OpenAI’s latest ambition is larger than a smarter chatbot and more consequential than a browser agent. According to recent reporting based on comments from chief scientist Jakub Pachocki, the company now treats a fully automated AI researcher as its “North Star” for the next few years: first an autonomous research intern, then a broader multi-agent system that can stay with a problem long enough to plan, test, revise, and keep going without constant human steering.
Table of Contents
That framing matters. A chatbot answers. An agent executes. A researcher has to sustain a line of thought. It has to decide what matters, separate weak evidence from strong evidence, notice when an approach is failing, and change direction before it wastes time. That is a very different standard from producing a polished reply in a chat window. OpenAI’s own product releases over the past year suggest the company has been building toward exactly that transition.
The timing is not accidental. Reuters reported that OpenAI has been redrawing parts of its roadmap under pressure from rivals and is trying to focus more tightly on products that can turn model capability into durable utility. An automated researcher fits that moment almost perfectly. It sounds visionary, but it also gives the company a concrete answer to a difficult question: after chatbots and first-generation agents, what is the next product category big enough to justify the scale of frontier AI?
A new destination after chatbots and agents
The easiest way to understand this bet is to look at the ladder OpenAI has already climbed.
From assistant to agent to researcher
| Stage | Core behavior | Human role |
|---|---|---|
| Chatbot | Answers, explains, drafts, summarizes | Directs each step |
| Agent | Browses, researches, uses tools, completes bounded tasks | Delegates a workflow |
| Automated researcher | Plans, investigates, tests, revises, persists over longer horizons | Supervises goals and verification |
That progression is more than branding. OpenAI’s deep research product already performs multi-step internet research and assembles documented reports with citations. Operator introduced computer use through a browser, using the same kind of interface a person sees. ChatGPT agent then combined research and action into one system that can browse, use tools, analyze files, and complete tasks with its own computer while still asking for confirmation on sensitive actions. The roadmap described in recent reporting extends that same arc outward, from “finish this task” to “own this problem.”
The pieces OpenAI already has on the table
The strongest reason to take this project seriously is that OpenAI is not starting from zero. It already has several technical pieces that look like precursors.
Deep research, launched in ChatGPT and exposed in the API, is designed to find, analyze, and synthesize hundreds of sources into a long, structured report. OpenAI describes it as an agentic capability for complex analysis, one that can use web search, files, and connected data sources. That is not autonomous science, but it is already a form of extended, plan-based knowledge work that goes well beyond prompt-and-response chat.
Operator added a different ingredient. It gave the model a browser, screenshots, a cursor, and a keyboard. OpenAI’s Computer-Using Agent system is built to perceive a graphical interface, take actions, recover from mistakes, and hand control back when a task reaches a sensitive step such as login or payment. That is the scaffolding for agency in the digital world: the ability to move from reasoning about a task to acting inside the software where the task actually lives.
ChatGPT agent is the clearest sign of convergence. OpenAI says it combines deep research’s multi-step investigation with Operator’s browser action and adds terminal-style execution for code, data analysis, and document generation. The company also says users can interrupt, redirect, or take over at any point. In other words, OpenAI is already building systems that reason, browse, act, pause, resume, and collaborate over time. A fully automated researcher would be the next jump in duration, autonomy, and internal coordination.
Research is the toughest useful job
Why aim at research rather than a dozen smaller agent products? Because research is where the value compounds.
OpenAI’s January 2026 science report makes the case in unusually direct terms. It argues that many fields are suffering from falling research productivity: more people, more time, and more money are needed to produce each additional insight. It points to semiconductors, where sustaining progress now requires far more researchers than it did decades ago, and to drug development, where moving from target discovery to approval in the United States still takes roughly 10 to 15 years on average. If AI can compress even part of that cycle, the payoff is not incremental.
OpenAI’s own usage data suggests the demand is already there. In that same report, the company says ChatGPT sees almost 8.4 million weekly messages on advanced science and mathematics topics from roughly 1.3 million weekly users worldwide, with advanced science and math usage rising sharply through 2025. The point is easy to miss because it hides behind a product metric, but it matters: researchers are not waiting for a perfect autonomous scientist. They are already using today’s tools anywhere the tools shave time off reading, coding, analysis, simulation, and experiment planning.
There is also a broader industry signal here. Nature published a peer-reviewed paper on The AI Scientist, an agentic system that generates ideas, writes code, runs experiments, analyzes results, drafts a manuscript, and performs automated review in machine-learning research. One AI-generated paper cleared the first round of peer review at an ICLR workshop, though the paper is careful about the limits: this was a workshop rather than the main conference, and the domain was machine learning, where experiments can stay entirely inside the computer. That is still enough to show that end-to-end automation is no longer science fiction.
A short roadmap with very large consequences
The reported roadmap is striking partly because it is so compressed. Recent reporting says OpenAI wants an autonomous AI research intern first, with a larger multi-agent researcher targeted for 2028. That is aggressive, though not out of line with OpenAI’s own public writing late last year. In November 2025, the company wrote that current systems already seem “more like 80% of the way to an AI researcher than 20%”, that AI has progressed from seconds-long tasks to tasks taking more than an hour, and that systems able to do days- or weeks-long work are expected soon. The same post says OpenAI expects AI to make “very small discoveries” in 2026 and more significant ones in 2028 and beyond.
That is why the “researcher” idea is more than a lab fantasy. It gives OpenAI a way to unify several strands of work that otherwise look separate: reasoning models, coding systems, browser agents, interpretability, safety tooling, and high-compute orchestration. Reuters’ reporting on the company’s recent refocus helps explain the business logic. OpenAI has been pushed to prioritize clearer product direction under competitive pressure. A durable AI researcher is both a technical objective and a product thesis.
If this works, the most obvious first impact will not be a machine replacing principal investigators. It will be a system that absorbs the long, grinding middle of research work: literature triage, hypothesis expansion, code iteration, replication attempts, benchmark design, data cleaning, exploratory analysis, documentation, and endless reformulation. Human researchers would still choose the questions and judge the stakes, but the tempo of the work could change dramatically.
Safety will decide whether the idea is credible
The hard part is not only capability. It is trust.
OpenAI’s own agent documentation is blunt about the hazards. ChatGPT agent and related agent-building materials warn about prompt injection, private data leakage, tool misuse, mistaken actions, and ambiguous instructions. The company says it uses explicit confirmations for consequential actions, watch-mode style supervision on certain sites, refusal behavior for disallowed tasks, and approval flows for tool use. That is a strong sign that OpenAI understands the problem. It is also a sign that the problem is very real.
The safety posture becomes even more revealing at the frontier. In its announcement for ChatGPT agent, OpenAI said it was treating the system as High Biological and Chemical capabilities under its Preparedness Framework and applying associated safeguards. Separately, the company’s safety materials stress that no superintelligent system should be deployed without robust alignment and control. Those are not throwaway caveats. They amount to an admission that greater autonomy multiplies both usefulness and risk at the same time.
This is where an automated researcher becomes qualitatively different from a chatbot. A chatbot can be wrong in a contained way. A researcher-like system can be wrong across a chain of decisions, each one reinforcing the last. The risk is not only hallucination in the old sense. It is confident procedural drift: the model pursues a plausible line, generates polished intermediate work, and buries the mistake deep enough that nobody notices until the conclusion looks persuasive. The Nature paper on AI Scientist and OpenAI’s own safety guidance both point toward the same lesson: validation, provenance, and human review have to scale with autonomy or the system will produce believable junk faster than institutions can absorb it.
The real meaning of OpenAI’s North Star
The most interesting part of this story is not the headline promise. It is the shift in what OpenAI appears to think counts as a meaningful frontier product.
For years, the industry taught people to think in terms of models. Bigger model, faster model, cheaper model, better benchmark score. The automated researcher reframes the contest around sustained cognitive labor. Can an AI stay coherent over hours or days? Can it decide what to do next without being spoon-fed? Can it decompose an open problem, work through dead ends, and return something a serious person would actually use?
That is a much harder bar than “impressive demo.” It is also a more honest one. Research exposes almost every weakness these systems still have: shaky judgment, brittle memory, overconfidence, weak falsification, poor handling of hidden assumptions. If OpenAI can build a system that holds up there, it will have built something more important than another assistant feature.
So the real significance of this project is simple. OpenAI is no longer just trying to make AI more helpful inside a conversation. It is trying to make AI independently productive across an entire line of inquiry. That ambition could accelerate science, software, finance, policy work, and any field whose core materials are text, code, models, or diagrams. It could also flood those same fields with faster error if the verification layer lags behind.
That tension is why this feels like a genuine turning point. The next era of AI will not be decided by who can talk the best. It will be decided by who can work the longest, steer the cleanest, and stay the most accountable while doing it.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
OpenAI is building fully automated AI researcher, says it is top priority project
Reporting on OpenAI’s reported “North Star” goal, citing comments from chief scientist Jakub Pachocki and the company’s plan for an autonomous AI researcher.
https://www.indiatoday.in/technology/news/story/openai-is-building-fully-automated-ai-researcher-called-north-star-2885120-2026-03-21
OpenAI shifts focus to building fully autonomous AI researcher, sets 2028 target
Industry coverage summarizing the reported roadmap from an autonomous research intern to a larger multi-agent research system.
https://enterpriseai.economictimes.indiatimes.com/news/industry/openai-pursues-fully-autonomous-ai-researcher-by-2028-revolutionizing-scientific-exploration/129777645
Introducing deep research
Official OpenAI product announcement describing deep research as a multi-step research agent that can synthesize hundreds of sources into documented reports.
https://openai.com/index/introducing-deep-research/
Introducing Operator
Official OpenAI announcement for Operator, explaining browser-based task execution, takeover mode, safety controls, and the move from passive assistant to active agent.
https://openai.com/index/introducing-operator/
Introducing ChatGPT agent
Official OpenAI announcement describing ChatGPT agent as the combination of research and action, with its own computer, tool use, supervision controls, and risk mitigations.
https://openai.com/index/introducing-chatgpt-agent/
AI as a Scientific Collaborator
OpenAI report on scientific usage of ChatGPT, research productivity bottlenecks, and the current role of AI in literature synthesis, coding, analysis, and experiment planning.
https://cdn.openai.com/pdf/f4b4a5da-b2de-418d-9fcd-6b293e9dc157/oai_ai-as-a-scientific-collaborator_jan-2026.pdf
AI progress and recommendations
OpenAI’s policy and research outlook, including its claims about proximity to AI-researcher-like systems and expected discovery capabilities in 2026 and 2028.
https://openai.com/index/ai-progress-and-recommendations/
Artificial Intelligencer OpenAI’s $852 billion problem finding focus
Reuters analysis of OpenAI’s strategic refocus under competitive pressure and its push toward tighter product direction.
https://www.reuters.com/technology/artificial-intelligence/artificial-intelligencer-openais-852-billion-problem-finding-focus-2026-04-01/
Towards end-to-end automation of AI research
Nature paper describing The AI Scientist, an autonomous research pipeline spanning idea generation, experimentation, manuscript writing, and automated review.
https://www.nature.com/articles/s41586-026-10265-5
