Reports
OpenAI Study Finds Scientists Rarely Request End-to-End AI Research

Key PointsEnd-to-end autonomous research remains rare at just 0.4% of the average scientist's…
View more
- End-to-end autonomous research remains rare at just 0.4% of the average scientist's messages.
- Scientists still check AI outputs heavily with 45.4% citing correction as their main difficulty.
- OpenAI will offer 50,000 new free researcher grants on a first-come first-served basis.
OpenAI released a study on November 16, 2026, covering 580,186 scientific messages from 1,819 academic scientists and survey responses from 251 participants. It examines how the scientists used ChatGPT and Codex across research work and how they expect AI to affect discovery.
End-to-end autonomous research, defined as requests for AI to coordinate literature review, ideation and downstream analysis, accounted for only 0.4% of the average scientist’s scientific messages. Yet 15.3% of scientists sent at least one such request, a gap the paper interprets as experimentation with full automation rather than routine use.
The full paper analyzes activity from August 20 through September 22, 2026, and a survey conducted September 30 to October 7, 2026. Researchers used privacy-preserving classifiers to categorize messages without reading the underlying exchanges, and program data were not used for model training by default.
Reading and Writing Dominate Research Requests
Reading and knowledge acquisition accounted for 27.7% of the average scientist’s scientific messages, while writing and communication represented 26.9%. Scientific analysis contributed 16.7% and research software and code development 13.8%; together, those four workflows represented 85%.
That concentration puts most current AI use in discrete knowledge-work tasks rather than coordinated research projects. Explaining scientific concepts was the leading individual task at 18.8% of messages, with 89.7% of scientists sending at least three such requests.
Ideation tells a different story about breadth. It represented 5.3% of messages, but about 68% of scientists sent at least three ideation messages during the study. Across 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% could not be categorized.
The paper also identifies a stage-dependent division of labor. Scientists often asked AI to propose directions during early exploration, then supplied their own questions or ideas for refinement when developing specific research problems.
Human Review and Productivity Expectations
The paper’s scientist-involvement score, measuring the average share of initial requests followed by clarification, checking or revision, was 31.8%. Writing and communication had the highest score at 38.0%, followed by scientific analysis at 34.4% and research software at 33.7%.
For comparable tasks, requests requiring advanced or frontier expertise had higher involvement in 28 of 32 cases, by an average of 6.5 percentage points. More specialized work therefore generally drew more follow-up, not less.
The score considers only the next scientific message in the same conversation within seven days and does not capture revisions made outside the AI application. Checking or correcting outputs was also the most commonly selected difficulty in obtaining value from ChatGPT, chosen by 45.4% of surveyed scientists.
Those review demands coexist with self-reported time savings. 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. Reported productivity ratios ranged from 1.36 times for administration to 2.85 times for research software and code development.
Expectations for discovery were also high. Of the 250 scientists answering 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, including 21.6% who expected it within three years.
Written explanations emphasized execution over full automation. Among 367 reasons identified in 224 responses, 52% concerned faster research execution, while automating research loops accounted for 2.2%. Another 23% concerned finding more promising research questions and 14% pursuing a wider range of questions.
The reported bottlenecks extend beyond model capability. Limited time or staffing was selected by 51.8% of respondents and limited funding by 43.8%; validating or reproducing findings affected 25.5%, while access to equipment, laboratories or computing resources affected 24.3%.
OpenAI cautions that participants were a selected group who applied for expanded AI access and chose to answer an AI survey, so the results may not generalize to all scientists. Usage shares describe requests rather than research value or time allocation, and the findings do not establish causal effects.
OpenAI Expands Free Researcher Access
Alongside the study, OpenAI announced an expansion of ChatGPT for Academic Researchers, saying the pilot had reached 20,000 researchers in 33 countries since July 2026.
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 and up to two collaborators with free ChatGPT Plus access, supporting as many as 150,000 researchers in this round. OpenAI’s goal is to reach 200,000 researchers across the program.
OpenAI said Plus replaced ChatGPT Pro for the new grants because most pilot usage would have remained within Plus limits.
The expansion broadens access to the tools, but the study’s central implication concerns the work around them: whether time saved on individual tasks survives the checking scientists still perform and the staffing, funding and experimental constraints they report.












