Research

Scientists Use AI for Tasks, Not Autonomous Research, OpenAI Finds

OpenAI's study of 1,819 academic scientists found end-to-end autonomous research requests at 0.4% of messages, heavy output checking, and 50,000 new free researcher grants.

OpenAI Study Finds Scientists Rarely Request End-to-End AI Research
OpenAI Study Finds Scientists Rarely Request End-to-End AI ResearchStockSnap / pixabay
By Sophie Lindqvist5 min read

Why it matters

  • End-to-end autonomous research made up 0.4% of scientists' AI messages, though 15.3% tried it at least once.
  • 45.4% of surveyed scientists cited checking or correcting AI outputs as their main difficulty.
  • OpenAI opens applications for 50,000 free researcher grants on November 23, 2026, first-come, first-served.
  • 98% of 250 surveyed scientists expected AI to accelerate discovery in their field.
  • Reading (27.7%) and writing (26.9%) dominated scientific AI use.

Only 0.4% of scientists' AI messages ask for end-to-end autonomous research, according to an OpenAI study published November 16, 2026. The paper analyzed 580,186 scientific messages from 1,819 academic scientists using ChatGPT and Codex, plus survey responses from 251 participants, and it lands at a moment when labs, funders and policymakers are debating how much of the scientific process AI should take over. The short answer from the data: scientists are handing AI discrete tasks, not the keys to the whole experiment.

The study covers activity from August 20 through September 22, 2026, and a survey run from September 30 to October 7, 2026. OpenAI's researchers used privacy-preserving classifiers to sort messages into categories without reading the underlying exchanges, and program data were not used for model training by default. That method matters — it lets OpenAI describe behavior at scale while limiting intrusion into scientists' actual work.

What are scientists actually asking AI to do?

Reading and knowledge acquisition led the pack at 27.7% of the average scientist's scientific messages. Writing and communication came close behind at 26.9%. Scientific analysis contributed 16.7%, and research software and code development 13.8%. Together, those four workflows accounted for roughly 85% of all scientific use.

That concentration tells you where AI sits in research today: discrete knowledge work, not coordinated projects. Explaining scientific concepts was the single most common task, at 18.8% of messages, and 89.7% of scientists sent at least three such requests. In plain terms, scientists mostly use these tools the way everyone else does — to read, understand and write.

Ideation looks different when you measure breadth rather than volume. It represented only 5.3% of messages, but about 68% of scientists sent at least three ideation messages during the study window. Within those messages, scientists supplied a specific question 35% of the time, asked ChatGPT to refine an existing idea 36% of the time, and asked it to suggest a research question 23% of the time. The remaining 6% couldn't be categorized.

The paper also finds a stage-dependent division of labor. Scientists often asked AI to propose directions during early exploration, then brought their own questions and ideas for refinement once a research problem sharpened. AI suggests; the scientist narrows.

How much do scientists check the AI's work?

Quite a lot, and that number is the study's most useful signal about trust. The paper's scientist-involvement score — the share of initial requests followed by clarification, checking or revision — averaged 31.8%. Writing and communication drew the highest involvement at 38.0%, followed by scientific analysis at 34.4% and research software at 33.7%.

The pattern gets sharper when you split by difficulty. In 28 of 32 comparable cases, requests requiring advanced or frontier expertise saw higher involvement, by an average of 6.5 percentage points. Harder work means more follow-up from the human, not less.

The score has limits worth noting. It counts only the next scientific message in the same conversation within seven days, so it misses revisions scientists make outside the AI application. Checking was also the top reported friction point: 45.4% of surveyed scientists selected correcting or checking outputs as their main difficulty in getting value from ChatGPT.

Takeaway: scientists are saving time on tasks while spending real effort verifying what the model produces.

Do the time savings hold up?

Scientists themselves think so, at least by self-report. Across eight research activities, surveyed scientists estimated that one hour of work with AI would have taken about 1.9 hours without it at the same quality. The ratios ranged from 1.36 times for administration to 2.85 times for research software and code development — the same coding category that showed heavy code use in the message data.

Expectations for discovery run high. Of the 250 scientists who answered the discovery-timeline question, 98% expected AI to accelerate discovery in their field. Asked to imagine a discovery that would take ten years without AI, 72.0% expected it at least three years sooner with AI, and 21.6% expected it within three years.

But when scientists explained their reasoning in writing, execution — not automation — dominated. Among 367 reasons identified across 224 responses, 52% concerned faster research execution. Automating research loops accounted for just 2.2%. Another 23% concerned finding more promising research questions, and 14% pursuing a wider range of questions.

What still blocks scientists?

The bottlenecks extend well beyond model capability, and this is where the study connects to policy. Limited time or staffing was selected by 51.8% of respondents, and limited funding by 43.8%. Validating or reproducing findings affected 25.5%, and access to equipment, laboratories or computing resources affected 24.3%. Faster AI doesn't fix a missing postdoc or an unfunded lab.

OpenAI itself flags the caveats. Participants were a selected group — they applied for expanded AI access and chose to answer an AI survey — so the results may not generalize to all scientists. The usage shares describe requests, not research value or time allocation, and the findings don't establish causal effects.

OpenAI is also handing out 50,000 grants

Alongside the study, OpenAI expanded its ChatGPT for Academic Researchers program. The pilot, running since July 2026, reached 20,000 researchers across 33 countries. Applications for 50,000 new grants open November 23, 2026, on a first-come, first-served basis for eligible researchers — replacing the pilot's lottery-based selection.

Each grant covers a recipient plus up to two collaborators with free ChatGPT Plus access, supporting as many as 150,000 researchers in this round. OpenAI's stated goal is 200,000 researchers across the program. The company said it chose Plus over ChatGPT Pro for the new grants because most pilot usage would have remained within Plus limits.

The expansion widens who gets the tools. But the study's central implication sits elsewhere: whether the hour scientists save on individual tasks survives the checking they still perform — and the staffing, funding and experimental constraints that no chatbot can remove.

Source: unite.ai

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. OpenAI Study: Researchers Say ChatGPT Cuts Their Work Time Nearly in Half
  2. OpenAI Predicts AI-Made Discoveries by 2026 as Intelligence Costs Plunge
  3. 81% of Australian students use AI weekly as universities respond
  4. OpenAI Launches FrontierScience Benchmark for AI Research Skills
  5. 67.8% of Workers Now Use AI Weekly — But 56.4% Get No Time to Learn It

« Previous articleNext article »