The question most people bring to AI and science is whether the machine can do the work. The more revealing question, buried in a new study of how 1,819 researchers actually spent their time with ChatGPT, is what the machine leaves for them to do once it has done the work. OpenAI’s Economic Research team analyzed 580,186 scientific messages sent between August 20 and September 22, 2026, and surveyed 251 of those researchers. [1] The picture that emerges is not of a machine replacing the scientist outright. It is of a machine quietly absorbing the parts of scientific labor that used to belong to a person, and leaving behind a residue of checking, correcting, and second-guessing.
The Narrow Band of Routine
Reading and knowledge acquisition accounted for 27.7 percent of the average scientist’s messages. Writing and communication took 26.9 percent. Scientific analysis drew 16.7 percent, and research software and code development 13.8 percent. Those four activities together made up 85 percent of everything the scientists asked the system to help with. Explaining scientific concepts alone was the single largest task at 18.8 percent of messages, and 89.7 percent of scientists sent at least three such requests during the study period.
This is the part of research that has always been tedious, and it is exactly the part the tool now handles. A scientist who once spent an afternoon drafting a literature summary or debugging a script can hand that work over and get something back in seconds. The source of the efficiency is not that the machine thinks better than the researcher. It is that the researcher no longer has to do the thinking that was never the point of the job in the first place. What used to be a person’s afternoon is now a prompt, and the person is free — in theory — for something else.
Where the Machine Draws a Line
End-to-end autonomous research, defined in the study as requests for the AI to coordinate literature review, ideation, and downstream analysis together, accounted for just 0.4 percent of the average scientist’s messages. That is not a rounding error away from routine use. It is a signal that researchers are not yet handing over the whole of a project. Yet 15.3 percent of scientists sent at least one such request, which the paper reads as experimentation rather than reliance.
The distinction matters for anyone worried about what gets displaced. The machine is not being asked to run the lab. It is being asked to do the discrete, bounded tasks that once filled a researcher’s day. The scientist keeps the coordinating role, the judgment about what to pursue, and the responsibility for what comes out. That is a real division of labor, and it is one the researchers themselves appear to be enforcing. The boundary is not technical. It is chosen, message by message, by people who still believe the whole belongs to them.
The Follow-Up as the Last Human Reflex
The study’s scientist-involvement score measures how often an initial request was followed by clarification, checking, or revision within the same conversation and within seven days. The average was 31.8 percent. Writing and communication drew the highest rate at 38.0 percent, followed by scientific analysis at 34.4 percent and research software at 33.7 percent. When researchers worked on tasks requiring advanced or frontier expertise, follow-up rose in 28 of 32 cases, by an average of 6.5 percentage points.
Here is the shape of the new job. The machine produces, and the person audits. Checking or correcting outputs was the most commonly cited difficulty in getting value from ChatGPT, selected by 45.4 percent of surveyed scientists. The work that remains is not the work of discovery. It is the work of verification, and it is expanding to fill the space the machine has vacated. A researcher who once built the argument now grades the argument someone else built. The skill being exercised is no longer construction. It is suspicion, disciplined and repetitive.
The Speed Claim and What It Leaves Out
Across eight research activities, each weighted equally, surveyed scientists estimated that one hour of work with AI available would have taken about 1.9 hours without it, at the same quality. And 98 percent of those who answered a question about discovery timelines expected AI to accelerate discovery in their field. The figure is striking, and it is also self-reported. The paper is careful to note that these findings do not establish causal effects, and that scientists themselves flagged constraints on staffing, funding, and experimental capacity.
The expectation of speed sits uneasily beside the follow-up data. If a third of every request requires a human pass to check or revise, the time saved on generation is partly spent on inspection. The 1.9 multiplier measures the self-reported time on a task, not the vetting that follows it. What the survey captures is a belief about where the hours went, not an accounting of where they now go. The researchers who expect faster discovery may be right, but the mechanism they describe — less time on the routine — is the same mechanism that produces more time on the checking. Whether the ledger balances is a question the study raises without answering.
Ideation, or the Appearance of It
Ideation represented 5.3 percent of messages, but about 68 percent of scientists sent at least three ideation messages during the study. Within those messages, 36 percent involved a scientist bringing an idea for the system to refine, 35 percent began with the scientist’s own specific question, and 23 percent asked the system to suggest a research question. The remaining 6 percent could not be categorized.
The paper reads this as a stage-dependent division of labor: the machine proposes directions during early exploration, and the scientist supplies the questions when the problem sharpens. That reading is generous, and it may be right. But it also describes a role reversal at the front end of research. The moment when a scientist decides what is worth asking — the moment that has always defined the vocation — is increasingly a moment of choosing among options the machine generated. The researcher still selects. The researcher still refines. What the researcher no longer does, at least some of the time, is originate. The idea arrives as a menu, and the human picks from it.
The Reasons Scientists Give
Scientists who expected faster discovery were asked to explain why in their own words. From 224 written responses, the researchers identified 367 distinct reasons. The study reports that three reasons stood out, and they cluster around the same theme: the machine widens the search, surfaces more promising questions, and lets researchers pursue a broader range of work than they could otherwise attempt.
Those are ambitions, not results. The study is explicit that usage describes requests, productivity is self-reported, and discovery outcomes are expectations. What the scientists describe is a hope that the machine’s breadth will compensate for the narrowing of their own role. If the human no longer originates every question, perhaps the machine’s wider net will supply better ones. That is a plausible bet. It is also a bet that the value of the scientist now lies in curation rather than creation, in the ability to recognize a good question rather than the ability to pose one.
What the Program Itself Reveals
The pilot enrolled 20,000 researchers across 33 countries, and participants received ChatGPT Pro. [1] OpenAI is now expanding the program with 50,000 new grants on a first-come, first-served basis, each covering a recipient and up to two collaborators, with a stated goal of reaching 200,000 researchers. New grants provide ChatGPT Plus rather than Pro, a switch the company attributes to a usage review finding that most participants stayed within Plus limits. [1] Data from the program is not used to train the company’s models by default.
The scale of the expansion is the point. Two hundred thousand researchers is not a pilot. It is a working assumption that AI assistance will be a standard part of scientific labor, distributed widely enough to become unremarkable. The projects the company highlights show the range: one researcher exploring whether the system can connect laboratory and patient data to speed antibiotic development for tuberculosis, another turning specimen photographs into searchable databases. In each case, the machine is doing a task that a person used to do, and the person is doing something adjacent — framing the question, judging the output, deciding what the result means.
The Asymmetry Nobody Is Measuring
The benefits of this arrangement are easy to name. The costs are structural, and they fall on the same people in a different currency. A researcher whose day was once filled with reading, writing, and coding now spends a measurable share of it checking a machine’s work. The 31.8 percent follow-up rate is not a failure of the tool. It is the price of trusting it, paid in the one resource the tool was supposed to conserve.
The deeper asymmetry is in what gets exercised. Generation, synthesis, and the slow construction of an argument are the activities that train a scientist. They are also the activities the machine now performs first, leaving the human to react. A generation of researchers who learn to verify rather than to build may be efficient. Whether they will be capable of the originating work the field depends on is a question the study’s own numbers cannot settle. The 98 percent who expect faster discovery are not asked what they expect to be doing while discovery accelerates.
The Next Step the Data Points To
The study’s authors are careful about causality, and rightly so. What they have is a snapshot of usage, a survey of expectations, and a set of follow-up rates that describe behavior rather than outcomes. The next step they name is more research on whether those expectations translate into faster discovery, and they note that progress will also depend on funding, staff, and experimental capacity. That is the honest position.
But the snapshot already shows something the framing does not emphasize. The machine has taken over the routine, and the routine was where scientists learned their craft. The follow-up rate is the measure of what remains: a human reviewing, correcting, and approving. That is a real role, and it is a necessary one. It is also a role defined by the machine’s output rather than the scientist’s initiative. The question the data leaves open is not whether AI can do science. It is whether a scientist trained mainly to audit a machine’s drafts will still be able to originate the work the field depends on.
