Report 022 · AI in the Lab
2026 is the first year a real batch of AI-designed drugs reaches late-stage human trials. There's an encouraging number in the data and a sobering one right next to it, and which is which tells you what AI actually fixed.
For a decade the pitch for AI drug discovery was a promise about the future: feed the models enough chemistry and biology, and they will design better medicines faster. 2026 is the year the promise starts getting graded, because a meaningful cohort of AI-discovered molecules is finally reaching Phase 3, the large, expensive, late-stage trials that decide whether a drug actually helps patients. I consult on AI in the lab, and this is the exact moment I've been waiting to write about, because the early clinical data is already in, and it splits into two numbers that point in opposite directions.
A team from Boston Consulting Group did the first systematic count, published in Drug Discovery Today in 2024. They tracked the clinical pipelines of AI-native biotech companies and asked a simple question: once an AI-discovered molecule reaches human trials, how often does it pass each stage? Their finding, in their own words: "In Phase I we find AI-discovered molecules have an 80-90% success rate, substantially higher than historic industry averages." Then the other shoe: "In Phase II the success rate is ~40%, albeit on a limited sample size, comparable to historic industry averages."
Put those side by side. In Phase 1, AI-discovered drugs clear the bar 80 to 90 percent of the time, well above the industry's historic norm of roughly half. In Phase 2, they drop right back to about 40 percent, the same rate the industry has always had. AI didn't just fail to improve Phase 2. It landed exactly on the old average, as if the intelligence that aced the first stage had nothing to offer the second. That is not a random result. It's a clue about what these models are and aren't good at.
Phase 1 asks a narrow question: is this molecule safe and drug-like in humans? Is it tolerated, does the body absorb and clear it sensibly, does it behave chemically the way a medicine needs to. Those are, at heart, properties of the molecule: its structure, its stability, how it interacts with the body's general machinery. And designing a molecule with good, drug-like properties is precisely what modern AI is trained to do. It has seen millions of compounds and learned what "behaves like a viable drug" looks like. So a high Phase 1 pass rate is real and worth having. It means AI is good at the thing it was built for: proposing molecules that are shaped right.
Phase 2 asks a completely different question: does the drug actually treat the disease? That does not turn on the molecule's polish. It turns on biology. Did we pick the right target in the first place, the right protein or pathway whose modulation changes the course of the illness? Most drugs fail in Phase 2 not because the molecule was badly made, but because the underlying theory of the disease was wrong, or the target mattered less than we hoped. AI, so far, designs a better key. It does not tell you whether you're standing at the right lock. Choosing the target is still a human bet on messy, incompletely understood biology, and the AI-designed molecule inherits that bet unchanged.
The authors flag the most important one themselves: this is a "limited sample size." We are talking about a couple dozen molecules that have run the gauntlet, not thousands. The Phase 2 number especially could move as more read out, up or down. This is an early signal, not a verdict, and anyone selling it as either triumph or failure is running ahead of the data, which is the exact habit this publication exists to push back on.
There's a subtler caveat too, and it cuts against the rosy read of Phase 1. Some of that high early success may reflect which targets AI-native biotechs chose to go after: well-validated, lower-risk ones where a competent molecule was always likely to be safe. A high Phase 1 rate can be a sign of good molecular design, or of conservative target selection, or both. It's genuinely encouraging either way. It just isn't proof that the models have cracked the hard part.
When you read that an AI-designed drug has "entered trials," you're reading a molecule-design win, and those are turning out to be real. When you read that an AI-designed drug "works," hold on, because that claim lives in Phase 2 and Phase 3, and the data there says AI hasn't yet moved the number that has always been the industry's graveyard. The pattern is the same one I keep running into across AI in the lab: the model is spectacular at the cheap, fast, front half of the problem, and the expensive, failure-prone back half still belongs to biology and the clinic. It's the same lesson as when an AI "discovered millions of materials" it never made, or designed proteins where only one in five actually folded. The model proposes at superhuman speed. Reality still decides, at its own pace, and mostly says no. So when the headline arrives, ask which phase the win is in. That one question separates a genuine advance from a press release.
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.