Science at the Frontier: OpenAI’s Study of 1,819 Scientists Using ChatGPT

OpenAI’s Economic Research team analyzed 580,186 scientific messages sent by 1,819 researchers on the ChatGPT for Academic Researchers plan between August 20 and September 22, 2026. The company published the results on November 16. A separate survey ran from September 30 to October 7, and responses from 251 participants appear in the analysis.
The team did not read the messages. Automated classifiers sorted them by research activity, and the researchers saw only the categories. Training on participant data is off by default on the plan.
What scientists asked for
Averaged across scientists, with each person weighted equally, reading and learning made up 28% of scientific messages and writing and communicating 27%. Analyzing evidence and theories came to 17%, writing research software to 14%, and developing ideas to 5%. Teaching and mentoring and research administration each accounted for 3%, reviewing research for 2%, and gathering data for 1%.
Ideas
Sixty-eight percent of scientists sent at least three messages about developing research ideas. Within those messages, 35% began with a specific question the scientist brought, 36% began with an idea the scientist asked ChatGPT to refine, and 23% asked ChatGPT to suggest a research question. The remaining 6% could not be categorized. César de la Fuente, who works on antimicrobial research, described the tool as “a playground for a scientist like myself.”
OpenAI also measured how often scientists followed a request with a clarification, a check, or a revision. The measure looks at the next scientific message in the same conversation within seven days and is averaged within each activity.
Time and discovery estimates
Across eight research activities, survey respondents estimated that an hour of work with AI available would have taken about 1.9 hours without it. Each activity counted equally, with 249 to 251 respondents per activity. A scientist who reported using no AI for an activity was counted as saving no time.
The survey also asked 250 scientists to imagine a discovery that would take ten years without AI and estimate how long it would take with AI. Of those who answered, 98% expected AI to speed up discovery in their field, and any answer of ten years or longer counted as no acceleration. Among 224 written explanations, OpenAI identified 367 reasons. Getting research done faster made up 52%, finding more promising research questions 23%, and pursuing a wider range of questions 14%. The other 11% fell outside those groups.
Limits stated by OpenAI
The company says usage data describes requests made to AI, productivity figures are self-reported, and discovery results are expectations. It states that the findings do not establish causal effects. Respondents also reported constraints on staffing, funding, and experimental capacity. By my calculation, the 251 survey respondents are about 14% of the 1,819 scientists whose messages were analyzed.
The full paper and a key-findings download are linked from the report page at openai.com/index/science-at-the-frontier. OpenAI says it is expanding ChatGPT for Academic Researchers with free access for more than 200,000 researchers.
Author bio
David Borish is the author of The Tony Hawk Paradox: When Video Games Predict Reality and writes at davidborish.com
