Skip to content

OpenAI

Scientists tell OpenAI work takes 1.9x as long without AI

4 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: openai.com

OpenAI’s Economic Research team reported on November 16 that 1,819 scientists in the summer pilot of ChatGPT for Academic Researchers sent more than 580,000 scientific messages, and that 251 surveyed participants estimated the same work, at the same quality, would have taken 1.9 times as long without AI. If you’re a researcher who wants in, applications for the next round open November 23, with 50,000 grants going to eligible researchers first come, first served, each giving ChatGPT Plus to you and up to two collaborators rather than the Pro plan the pilot handed out.

What the next round offers

The round terms are on OpenAI’s program page, published the same day as the study. This round supports up to 150,000 researchers, and OpenAI’s goal for the program as a whole is 200,000, double the goal it announced when the pilot launched in July. So far, 20,000 researchers in 33 countries have joined.

Pilot participants each got a workspace and could invite up to four colleagues from their institution, according to the paper behind the study. A new grant covers only two collaborators, and on Plus. OpenAI says that when it reviewed pilot usage, most participants would have stayed within ChatGPT Plus usage limits, so offering Plus lets it extend access to more research teams. Getting in has changed as well: the pilot’s first round, 10,000 seats, was allotted by lottery, while this round goes to eligible applications in the order they arrive, so the date you apply matters.

By default, OpenAI does not use participants’ data from the program to train its models. For the study, automated classifiers sorted messages by type of research activity, such as reviewing literature, writing code or analyzing data; the research team saw those categories but did not read the messages or ChatGPT’s replies.

How scientists used it

68% of scientists used ChatGPT to develop research ideas. In messages about research ideas, 35% of the time the scientist brought a specific question for ChatGPT to work with, 36% of the time an idea they asked it to refine, and 23% of the time they asked ChatGPT to suggest a research question.

98% of those who answered the question about discovery timelines expect AI to speed up discovery in their field. Those who expected faster discovery were asked why: of 367 reasons found in 224 written answers, 52% were about getting research done faster, 23% about finding more promising research questions and 14% about pursuing a wider range of questions.

OpenAI itself says it will take more research to know whether those expectations turn into faster discovery, and that progress also depends on the funding, staff and experimental capacity scientists need to test their ideas.

Our analysis

What we checked

On November 16 we read the 76-page paper behind the post, pulled the data for its field-by-field chart out of the page, and compared the post with OpenAI’s program page.

How we check stories

The 1.9 times is an average built from fixed answer options, and the typical respondent expects a discovery that would take 10 years without AI to take 4 to 7 years with it. The post’s free access for “more than 200,000 researchers” overstates the program: the program page calls 200,000 the goal for the whole program.

A 10-year discovery, with AIRespondents (of 250)
Within 3 years54 (21.6%)
4–7 years126 (50.4%)
8–9 years48 (19.2%)
Just under 10 years16 (6.4%)
No acceleration6 (2.4%)

That puts the median respondent at a speed-up of between 1.43x and 2.5x, by our calculation from the band’s bounds. The “within 3 years” share ran from 31% in math and statistics (87 people) to 10% in engineering (51), in field groups that overlap.

The paper describes its respondents as scientists who “applied for expanded AI access and chose to answer a survey about AI use”; 251 of 270 valid completed responses consented to analysis. For the 1.9 times, they picked from nine fixed options per activity: “less time” was coded as 0.5, “400% or more” was capped at 5, and not using AI for an activity counted as no gain, before eight activity averages were weighted equally. By activity, the ratio runs from 1.35x for administrative work to 2.83x for research software and code.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.