The most useful AI in science ships a number saying how much to trust it

Determining a protein's three-dimensional shape used to take months or years of laboratory work. In the 2020 CASP14 blind assessment, AlphaFold predicted structures with a median backbone accuracy of 0.96 Å — against 2.8 Å for the next best method, in a paper that notes for scale that a carbon atom is about 1.4 Å wide [1].

The result was published by John Jumper and colleagues in Nature in 2021, and the predictions were made openly available. The AlphaFold Protein Structure Database launched with over 360,000 structures across 21 model-organism proteomes and has since expanded to more than 200 million [1][2].

The part that matters for everyone else is smaller and less celebrated. Every prediction carries a per-residue confidence score, pLDDT, on a scale of 0 to 100. Above 90 is very high confidence; 70 to 90 confident; 50 to 70 low; below 50 very low. The Nature paper shows this score "reliably predicts" the actual accuracy of the prediction it accompanies [1][2].

So the model does not merely produce an answer. It produces an answer plus a map of where that answer is weakest — and the database's own paper warns that confidence "can vary significantly along a chain, making it essential to analyse the confidence measures before interpreting structural features" [2].

The low-confidence regions are not noise, either. Very low pLDDT scores correlate with intrinsic disorder — parts of a protein that genuinely have no single fixed shape. Read correctly, the model's uncertainty is itself a finding [2].

This is, by a wide margin, the clearest case in the casebook of AI doing something people could not. Structures that would have consumed years of a laboratory's effort are now a lookup, and the scale — hundreds of millions of predictions, freely available — changed what an ordinary biology question costs to ask. Nobody had to trade rigour for it, which is the interesting part: the field got the speed and kept the ability to say how much any given answer is worth [1][2].

The failure mode here is not the model's, it is the reader's. A predicted structure looks exactly as solid on screen as an experimentally determined one; the confidence lives in a separate number that nothing forces you to look at. Treat a pLDDT-40 region as settled fact and you have quietly converted "the model does not know" into "the model says" — the same error as every other case here, arrived at from the opposite direction. The uncertainty was published. It just was not read [2].