OpenAI’s study of 1,819 scientists found that reading, writing, scientific analysis, and research software development dominated their scientific ChatGPT use. Published on November 16, 2026, Science at the Frontier: How Scientists Use ChatGPT combined classified message activity with a survey about productivity and future discoveries.
The distinction matters: the message analysis tracked what participants asked the tools to do, while the productivity figures came from participants’ estimates. OpenAI says further research is needed to determine whether those reported gains translate into faster scientific discovery.
Where scientists put ChatGPT to work
Reading, writing, scientific analysis, and research software together accounted for 85% of the average scientist’s scientific messages during the study period. Reading and writing alone represented more than half, with reading and knowledge acquisition contributing 27.7%.
OpenAI calculated each scientist’s share of messages in each activity, then averaged those shares with equal weight. A participant sending thousands of messages therefore had the same weight in that average as someone sending far fewer.
The largest individual task was explaining scientific concepts, at 18.8% of the average participant’s scientific messages. That use was widespread within the cohort: 89.7% of scientists sent at least three messages in that category during the observation window.
The picture is recognizable if you spend time around research: understanding unfamiliar material, communicating results, analyzing a problem, and getting code to cooperate. Much of the recorded activity concerned the work surrounding a research result.
How researchers developed ideas
OpenAI reported that 68% of participating scientists used ChatGPT to develop research ideas. Within messages classified as idea development, the starting points varied:
- 35% supplied a specific research question.
- 36% involved refining an idea the scientist supplied.
- 23% asked ChatGPT to suggest a research question.
- 6% were uncategorized.
Those percentages describe ideation messages. In most of those messages, the scientist contributed a question or an idea for further development, placing the model inside an existing thought process.
Requests for end-to-end autonomous research were a much smaller part of activity. About 15% of scientists tried at least one such request, but these requests accounted for just 0.4% of the average scientist’s scientific messages. The paper also reported low rates of returning to that type of request.
Who was in the study, and how were messages classified?
OpenAI analyzed messages sent from August 20 through September 22, 2026. The cohort generated 1,262,893 messages in total, of which 580,186 were classified as scientific or scholarly.
Participants came from ChatGPT for Academic Researchers, a pilot that began on July 29, 2026, at selected academic institutions. Eligible applicants received access through a lottery and could invite up to four colleagues from their institution. The pilot offered access to frontier models and higher usage limits.
For the usage analysis, OpenAI included 1,819 scientists who had sent at least one scientific message by September 2. That cutoff gave each included participant at least 21 calendar days of observation, counting their first day of use.
The analysis covered messages in ChatGPT’s Chat and Work modes and in Codex, including its command-line interface. GPT-5.6-Terra classified messages using a taxonomy of 10 workflows and 40 tasks.
According to OpenAI, the researchers analyzed privacy-preserving classifier outputs without reading the underlying messages or model responses. The company also says program content is not used to train its models by default.
OpenAI cautions that this selected pilot population, and the participants who chose to answer its survey, limit how broadly the findings can be generalized. These were researchers who had sought expanded AI access in the first place.
How much time did researchers think AI saved?
Survey participants estimated that completing the same amount of work, at the same quality, would have taken 1.9 times as long without AI. OpenAI conducted the survey from September 30 through October 7, 2026, and included 251 respondents who consented to the analysis.
The comparison held both the work and its quality fixed. As a hypothetical illustration of that ratio, work taking 10 hours with AI would be estimated to take 19 hours without it. The underlying inputs were researchers’ retrospective judgments about their work.
Those estimates differed substantially by activity. The mean estimated time without AI was 1.35 times the time with AI for administration, compared with 2.83 times for research software and code development. Coding stood out as an area where respondents perceived particularly large gains.
Checking the output also took time
Reviewing AI-generated work was a recurring concern: 45.4% of survey respondents selected the time needed to check or correct outputs as a difficulty in getting value from ChatGPT.
OpenAI separately measured follow-up behavior across 125,947 eligible initial requests from 1,809 scientists. The average scientist requested clarification, checking, or revision after 31.8% of those initial requests.
That measure had a specific boundary. It examined the next classified scientific message in the same conversation within seven days, excluding extensions and reactions. Checking or editing performed outside the application fell outside the measure.
For a researcher deciding where an assistant might help, the practical trade-off is straightforward: generating an answer and making that answer usable are both part of the job. The survey’s checking burden helps explain why a quick response can still leave substantial work for the person using it.
Research also had constraints beyond individual tasks. Among the 251 respondents, 51.8% selected limited time or staffing and 43.8% selected funding as research bottlenecks; participants could choose up to three constraints. OpenAI argues that scientific progress also depends on complementary resources such as staff and experimental capacity.
What researchers expected AI to do for discoveries
The discovery-timing question asked participants to imagine an important discovery in their field that would take ten years without AI. Of the 250 scientists who answered that question, 97.6% expected AI to bring it forward.
More specifically, 72% expected the hypothetical discovery at least three years sooner, while 21.6% expected it seven to ten years sooner. These were forecasts about an imagined research milestone, collected in the September–October survey.
Participants’ written explanations offered a useful view of why they were optimistic. OpenAI classified 367 separate claims from 224 written responses: 52% concerned faster research execution, 23% finding more promising questions, 14% pursuing a wider range of questions, and 11% other reasons. The percentages counted classified claims, so one respondent could contribute more than one.
OpenAI also illustrated that optimism with a testimonial:
“Tools like ChatGPT are helping us accelerate scientific discovery. It’s really a playground for a scientist like myself.”— César de la Fuente, antimicrobial researcher, in a testimonial published by OpenAI
The explanations extended beyond finishing existing tasks sooner. Some respondents expected AI to help them choose more promising questions or explore a broader set of possibilities. That distinction puts research direction alongside execution speed in their expectations for the technology.
What OpenAI announced for researcher access
Alongside the study, OpenAI announced a new round of 50,000 grants on November 16, 2026. The grants would provide free ChatGPT Plus access to eligible researchers, with each award covering the recipient and up to two collaborators.
OpenAI said the new round would allocate grants on a first-come, first-served basis, changing the lottery-based selection used in the pilot. The company said it chose Plus because most pilot participants’ usage would have fit within that plan’s limits.
The company’s stated goal for the overall program is to reach 200,000 researchers worldwide. The 50,000 announced grants are the next round within that broader program.
OpenAI scheduled applications for the new round to open on November 23, 2026.