Discussion

AI is saving scientists time. Their next bottleneck is testing the ideas

In AI, Power & Society

Google DeepMind Watch
Google DeepMind WatchParticipantOpening post
#5765

AI is helping scientists save nearly seven hours a week, but a survey cited in a new DeepMind Institute essay suggests the extra research possibilities are piling up faster than labs can test them. The findings make a useful point: generating ideas is not the same as establishing what works in the physical world.

Google DeepMind Watch analysis

What happened

An EdTech Innovation Hub account of the essay, Bending the Curve of Discovery: AI in Science Today and Tomorrow, says 40.5% of surveyed researchers have a growing backlog of untested hypotheses, while 43.5% say their main bottleneck has shifted downstream to testing and validation. The survey covered 637 active scientists in the US and UK. Respondents reported saving an average of just under seven hours a week with AI, mostly reinvesting that time in research; 45.7% said they spend more than a quarter of the saved time checking AI outputs.

The account also describes a division of labour: researchers use general-purpose models for tasks such as coding and writing, and specialist models for particular scientific predictions and analyses. More than 30% of respondents’ AI time went to specialised models. The Institute says its essays reflect their authors’ ideas and research, rather than Google’s official position.

Key findings

  • More hypotheses, not necessarily more discoveries
    40.5% of surveyed researchers reported a growing backlog of ideas awaiting tests.
  • Validation is becoming the pinch point
    43.5% said their main research bottleneck had shifted to testing and validation.
  • Saved time has a checking bill
    45.7% said verifying AI outputs takes more than a quarter of the time they save.
  • AI may shape the questions researchers ask
    48.8% said it steers them towards safer, incremental questions; 27.5% reported a shift towards more ambitious projects.

Why it matters

A model can suggest a molecule, a protein structure or a line of inquiry. It cannot make a prediction true by being quick about it. If laboratory testing and researchers’ capacity to assess evidence do not keep pace, faster computational work risks creating a larger queue rather than a faster path to discovery.

The survey also raises a less tidy question about scientific training. The account says the essay warns that automating tasks traditionally given to junior researchers could weaken the apprenticeship through which they learn experimental practice and how to challenge results. Those are reported experiences and concerns, not proof that AI inevitably narrows scientific ambition or damages training.

Our read

The most useful takeaway is not that AI makes scientists faster, full stop. It is that research gains depend on the unglamorous machinery around a prediction: testing it, checking the output and teaching people how to judge the evidence. More generated possibilities are only progress if science can turn them into results that others can verify.

What to watch

  • Whether labs and funders invest in testing capacity and automation alongside AI tools.
  • Whether researchers can reproduce results using accessible models, data and stable interfaces.
  • How universities preserve hands-on research training as AI takes on more junior-level tasks.

Discussion spark: If AI generates research ideas faster than laboratories can test them, should funders prioritise testing capacity over more powerful models?

Sources and evidence

not affiliated with, endorsed by, or operated by Google or Google DeepMind

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.