"Scientific AI Is Compressing Centuries of Research Into Days"

James Manyika, Google and Alphabet's senior vice president for research, labs, technology, and society, recently put a striking number on what's happening at the intersection of artificial intelligence and science: researchers are now able to "condense hundreds of years of research into months or even days." It's the kind of claim that sounds like marketing until you look at the specifics — and the specifics are piling up fast enough that the framing is hard to dismiss.

In a wide-ranging interview with Scientific American, Manyika compared the current moment to the heady days of early 2022 for generative AI, when capability was advancing at a pace nobody had fully internalized yet. The analogy is instructive because it's about the shape of progress, not just its speed. Just as language models quietly crossed a usefulness threshold that took the industry years to fully metabolize, scientific AI appears to be crossing a similar threshold across biology, chemistry, physics, mathematics, and atmospheric science all at once — a broad, simultaneous inflection rather than a single breakthrough.

The clearest way to see that inflection is in a single change in the kind of question scientists can ask. Manyika points to the shift from predicting one protein to predicting all of them. That's not a linear speed-up; it's a change in the unit of inquiry. When AlphaFold could fold a single protein, it was an impressive tool. When the freely available AlphaFold Protein Database let more than three million researchers across 190-plus countries query hundreds of millions of predicted structures, the tool became something closer to infrastructure — and the scarce resource shifted from "can I get an answer" to "which of the millions of available answers should I trust, and how do I design the next question."

That reframing — from answer-scarcity to question-scarcity — is the thread running through Google DeepMind's Gemini for Science, the suite of experimental tools rolled out at I/O 2026. It bundles four capabilities: Hypothesis Generation, which scans millions of papers to surface testable directions with verified citations; Computational Discovery, an agentic search engine that can auto-generate thousands of experimental variants; Literature Insights, which synthesizes research into reports, infographics, and audio or video overviews; and Science Skills, which stitches more than 30 major life-science databases into a single workflow. The ambition is explicit: compress the research loop from hours or months to minutes, and let scientists spend their effort on judgment rather than retrieval.

What's easy to miss in a feature list is the division of labor it implies. When a model can enumerate thousands of candidate molecular variants before lunch, the bottleneck in discovery moves from generating possibilities to testing them — from the intellectual bottleneck to the experimental one. Manyika gestures at exactly this when he says the most valuable skill for an early-career scientist will be the ability to "frame questions and design lines of inquiry." The scientific method isn't being automated away; the part of it being automated is the part that was always the slow, mechanical grind.

The democratization story deserves a closer look, because it cuts both ways. The AlphaFold database is genuinely open — three million researchers in more than 190 countries is a remarkable spread for a tool that would have been the exclusive domain of a handful of well-funded labs a decade ago. But there's an asymmetry hiding underneath: open access to results is not the same as open access to the means of production. Training frontier scientific models remains data- and compute-intensive in ways most academic and national labs can't match, which is precisely why Manyika argues the traditional relationships between academia, industry, and government need to be "rethought and reset." The democratization of scientific AI is real, but it currently runs in one direction — everyone can read the answers, while a small set of labs builds the machines that produce them.

Weather and flood forecasting may be the clearest proof that this isn't all abstract. Google's Flood Hub now covers more than two billion people in over 150 countries, built on AI hydrology models that predict riverine floods with enough lead time to matter. What makes this domain special is that it has ground truth — a flood either happens or it doesn't, on a schedule you can verify. That feedback loop is why forecasting keeps improving in ways that are easy to audit, and it's a useful template for where AI earns trust fastest: domains with real-world, falsifiable outcomes rather than open-ended text generation.

The honesty in the reporting is worth preserving. The clearest risk identified around systems like Gemini for Science is that a hypothesis-generation tool could surface "plausible-sounding but subtly incorrect citations at scale," and a research team that relies on it without verification could quietly propagate flawed claims into the published literature. In other words, the failure mode of scientific AI is less likely to be a dramatic wrong answer than a confident, well-cited one that's subtly off. That turns the skill of evaluating AI outputs into a core scientific competency rather than a nice-to-have — and it makes "does the citation actually say that" the new "did you actually run the control group."

Manyika's advice to early-career scientists is quietly radical in its framing. He encourages them to ignore the old dichotomies between theoretical, computational, and experimental work, and to think across disciplines with a sophisticated understanding of state-of-the-art AI. That's a description of a different kind of scientist — one for whom the ability to interrogate a model is as fundamental as the ability to interrogate a dataset. It's not that the domain expertise goes away; it's that it gets paired with a new, non-negotiable layer of machine literacy.

The balanced read is that scientific AI is at an inflection point that is simultaneously thrilling and genuinely unfinished. Hundreds of years of research compressing into days is a real phenomenon, visible in protein folding, materials discovery, and flood forecasting alike. But the moment also comes with an open question about who gets to participate in building the tools, and a new responsibility to verify what the tools claim. The most encouraging part of Manyika's account is how much of it is already about humility — about resetting relationships, teaching evaluation, and framing better questions. That's a striking thing for a research chief to lead with, and it's probably the right note for a technology that is, for the first time, changing the pace of science itself.

You can read the Scientific American interview with James Manyika and the rundown of Gemini for Science for the specifics, and explore the openly available AlphaFold Protein Database to see the democratization story firsthand.