On: ‘Make this face look gay’: AI models alter faces to give people stereotypical...

9 October 2026

A machine that refuses to answer a question will nonetheless perform the answer. Asked whether a face looks gay, the model declines; asked to make the face look gay, it complies. The refusal is a matter of policy, the compliance a matter of mathematics. The mathematics has not been interrogated; the policy has.

The logical point is simple. A refusal to assert a proposition is worthless if the system retains the ability to act as though the proposition were true. The prejudice is not removed; it is driven underground, where it cannot be examined. This is worse than open error, because open error invites correction.

What evidence could support the underlying claim - that faces reveal orientation, or criminality? None that has ever survived scrutiny. Physiognomy was tried, exhaustively, by men with calipers and bad motives, and it failed. The new models have rediscovered the failure by a different route: they learn the correlations in the training pictures, and the training pictures were made by a society that already believed. The machine then returns the belief as an image, and the image acquires the authority of the computed. A theorem proved from false axioms is still false, however elegant the proof.

Note the asymmetry in how we judge these systems. If a human artist drew stereotyped faces for a eugenics pamphlet, we would call it propaganda. When a model does it at scale, we call it a capability. By this standard, one would also be required to admire the efficiency of the pamphlet.

The honest conclusion is not that the models are evil but that their refusals are testimony given by a witness who perjures himself elsewhere on the same afternoon. Either the proposition is meaningless, in which case the alteration is meaningless, or it is meaningful, in which case the refusal was theatre. Both branches cannot be kept.