Short answer: On October 8, 2026 the DeepMind Institute published an essay, Bending the Curve of Discovery: AI in Science Today and Tomorrow, that puts real numbers on how researchers use AI. Scientists use it more than other professions, nearly half use it daily, and they report saving just under seven hours a week. The essay's sharper point is what comes next: the bottleneck is moving from computing to the lab bench, and AI may be steering some researchers toward safer questions.
What DeepMind published and what it is built on
The essay comes from the DeepMind Institute, the essay-and-research outlet Google DeepMind launched in September. We covered that launch, including its disclaimer that essays reflect individual authors' arguments rather than company policy, in our DeepMind Institute launch explainer. This piece is more empirical than the launch slate: it summarizes a study that the essay cites as Codreanu et al., 2026.
According to the essay, the study combined three data sources:
- 15 million interactions across Gemini surfaces, used to see what researchers actually ask general-purpose models to do.
- Bibliometric data on more than 2,600 specialized AI models, ranging from protein structure prediction to weather forecasting and materials discovery.
- A new survey of more than 600 scientists in the US and UK (the figure captions give N = 637).
Tasks were mapped against a taxonomy of scientific work developed by MIT FutureTech, which the authors say gives a granularity earlier labor research lacked.
One caveat up front: the data is partly Google's own product logs and partly self-reported. That does not make it wrong, but it is not an independent audit.
The headline findings
The essay lists several results directly:
- Scientists are over-represented in Gemini usage relative to their share of employment, so they use AI more than most professions.
- Nearly half of surveyed scientists use AI every day in their workflow.
- Specialized models cover every scientific field, and in almost half of cases they sit in the top 1% of citations for their domain.
- Adoption is geographically concentrated, scaling with a country's scientific workforce.
- Productivity gains are large on paper: an average of just under seven hours saved per week, mostly put back into research.
A detail worth underlining: general LLMs and specialized models do different jobs. LLMs handle broad tasks such as coding and writing, while specialized models like AlphaFold, GNoME and MatterGen support domain-specific data analysis and modeling. Scientists report spending over 30% of their "AI time" on specialized models, and the essay says task overlap between the two classes is limited, so they behave like economic complements rather than substitutes.
That framing matters for builders. If you are shipping an AI product for researchers, the evidence here says a chat window alone covers only one half of the workflow.
From "general purpose technology" to an "invention of a method of invention"
The most interesting conceptual move is the idea of AI as an invention of a method of invention (IMI). A general-purpose technology cuts the cost of existing routines. An IMI, like the microscope or telescope, permanently changes how discovery itself is produced.
The essay uses AlphaFold as the example. Experimental crystallographers resolved roughly 170,000 protein structures over decades. AlphaFold 2 predicted more than 200 million, and the open database built from it has been used by more than 4 million researchers in 190 countries. The essay adds that basic research leveraging AlphaFold predictions rose by 15 to 40 percent.
The authors call the difference "more is more" versus "more is different." Saving a lab years on one protein is the first. Clustering all 200 million structures at once to reconstruct evolutionary histories that sequence alignment cannot see, or running exhaustive all-by-all interaction screens, is the second. It places prediction inside the inner loop of large-scale search, which is also the logic behind agentic systems such as Google's co-scientist work we covered in DeepMind's co-scientist moving into real-world labs.
Other examples the essay cites: GNoME predicting millions of candidate crystal structures, the complete map of neuron connections in the adult fruit fly central nervous system, a more representative human pangenome, and deep learning helping to control plasma in experimental fusion reactors.
The friction points: the lab is now the slow part
Here the essay turns candid. Speeding up the front end of research "is triggering an inversion of the research process." Historically the wet lab generated ideas and computers assisted. Now in silico simulation increasingly generates the ideas and the physical lab verifies them.
Survey numbers quoted in the figure captions:
- 45.7% report spending over 25% of their AI-saved time auditing AI output.
- 43.5% say their primary bottleneck moved downstream over the past two years.
- 40.5% report a growing backlog of untested hypotheses.
The essay also notes a sobering result from a recent study of AlphaFold's impact: less-studied proteins did attract more basic research, but there is so far little evidence they have fed downstream applied research or early drug discovery, with exceptions in neglected diseases such as Chagas disease and leishmaniasis. Understanding a protein's biological context is slow and physically constrained.
Quoting Herbert Simon, the authors argue that "a wealth of information creates a poverty of attention": the scarce resource is no longer hypothesis generation but the bandwidth of theorists and experimentalists who interpret and test the output.
Does AI make science safer and narrower?
Two findings pull in opposite directions:
- Over two-thirds of surveyed scientists say AI tools expanded their access to insights outside their primary field.
- Almost half say AI encourages them to focus on incremental questions, concentrated among junior researchers, against about 30 percent who say it lets them take riskier ones.
The essay connects this to related work finding that AI-enabled science expands individual impact (tripling publications and quintupling citations) while narrowing the collective range of inquiry toward data-rich, easily verified domains. It also raises the training-pipeline concern: if AI automates paper generation and junior-level lab execution, peer review could be overwhelmed and the apprenticeship that trains future scientists hollowed out. Pushmeet Kohli is cited warning against "epistemic complacency," scientists becoming passive consumers of black-box models.
The authors' counterintuitive suggestion is that more capable AI, specifically agents that synthesize disparate knowledge into prioritized and unexpected hypotheses, could help break the inertia. That is a claim, not a result.
Claimed versus verified
Because this is a lab studying its own users, it helps to separate what is documented from what is argued.
| Item | Status |
|---|---|
| 15M Gemini interactions, 2,600 specialized models, 637-scientist survey | Stated by the essay; the underlying paper is cited as Codreanu et al., 2026 |
| About 7 hours saved weekly | Self-reported survey average, not a measured time study |
| Nearly half use AI daily | Self-reported by respondents in the US and UK only |
| AlphaFold numbers (200M structures, 4M users, 190 countries) | Consistent with DeepMind's earlier public statements |
| AI as an "invention of a method of invention" | An argument, not an empirical finding |
| Agents can offset incremental thinking | A hypothesis the essay itself frames as a possibility |
How to check yourself: the same dataset was released earlier as Google's "AI in Science" report, which we read in full in our report breakdown (including its own caveat that the sample is not representative). This essay, by Alex Imas and James Manyika, is a later, interpretive layer on that data. Compare against independent surveys too. Google's separate February 2026 research on Accelerating Scientific Research with Gemini gives case-study evidence from the same direction, though it is a different paper.
Why it matters for people building with AI
- The next product gap is validation, not generation. If scientists already drown in candidate hypotheses, tools that rank, de-duplicate, reproduce and auto-verify have more value than another idea generator. See the shape of this problem in Terminal-Bench Science, which tries to measure agents on real research tasks.
- Pair general and specialized models. The complement finding suggests agents that can call AlphaFold-class tools beat chat-only assistants. Compare how Anthropic's Claude for science workbench and OpenAI's Rosalind life-sciences workbench approach the same split.
- Expect lab automation to become the strategic layer. The essay explicitly names the lag in automated lab technology as a cause of the bottleneck.
- Watch the evidence standard. Self-reported hours saved are the same kind of metric critics question in coding assistants. Treat 7 hours as directional.
- People are moving. The talent market in AI for science is shifting too; we tracked one example in John Jumper leaving Google DeepMind.
What to watch next
The essay is titled "today and tomorrow," and the tomorrow half is where the institute is likely to publish more: institutional reforms for peer review, training, benchmarks and compute access. Also watch whether the underlying study is released for independent replication, and whether other labs publish comparable usage data for their own science users. For the strategic context behind Google's science push, see Jeff Dean's discovery loop and the Google DeepMind shakeup and the earlier Google AI scientist work with chain-of-evidence.
Related reading
- The DeepMind Institute launch explained
- DeepMind co-scientist in real-world labs
- Terminal-Bench Science
- Claude as an AI workbench for scientists
- OpenAI Rosalind life-sciences workbench
- John Jumper leaves Google DeepMind
- Jeff Dean's discovery loop
- Official: Bending the Curve of Discovery · DeepMind Institute
Figures here are drawn from the essay as published on October 8, 2026 and are the authors' claims, not independently audited results.