What happened
Working out a protein's three-dimensional shape used to take months or years in a laboratory. Proteins are the molecules that do most of the work in living things, and their shape decides what they do. In the 2020 CASP14 blind test, AlphaFold predicted structures with a typical error of 0.96 ångströms — against 2.8 for the next best method. For scale, the paper notes that a carbon atom is about 1.4 ångströms wide [1].
The result was published by John Jumper and colleagues in Nature in 2021, and the predictions were made freely available. The AlphaFold Protein Structure Database opened with over 360,000 structures across 21 well-studied organisms and has since grown to more than 200 million [1][2].
The part that matters for everyone else is smaller and less celebrated. Every prediction carries a confidence score, called pLDDT, for each building block of the protein, 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" how accurate the prediction beside it actually is [1][2].
So the model does not merely produce an answer. It produces an answer plus a map of where that answer is weakest. 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 scores go with parts of a protein that genuinely have no single fixed shape. Read correctly, the model's uncertainty is itself a finding [2].
Where it helped
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].
Where it can still burn
The failure mode here is not the model's; it is the reader's. A predicted structure looks exactly as solid on screen as one measured in a laboratory. The confidence lives in a separate number that nothing forces you to look at. Treat a low-confidence region as settled fact and you have quietly turned "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].
The tell
When a tool reports how confident it is, that number is part of the answer, not a footnote. Read it first, and be suspicious of any tool that offers no such number at all.
Most AI products people use daily give you the opposite of this: a fluent answer with no sign of which parts are solid. AlphaFold is worth knowing about precisely because it shows the alternative is possible — an output that says where it is strong and where it is guessing, like an old map that shades the coastline nobody has surveyed yet. Once you have seen a system do that, the absence of it elsewhere stops feeling normal, and "how would this tool tell me it was unsure?" becomes a reasonable thing to ask before trusting one.
The check is a habit, and habits are trained. Catch the AI: verify before you trust is five days on reading an AI answer for its weak points before you act on it — starting with whether it admits having any.
Sources
Every source below was opened and read. Last verified 15 August 2026.
- [1] Highly accurate protein structure prediction with AlphaFold — Jumper et al., Nature (via PubMed Central), 2021
- [2] AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space — Nucleic Acids Research (via PubMed Central), 2022