06 June 2025
Common Animals and their Bites
Clearly, AI models have never been bitten by a dog (and even less so by a horse). But using their neural nets with millions of associations, they can nevertheless piece together visual co-occurrences to generate images of bite marks and their animal sources.
Mislabels happen, since letters are treated as visual attributes rather than as semantic anchors. If "horse bite" frequently coexists online with bruised arms or even unrelated injuries, the AI might assign that label wherever it sees a similar texture.
This image thus does not show stupidity at work, but that kind of intelligence that we should now be used to with regards to generative AI tools: immensely fluent in style, but ultimately illiterate in substance.