OpenAI's own data shows scientists mostly use ChatGPT to read and write, not to do research for them
OpenAI is widening its giveaway to academics. On November 16 it said it will hand out 50,000 new grants of free ChatGPT Plus, each covering a researcher and up to two collaborators. Its goal is now 200,000 researchers, double the target it set in July.
The same day, OpenAI published a study of how scientists in its pilot actually used the tool. The answer is far more ordinary than the industry's talk of AI making discoveries on its own. Most of the work was reading papers, writing and coding.
Requests for the AI to carry out a research project from start to finish made up 0.4% of the average scientist's messages. For investors and buyers, the study is a rare look at real usage data from a leading AI lab, rather than test scores. It points to value in speeding routine knowledge work, with a review cost the headline time savings leave out.
What scientists actually asked for
OpenAI analyzed 580,186 science-related messages from 1,819 scientists in the pilot, sent between August 20 and September 22. Automated tools sorted the messages by activity. The Rundown notes the research team did not read the underlying content.
Four kinds of work made up 85% of messages. Reading and learning led at 27.7%, followed by writing at 26.9%, scientific analysis at 16.7% and research software at 13.8%. Brainstorming new ideas was only about 5% of messages, though roughly two-thirds of scientists tried it at least three times.
About 15% of scientists asked at least once for fully autonomous research, meaning the AI plans and runs a project with little human input. Few did so regularly, so its share of the average scientist's messages stayed at 0.4%.
Younger researchers leaned harder on coding help. Code was 17.9% of messages for those aged 18 to 29, against 9.8% for those 40 and over, according to The Rundown's reading of the data.
The time savings are the scientists' own estimates
The headline number is soft. In a survey of 251 participants, scientists estimated that work done with AI would have taken nearly twice as long without it. By activity, OpenAI's paper puts the estimates at about 1.35 times for administrative work to 2.83 times for research software.
These were guesses, not measurements. There was no comparison group, and respondents were asked to assume equal quality; no one checked whether the AI-assisted work was as good. The Rundown also notes that the roughly 1.9 times figure counts all eight activities equally, rather than by how much time each one takes.
OpenAI's paper flags its own limits. It says respondents are scientists who applied for more AI access and chose to answer, and that message shares describe use, not the value AI provides.
The biggest complaint was the cost of checking. Some 45.4% of scientists named the time needed to check or correct the AI's answers as an obstacle, the most common response. Roughly a third of requests were followed within a week by clarification, checking or revision.
That matters well beyond universities. Any company pricing in AI productivity gains should count the human review time that self-reported savings may leave out.
Scientists also said AI is not what holds them back most. In the survey, 51.8% cited limited time or staff and 43.8% cited limited funding. Faster reading and writing do not pay for more lab staff or equipment.
More seats, at a cheaper tier
The expansion also cuts OpenAI's cost per researcher. The summer pilot handed out ChatGPT Pro, OpenAI's premium tier. New grants are for Plus, the roughly $20-a-month plan. OpenAI says a usage review found most pilot users would have stayed within Plus limits. StartupHub notes the heaviest users will hit caps sooner.
At that list price, up to 150,000 Plus seats for a year would be worth about $36 million. That is machine.net's estimate. OpenAI gave no dollar value, and its actual cost of serving those users is not public. CASRAI reports the grants run for 12 months and do not include credits for building software on OpenAI's models.
Selection is also changing, from a lottery to first come, first served, with applications opening November 23. StartupHub notes that this favors well-organized labs with grant administrators ready to file on day one. About 20,000 researchers in 33 countries have joined since July.
The business logic is familiar from software: get people using a product early in their careers, and their institutions may buy licenses later. OpenAI already sells ChatGPT Edu, a campus-wide plan, to universities. It says it does not use data from the program to train its models "by default."
The study does not address the question investors care about most: whether free seats turn into paid university contracts. It also comes from a company that sells the product and published the results alongside plans to give more of it away. The usage data is useful. The productivity claims need independent testing before anyone budgets around them.