TL;DR

An early-findings paper from Google, Google DeepMind and MIT FutureTech reports that scientists using AI save an average of 6.9 hours a week. It also identifies a “verification tax”: for 46% of those saving time, checking AI output used up more than a quarter of the hours saved. The authors’ biggest worry is different again, with 49% saying AI steers them towards safer, more incremental work.

The survey

The paper, published on 16 September, draws partly on a survey of more than 600 scientists in the UK and US conducted over July and August. Almost three-quarters said AI saves them time. Around 80% said their lab produces more than it did three years ago, and 89% expect that to keep rising.

The paper describes the time savings flowing back into more research. But almost everyone who saved time also reported spending a “meaningful share” of it verifying what the AI produced, and nearly half of that group lost more than a quarter of their gain this way.

Mihai Codreanu, a Google senior economist who co-led the work, said that “as some tasks become easier, bottlenecks shift downstream”, pointing to growing backlogs of untested hypotheses.

The riskier finding

The paper calls a different result its “most concerning” for the long run. Nearly half of respondents said AI nudges them towards projects where benchmarks exist and results are dependable, while only 28% said it helps them take on riskier questions. The authors warn this could raise the number of papers without moving the scientific frontier.

UCL research fellow Basil Mahfouz, who also works with the Research on Research Institute, said big productivity gains depend on solving validation. His own research found AI answering policy questions drew on a narrower spread of research than humans did, which he said points to “less diverse, less risky and less innovative science”.

Why this travels beyond the lab

The verification tax is not unique to science. A tribunal judge said last week that litigants in person must check their AI output, and the same checking burden applies in legal, accountancy and consulting work, where AI time savings are often quoted without it. For UK firms building business cases, this is a useful corrective: a headline hours-saved figure is a gross number, and the net figure depends on how much review the output needs.

Looking forward

The paper is labelled early insights and is based partly on self-reported survey data, so the figures are indicative. Google is also among the companies putting significant effort into AI tools for science. The more durable point is the shift in where the bottleneck sits. Organisations that measure only generation speed will overstate their gains, and those that invest in faster, more reliable verification may capture more of them.