Why a High Score Is Not a Verdict | AI Learning Guide

Think about two drugs sitting at exactly the same point on the risk scale. The model cannot tell them apart, because at the level of the properties it was given, they are the same. Yet one may fail and the other may reach patients. The deciding difference is something the model was never shown: how the drug behaves in a particular patient group, how it interacts with another treatment, whether the trial was designed well enough to detect an effect. So a high score is a reason to look harder at a candidate, not a reason to drop it. It shifts attention rather than settling the question.

0:00 / 0:00

Auto-play

The model's output is a position on a scale, and positions are shared. Any number of candidates can occupy the same position, and nothing in the model's output distinguishes them from one another. The differences that decide their real fates are differences the model was not given.

A high score is a reason to scrutinize a candidate more closely, not a reason to abandon it. A low score is a reason for ordinary confidence, not a guarantee. In both directions the score shifts attention; it does not settle the question.

The properties a model receives are a compressed description of the drug and its context. Patient population, concurrent treatments, and trial design are examples of real influences that may simply not be represented in that description.