What Human Judgment Still Decides | AI Learning Guide

The point of this page is not that humans are better at predicting binding. It is that the model optimizes inside a space someone else defined, and three decisions sit outside that space. First, which target is worth going after — that is a clinical and biological judgment, and it is not in the training data. Second, what counts as acceptable: a binding score is not a specification, because a therapeutic antibody also has to be stable, well tolerated, and manufacturable, and those requirements pull against each other. Third, how to read a surprising result — when a model reports a very high score for an unusual candidate, someone has to judge whether that is a real find or a case unlike the training data.

0:00 / 0:00

Auto-play

Three decisions the model cannot make for itself

The AI pipeline optimizes within a space that someone else defined. Target choice, the acceptance criteria for a candidate, and the interpretation of surprising predictions are all outside that optimization. They require knowing what the antibody is for, which is context the model does not have.

Why "binds the target" is not a specification

Suppose a model returns a candidate with an excellent predicted binding score. That score says nothing about whether the antibody stays folded in a concentrated formulation, whether it triggers an immune response in patients, or whether it can be produced at the scale a clinic needs. A human designer has to weigh those requirements against each other and decide which candidate is acceptable for the intended use — a decision the binding score cannot make.

This is not a claim that human designers are more accurate. It is a claim about which questions each side is answering: the model answers "does this candidate look like a binder," while the human answers "is this the right thing to build."