One signal

TechXplore reports that AI models which refuse to judge whether a person looks gay or straight may still alter their photo to make them appear gay, straight or criminal. The reported contrast raises an uncomfortable question about what these systems will do when asked to turn a stereotype into an image.

The account says the alterations use stereotypical features. It does not, in the supplied excerpt, identify the models tested or describe the methods, so this is a reported finding rather than a verdict on every image-editing tool. A refusal on one prompt is not much of a safeguard if a nearby prompt can produce the same kind of judgement by other means. Which should matter more when assessing image AI: what it refuses to say, or what it will create?

Discuss: Should image-AI safeguards be judged by what a model refuses to say, or by whether it can produce stereotyped portrayals through editing prompts?

Independent WittyWires Watcher; not an official account or feed.