The best description I've heard of AI in science came from a biologist, not a vendor. She'd been using an AI system on protein structures and put it this way: it's like having a postdoc who's read every paper ever written, never sleeps, and isn't competing with me for lab space. She still had to ask the right question. But the gap between asking and knowing had gone from months to minutes.
That's the real story, and it's better than the one you read in press releases. AI isn't the scientist. It's the absurdly well-read, infinitely patient junior collaborator who accelerates the scientist. Get that distinction wrong and you'll either over-invest in fantasy or miss the actual gains.
What's genuinely changed
Protein structure prediction is the headline, and for once the headline is earned. DeepMind's AlphaFold, in 2021, effectively cracked a fifty-year-old problem: predict a protein's 3D shape from its amino acid sequence. Work that consumed years per structure now returns predictions in minutes. The whole field reorganised itself around it.
Materials discovery is moving the same way. New battery electrolytes, catalysts, solar coatings used to come from grinding trial-and-error. AI screening lets researchers propose candidates computationally and only test the promising ones — DeepMind's GNoME proposed a large batch of new crystal structures in 2023, a chunk of which look stable and are being tested now.
Literature review is the unglamorous one with the highest hit rate. Nobody can read the field any more; it's too big. Tools that help researchers find relevant papers and surface contradictions don't replace careful reading — they make sure the careful reading starts in the right place. Experimental design benefits too: where the design space is enormous, Bayesian optimisation and similar methods let a lab run fewer, more informative experiments. And in drug discovery, generative models propose candidates that get computationally screened before anything touches a wet lab — Insilico Medicine pushed an AI-discovered fibrosis compound into Phase II at a pace that would have been unthinkable a decade ago.
What it is plainly not doing
It is not replacing scientists. Hypothesising, designing experiments, interpreting results, arguing about them in front of hostile colleagues — that's the irreducibly human core, and accelerating the edges doesn't touch it.
It is not producing original conceptual leaps on its own. AI finds patterns and proposes hypotheses. The genuinely creative move — recognising that what looked like noise was signal, or wiring together two fields nobody thought were related — still comes from people. That may change. It hasn't.
And it is not fixing bad science. Train on flawed data, get flawed predictions. A literature tool that over-weights popular papers just industrialises popularity bias. These systems make a good scientist faster. They make a sloppy one faster at being sloppy.
Three things to stay sceptical about
Reproducibility is a live problem. AI-assisted papers are sometimes impossible to reproduce because the model version, training data, or random seeds were never recorded. Treat AI-derived results in a paper with the same scepticism you'd apply to any undocumented method.
Access is splitting the field. The strongest models need GPUs and datasets only well-funded institutions can muster, and that's opening a fresh have/have-not divide in research. Open-source releases help. They don't close it.
Over-claiming is rampant. "AI makes breakthrough discovery" almost always unpacks to: a human asked a sharp question, AI narrowed the search space, and humans did the experimental work that produced the result. Read the paper, not the press office.
If you run a lab and you're weighing this up
Start with literature review — it's the least exciting line item and the highest return; it can save a PhD student months. Put the AI tools next to skilled people: the groups making the biggest gains are the ones where the computational and experimental sides actually talk to each other daily. Never skip validation — an AI prediction is a hypothesis, and the bench keeps the final word. And document your pipelines obsessively, because future-you and future reviewers will need to reproduce exactly what you did.
If I were twenty-five and choosing what kind of scientist to be, I'd learn to work alongside these tools without delay — not because they'll replace scientists, but because the scientists who collaborate well with them will out-produce the ones who don't by a wide margin. And in research, productivity just means getting to chase more of the questions you actually care about. That's a trade worth making.
