What the 7 hours AI saves scientists measures

I consult on lab operations and AI, so "how much time does this actually save?" is a question I get asked for real money. A credible number would be valuable. That is exactly why it's worth reading how this one was made before repeating it.

Where the 7 hours comes from

The paper is "AI in Science: Early Insights," posted to arXiv in September 2026 (version 2, 25 September) by 18 authors from Google, Google DeepMind, MIT FutureTech and several universities. It combines three data sources: about 15 million Gemini interactions (of which roughly 360,000 were classified as science), an inventory of 2,690 specialized scientific AI models, and a survey of scientists.

The 7 hours comes only from the survey. Per the paper's appendix, the polling firm More in Common interviewed 637 scientists in the US (379) and UK (258) online between 27 July and 11 August 2026, recruited "through specialist research panels" and screened to people who actively do research. The result, in the paper's words: "just below three quarters of respondents report that AI saves them time on net in their working week, compared to only 6% reporting net time lost. The average scientist's time savings are around 6.9 hours per week."

Note the verb. Nobody timed anyone. Scientists were asked to estimate their own time savings, and 6.9 hours is the average of those estimates.

What the authors say about their own sample

To their credit, the authors are direct about this. On representativeness: "There is no reliable population frame for scientists in either the UK or the US, so we think of this as a screened non-probability sample rather than a representative one, and the data are therefore reported unweighted with no margin of error quoted."

On which direction the error probably runs: "those scientists who are already enthusiastic about or frequently use AI might be more likely to respond to a survey about AI in science. This could potentially overestimate the real adoption rate and the average weekly time savings."

The sample also skews senior and toward industry. By my count from the appendix, 356 of 637 respondents hold senior roles (PIs, professors, lab directors, R&D managers), 234 are mid-career (postdocs and staff scientists), only 47 are early-career (junior researchers and PhD candidates), and 368 work in industry.

So the honest reading is: among a screened, senior-leaning online panel of US and UK scientists, the average self-estimate was about 6.9 hours a week. That is a real finding. "AI saves scientists 7 hours a week" is a different, bigger claim.

Why self-reported time savings need a check

People are not great at estimating how long things would have taken without a tool. The clearest test I know of is a 2025 randomized trial by METR (Joel Becker, Nate Rush, Elizabeth Barnes and David Rein). Sixteen experienced open-source developers did 246 real tasks, each randomly assigned to allow or forbid AI tools. Before starting, developers forecast AI would cut completion time by 24%. Afterward, they estimated it had cut time by 20%. The measured result: "allowing AI actually increases completion time by 19%."

That study has real limits for this comparison. It is software developers, not scientists, working in codebases they knew deeply, with early-2025 tools. It does not prove scientists are wrong about their 7 hours. What it does show is that self-estimates and stopwatch measurements of AI speed-ups can disagree in sign, not just in size. Until someone measures scientists' time directly, the 6.9 hours is an upper-bound-flavored perception.

The findings I'd actually pay attention to

The same survey contains results that are less quotable and, to me, more useful, because they describe where the time goes.

A verification tax. Of the scientists who said they saved time, 89% spend more than a tenth of that saved time checking AI outputs, and 46% spend more than a quarter of it. In a lab, that matches my experience: AI output that touches a measurement or a calculation has to be checked by someone who could have done it.

The bottleneck moves to the bench. About 44% said their main bottleneck moved downstream (to wet-lab work, validation, or manuscript drafting), against about 14% who said it moved upstream. And 41% said their backlog of untested hypotheses has grown, against about 25% who said it shrank. AI can generate ideas faster than instruments and people can test them. A faster idea pipeline into a fixed number of instrument-hours does not produce more results; it produces a longer queue.

A drift toward safer questions. 49% said AI pushes them toward safer, more incremental projects, against 28% who said it lets them take on riskier ones. The authors' own summary calls this "Most concerning for the long run." I agree it's the sentence that deserved the headline.

All of these are self-reports from the same panel, with the same caveats. But they are self-reports about friction, which enthusiasts have less reason to inflate.

How to read the next "AI saves X hours" number

Measured or estimated? A timed comparison or a randomized design beats a survey answer. Most published figures are estimates.

Who answered? An opt-in panel about AI will draw people who like AI. Look for whether the authors weighted the sample or say it is non-representative (these authors did say so).

Net of checking? Ask whether "time saved" already subtracts the time spent verifying the output. Here, the verification share was asked separately.

Did output change? Time saved on drafting is not the same as more experiments run. In a lab, the binding constraint is usually instrument time, samples and people at the bench.

The paper itself is careful and calls these "early insights," which is what they are. The problem is the trip from appendix to headline, where an average of self-estimates from a non-random panel becomes a measured fact about how science works.

Disclosure, plainly: I use AI tools in my own work, including Claude (made by Anthropic, a Google competitor) for research and drafting for this publication. I am a customer of Anthropic's products and have no other relationship with the company, Google, Google DeepMind, MIT, More in Common, METR or any of the authors; none of them asked for, saw or reviewed this report. Nothing here is sponsored and no link earns a commission; here's the full policy.