OpenAI's ChatGPT Cuts Research Time in Half for 1,819 Scientists
OpenAI's Economic Research team analyzed 580,000 messages from 1,819 scientists in a pilot program, finding researchers finish work nearly twice as fast.
- OpenAI studied 580,000 messages from 1,819 scientists in its academic pilot.
- Surveyed researchers said comparable work would have taken nearly twice as long without ChatGPT.
- 68% used ChatGPT for ideation, usually bringing their own specific question to refine.
- Top use cases: exploring ideas, analyzing evidence, writing code, and communicating findings.
- OpenAI is expanding the program to 200,000 researchers, double the original target.
- Pilot used ChatGPT Pro, but next cohort gets Plus since most users fit within Plus limits.
OpenAI analyzed more than 580,000 messages from 1,819 scientists and surveyed 251 participants about their work with ChatGPT. Survey respondents estimated that completing the same research tasks at the same quality would have taken nearly twice as long without the model. The estimate captures perceived time savings, and the study’s observational design cannot establish a causal productivity effect.
The sample behind the claim
OpenAI’s Economic Research team published the study report after examining one month of activity from the ChatGPT for Academic Researchers summer pilot. The researchers combined message-level usage data with survey responses to study how scientists incorporated the chatbot into active projects.
| Measure | Study detail |
|---|---|
| Participants | 1,819 scientists |
| Messages analyzed | More than 580,000 |
| Survey respondents | 251 |
| Observation period | One month |
| Main estimate | Respondents said the same work would have taken nearly twice as long without AI |
| Primary limitation | The estimate is self-reported and does not establish causation |
The survey asked participants to estimate a counterfactual: how long the same work would have taken without ChatGPT while maintaining the same quality. Researchers did not independently time an AI-assisted group against a control group or grade the resulting work, so the study leaves actual time saved and output quality unverified.
Ideation led the workload
Scientists used ChatGPT for literature review, evidence analysis, coding, idea development, and communication. Ideation was the most common category: 68% of participants sent at least three messages about developing research ideas, which was OpenAI’s threshold for counting that use case.
The ideation conversations followed a clear order:
- Scientists most often supplied a specific question for ChatGPT to explore.
- They less often supplied an existing idea and asked the model to refine it.
- They least often asked the model to propose a research question.
Those patterns show researchers using the model mainly to examine and refine their own questions. Participants also continued reviewing outputs after delegating tasks, keeping human judgment inside the workflow.
What telemetry adds
The pilot connects a defined population of expert users with aggregate records of their actual ChatGPT activity. That combination provides more behavioral detail than a survey alone, including the tasks scientists attempted and the ways they initiated ideation sessions.
OpenAI says data from the academic program is excluded from model training by default. Its Economic Research team aggregated the findings with privacy-preserving methods and did not read individual conversations.
An earlier field experiment by Dell’Acqua and collaborators provides a reference point. Consultants using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and produced work rated 40% higher in quality on assignments inside the model’s “capability frontier,” meaning tasks the system could handle reliably. Differences in participants, tasks, outcome measures, and study design limit direct comparisons with OpenAI’s scientist pilot.
The scientists reported a larger time effect, although the study did not test the reason. Their workload included substantial writing, coding, and summarization, while physical laboratory work remained outside the chatbot’s execution capabilities.
Twice the seats, a lower tier
In its July 29, 2026 program announcement, OpenAI said ChatGPT for Academic Researchers would provide free frontier-model access to as many as 100,000 scientists, mathematicians, and engineers by the end of 2027. Selected researchers receive GPT-5.6 Sol Pro through ChatGPT, ChatGPT Work, and Codex.
OpenAI now plans to serve 200,000 researchers worldwide, doubling the original target. The company also plans to move away from lottery-based selection following feedback from participants.
Pilot participants received ChatGPT Pro, but OpenAI found that most stayed within ChatGPT Plus usage limits. The next cohort will receive Plus, a shift that the company says will support the larger seat count.
The project extends OpenAI’s broader effort to measure workplace adoption. Its earlier Work at the Frontier study analyzed more than 800,000 work-related messages from U.S. ChatGPT Business users across occupations; Science at the Frontier applies a similar approach to academic research.
Product lessons from the prompts
- Design for researcher-led questions. Most ideation sessions began with a specific question or existing idea, so tools should support iterative exploration and refinement.
- Keep outputs inspectable. Participants continued reviewing delegated work, making editable responses and visible reasoning steps relevant to the observed workflow.
- Measure outcomes separately. Message telemetry can reveal adoption and interaction patterns, but evaluations still need independent measures of speed, accuracy, and research quality.
- Scope claims by task. The pilot covered knowledge work such as analysis, coding, and writing. It provides no evidence that physical experiments run faster.
The boundary of the result
OpenAI’s data shows where a general-purpose chatbot entered scientific work during one month: ideation, literature review, evidence analysis, coding, and communication. The survey indicates large perceived time savings, while its method leaves the size of any causal productivity gain unresolved. For researchers and developers, the clearest findings concern workflow fit, prompt behavior, and the continued role of human review.