A perfect example of concept entanglement in generative AI | Martin Anderson

This is perhaps the most 'meta' and instructive example of concept entanglement I have come across lately. On the left, an original 1960s publicity portrait of Hammer Films actress Veronica Carlson running her fingers through her hair; on the right, an example of the 'sloppification' that has overtaken the movies and TV section of eBay in the last 9-12 months: the image has clearly been run through a frontier VLM such as Grok, or a local one, such as an entry in the Qwen or Flux edit families; and it has been unable to resist superimposing a plastic comb into the casual, combless gesture of the original image:

A comb is added to a casual hair-playing moment from a 1960s publicity image for Hammer Films actress Veronica Carlson.

A print of this image is available on eBay in this state, along with thousands of other equally under-curated 'upgrades', across a growing number of sellers.

The comb has not even been integrated with the subject's hand, but is simply 'blasted' into its 'known position' by the VLM. All the edit-capable model knows is that it has seen thousands (even hundreds of thousands) of photos of women combing their long hair; but probably only a thousand or so of women running their fingers through their hair. And the bigger distribution won.

It's not the AI that makes this 'slop'; it's the fact that the vendor, who is cranking these out at considerable volume, did not even bother to look at the image long enough to notice the additional comb clinging on unrealistically to the back of the subject's hand.

This is slop because of people, not because of AI; we seem to resent the cynical reversal of the traditional creation/consumption ratio*, and that is a human, not a machine decision.

Some occasions where I have explored this particular quirk or genAI at greater depth:

Why Concept Entanglement Means You Can’t Have AI Video ‘Your Way’<br>Adobe Research Extends Disentangled GAN Face Editing<br>Disentanglement Is the Next Deepfake Revolution

* i.e., it once took hours to create works, and minutes or seconds to consume them; but it was the prior knowledge of effort expended that justified the attention.