Report 009 · AI in the Lab
Google DeepMind announced its AI had found 2.2 million new crystals, "nearly 800 years' worth" of discovery. Then working chemists opened the list and asked a simple question: how many of these are actually new, real, and useful? The gap between those two sentences is the whole lesson.
In late 2023, Google DeepMind made one of the biggest-sounding claims in modern materials science. Its tool, GNoME, had predicted 2.2 million new crystal structures, and 380,000 of them were flagged as stable enough to be "promising candidates for experimental synthesis." The company called it "equivalent to nearly 800 years' worth of knowledge" and reported that outside labs had already made 736 of the structures. A companion effort with Lawrence Berkeley National Laboratory paired the predictions with a robotic "A-Lab" that synthesized dozens of new compounds. The headlines wrote themselves: AI discovers millions of materials.
I work with AI in a lab context, and I use tools like this. So I want to be clear that GNoME is a genuine achievement of computation. But there's a verb in that headline doing an enormous amount of work, and it's worth slowing down on. What did "discover" mean here?
GNoME didn't make 2.2 million materials. It predicted 2.2 million structures: arrangements of atoms that a model calculates should hold together. That's real and useful information. It is also not the same thing as a material. A material is something that exists, that you can synthesize, hold, and put to work, that does something. The distance between "the math says this crystal is stable" and "here is a substance that matters" is exactly where the science lives, and it's the distance the headline collapses to zero.
This isn't pedantry. It's the difference between a list of possible chess positions and a game actually won. The list can be astronomically large and still contain very little you'd want.
The most useful response came from people who do this for a living. In April 2024, Anthony Cheetham and Ram Seshadri of UC Santa Barbara published an analysis in the journal Chemistry of Materials examining what was actually in the GNoME output. Their verdict, quoted verbatim: they found "scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility." In plainer words, when you sampled the predicted structures, they tended to be either not new, not plausibly makeable, or not good for anything in particular, and often more than one of those at once. As reported by 404 Media, the authors noted they had "yet to find any strikingly novel compounds" in the listings, and that "many of the new compositions are trivial adaptations of known materials," small swaps on things already known. One of the authors put the practical upshot bluntly: the work was "not particularly useful to experimentalists such as ourselves."
Sit with that. Not fraud, not error, just a category slip. Swap one element for a chemically similar one across a database of known crystals and you can generate millions of "new" structures that are stable on paper and almost entirely uninteresting. A count of candidates is not a count of discoveries, and the gap between the two turned out to be most of the number.
An earlier report here covered a case where an AI missed a real, hidden drug pocket that a crystallographer found by hand. This is that lesson's mirror image. There, the model saw too little. Here, it produced far too much, an ocean of candidates with the signal thinned almost to nothing. Both failures share one root: today's models are strong at generating and interpolating what's plausible, and weak at the thing that actually defines discovery, which is judging what is genuinely new and genuinely worth the bench time to chase. That judgment still comes from the humans and the instruments.
None of this means AI is useless for materials, and I'd push back on anyone who reads it that way. A shortlist of computationally stable candidates is a real head start, and a robotic lab that can grind through synthesis is a real accelerator. But the headline number was a measure of the model's output, not of scientific progress, and those are not interchangeable. The honest scoreboard isn't structures predicted. It's compounds made, characterized, and put to use, and that scoreboard moves slowly because reality is the bottleneck, not imagination. When the next "AI discovered X" story lands, read the verb. Ask whether the AI proposed X or whether someone in a lab confirmed X is real and matters. That one question separates a press release from a result.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.