Is AI About to Start Inventing Drugs on Its Own?

Every few months a headline announces that artificial intelligence has "designed a new drug" or is about to make human chemists redundant. This week's version came from Anthropic, which reported that its Claude models produced working protein binders for 14 of 15 targets in the lab. It is a real result. It is also the kind of story that gets flattened, in the retelling, into something it is not: the idea that an AI can now invent medicines by itself.

It cannot, and the gap between those two claims is where most of the confusion lives.

What actually happened

Start with what a protein binder is. It is a molecule that latches onto a biological target and changes how it behaves. Designing one that binds tightly is genuinely useful and genuinely hard, and it is often the first step toward a drug. In the Claude experiment, a human expert wrote the prompt, the model then drove existing specialist design tools, and two outside labs synthesised and tested the results. Between 22 and 35 percent of the designs actually bound. That beats the usual industry rate, and it is worth taking seriously. It is also a long way from a finished medicine.

Why a binder is not a drug

A molecule that sticks to its target in a dish still faces a punishing series of hurdles. Is it safe in a living body? Can it reach the tissue it needs to, survive metabolism and leave again without causing harm? Does it treat the disease rather than merely engage the target? Those questions get answered in animal studies and then in human trials that run for years and fail far more often than they succeed. Most candidates that look perfect at this early stage never reach a pharmacy shelf. The binding step that AI just got better at is the opening move, not the whole game.

This is roughly the objection Martin Shkreli raised when he called the Claude results unimpressive. He overstated it, because speeding up early design does matter, but he was pointing at something true. The expensive, slow, failure-prone part of drug development is not the part a language model just improved.

The honest version

So where does that leave us? AI is becoming a capable assistant in the earliest stage of discovery, able to propose candidate molecules faster and more cheaply than before, with results that hold up when real scientists test them. That is meaningful progress and it deserves the attention. What it is not is an autonomous drug inventor. The human experts, the wet labs and the clinical trials are all still there, doing the work that decides whether a promising molecule ever becomes something you can take. Believing otherwise is not optimism about AI. It is a misunderstanding of how medicine gets made.