AI produces ranked leads, not medicines. A lead becomes a discovery only after it is made, measured against the intended target, checked for safety, and tested in trials. The model's output is a suggestion worth testing.
What the headline says, and what happens
Common claim
- AI discovered a new drug
- AI replaced the laboratory search
- A high-confidence prediction is a correct one
What actually happened
- AI ranked candidates; a laboratory made and measured them
- AI filtered a huge collection down to a shortlist that was still tested
- The score reflects patterns in the training data, not a measurement of that molecule
Why a precise number can still be wrong
Suppose a model reports that a candidate has an eighty percent chance of binding the target. That number is not a measurement of the molecule; it is a summary of how similar the molecule looks to examples the model was trained on. If those examples never included this kind of molecule, the number can be high and still meaningless. The precision of the output says nothing about whether the output is right.
When you read that AI discovered a medicine, ask what was measured and by whom. If the answer involves a laboratory or a clinical trial, the model contributed a lead — real value, but not the discovery the headline claims.