The AI Incident Database (AIID) is a public, curated archive of real-world events in which the deployment of an artificial intelligence system caused or nearly caused harm. In AI ethics, it serves as the collective memory of AI failures: each entry links an incident to the news articles, reports and analyses that describe it, so researchers, developers and policymakers can study what went wrong, to whom and why.
The project was launched in 2020 under the Partnership on AI and is now maintained by the Responsible AI Collaborative, an independent nonprofit. Its design borrows directly from aviation and cybersecurity safety practice, where shared incident records have long been the basis for preventing the same failure from happening twice.
How the AI Incident Database records harms
The AI Incident Database works by separating incidents from reports. An incident is a single real-world event, such as a facial recognition system leading to a wrongful arrest, and it receives a permanent incident number. A report is any published account of that event; one incident may collect dozens of reports from journalists, courts, regulators and academics over several years.
Anyone can submit a report through a public form, but entries do not appear automatically. Human editors review each submission, decide whether it describes a new incident or adds to an existing one, and write a short neutral description. This editorial review step keeps duplicates down and protects the archive from speculative or promotional claims.
The following table summarizes the main record types and what each one contributes to the archive.
| Record type | What it represents | Typical content |
|---|---|---|
| Incident | One real-world harm event | Number, date, description, alleged deployer and developer |
| Report | One published account of an incident | Article title, source, date, text snapshot |
| Variant | A near-identical repeat of a known failure | Input, output and link to the parent incident |
| Classification | Structured tags applied by analysts | Harm type, sector, AI technique, severity |
| Response | A statement from an involved party | Company or agency reply linked to the incident |
Each incident also names the alleged deployer, the alleged developer and the alleged harmed parties. The word “alleged” is deliberate: the database documents what credible sources report, and it does not issue legal findings of fault.
Taxonomies that make incidents comparable
Raw incident narratives become useful for analysis only when they are classified consistently. The archive therefore applies several independent taxonomies, each maintained by a different team with its own questions and definitions. A single incident can carry tags from more than one taxonomy at the same time.
| Taxonomy | Main question it answers | Example attributes |
|---|---|---|
| CSET taxonomy | What harm occurred and to whom | Harm domain, tangible harm, affected group |
| GMF taxonomy | Why the system failed technically | Goals, methods, known failure modes |
| MIT risk taxonomy | Which risk category the event fits | Discrimination, misinformation, safety, privacy |
| Editor tags | Quick navigation and search | Sector, technology, country |
The Goals, Methods and Failures (GMF) taxonomy is especially valuable for engineers because it links each event to a known technical failure mode, such as distributional shift, adversarial input or underspecified training data. The CSET work, by contrast, serves policy analysis by asking whether a tangible harm occurred and which population bore it.
How it compares with other incident sources
The AI Incident Database is one of several sources that track AI harms, and each has a different scope, curation model and legal status. Understanding the differences prevents analysts from treating one archive as a complete census.
| Source | Curation model | Scope | Legal status |
|---|---|---|---|
| AI Incident Database | Crowdsourced, editor-reviewed | Global, all sectors | Voluntary, public |
| OECD AI Incidents Monitor | Automated news analysis | Global media coverage | Voluntary, public |
| EU AI Act incident reports | Provider submissions to authorities | Serious incidents from high-risk systems | Mandatory, mostly confidential |
| CVE records | Coordinated vulnerability disclosure | Software security flaws | Voluntary, public |
| Internal incident logs | Company post-incident reviews | One organization’s systems | Private |
The key distinction is voluntary public documentation versus mandatory regulatory reporting. Under the EU AI Act, providers of high-risk systems must report serious incidents to market surveillance authorities, but those filings are not designed for public learning. The AIID fills the gap by making publicly reported events searchable for everyone.
Practical uses in AI governance and engineering
Practitioners use the archive mainly as an evidence base for anticipating harm. Before deploying a model, a team can search for incidents involving the same technique, sector or user population and turn past failures into concrete test cases and risk controls.
| User group | Typical use | Resulting output |
|---|---|---|
| Product teams | Search incidents for similar systems | Pre-deployment test scenarios |
| Auditors | Benchmark known failure patterns | AI audit findings and risk ratings |
| Policymakers | Identify recurring harm categories | Evidence for rules and guidance |
| Researchers | Analyze trends across incidents | Studies of bias, safety and misuse |
| Educators | Select documented case studies | Course material on responsible AI |
The archive fits naturally into an algorithmic impact assessment: assessors list foreseeable harms, and incidents from comparable deployments show which harms are realistic rather than hypothetical. It also supports AI accountability, because a documented public record makes it harder for organizations to treat a repeated failure as unforeseeable.
Limitations and common misreadings
The most important limitation is reporting bias. The archive captures only events that someone noticed and wrote about, so it overrepresents high-profile consumer products and English-language media, while harms in closed systems such as hiring or credit decisions often remain invisible.
- Counts are not prevalence rates. A rising number of incidents may reflect more media attention, not more failures.
- Allegations are not verdicts. Entries summarize credible reports and do not establish legal liability.
- Definitions remain contested. What counts as one incident versus several, and when a near miss qualifies, involves editorial judgment.
- Coverage lags events. Harms discovered years later, for example through litigation, enter the archive late.
Used carefully, the database is best treated as a qualitative library of failure patterns combined with rough trend signals, not as a statistical sample of all AI harm.
Put simply
The AI Incident Database is a shared notebook of AI mistakes that hurt people. When an AI system causes a problem in the real world, such as wrongly flagging someone as a criminal or giving dangerous advice, the event is written down, given a number and linked to the news stories about it.
Airlines learn from every crash and near miss so that the same error does not happen again. This archive does the same for AI. The key idea is simple: learn from past failures before building the next system.
In practice
A regional bank plans to introduce an automated loan screening model and asks its responsible AI team to review the design. Before testing begins, the team searches the AI Incident Database for incidents involving credit scoring, automated eligibility decisions and proxy discrimination.
The search returns a cluster of earlier cases in which lenders and public agencies used models that disadvantaged applicants by postal code or gender through correlated features. The team converts six of those patterns into test scenarios, checks approval rates across groups with the four-fifths rule and removes two features that acted as proxies for protected attributes.
The final launch documentation cites the reviewed incidents as known risks and lists the controls that address each one. The result is a model that passed internal audit on the first review, with a clear record showing that foreseeable harms were considered before deployment rather than after a complaint.