI deployments including the Myanmar earthquake and Sri Lanka cyclone assessments.

xBD: A dataset for assessing building damage from satellite imagery The research paper describing the open dataset of over 850,000 annotated buildings underpinning modern damage-assessment models.

PAWS: artificial intelligence fights poaching National Geographic’s account of the Protection Assistant for Wildlife Security and its use in routing anti-poaching patrols.

Patrick J. McGovern Foundation 2024 grants Announcement of $73.5 million in grants to 144 organizations across eleven countries for AI and data capability.

Donor Perceptions of AI 2025 Survey of 1,031 donors conducted in August 2025, finding conditional acceptance of nonprofit AI use and identifying fundraising results as the top perceived benefit.

Donor perceptions of AI Fidelity Charitable’s donor research on comfort levels with AI across nonprofit functions and the transparency conditions attached.

The new currency of fundraising: trust in the age of AI The Association of Fundraising Professionals on disclosure, authenticity, and maintaining donor confidence as AI use spreads.

AI marketing and fundraising statistics for nonprofits Compiled sector statistics including online donor attitudes toward nonprofit AI use across marketing and administration.

EU AI Act omnibus agreement: postponed high-risk deadlines and other key changes Legal analysis of the 2026 Digital Omnibus, detailing the deferral of Annex III obligations to December 2027 and the transparency duties that remain on schedule.

EU agrees Digital Omnibus deal to simplify AI rules White and Case’s briefing on the legislative sequence and substance of the AI Act amendments agreed in 2026.

Charity and disaster fraud FBI guidance on fraudulent charitable appeals, including warnings about AI-generated content in disaster-related scams.

FTC proposes new protections to combat AI impersonation The Federal Trade Commission’s rulemaking on impersonation, relevant to synthetic charity scams and deepfaked appeals.