Worst AI image fails: the ones with real consequences

Anyone can post a picture of a hand with seven fingers. What follows are failures that produced a consequence in the world: a cancelled event, a newspaper apology, a terminated contract.

They are described rather than shown, because none of these images belong to this site, and because in every case, the story is better than the picture.

The Wonka event that could not exist

Willy’s Chocolate Experience, Glasgow, February 2024.

An unlicensed event inspired by Charlie and the Chocolate Factory was promoted with lavish AI-generated imagery: candy-coloured halls, elaborate installations, dreamlike sweeping spaces. Families paid for tickets on the strength of it.

The copy was generated too, and nobody read it. The promotional material offered “cartchy tuns” and “a pasadise of sweet teats”, alongside genuinely invented words, “catgacating”, “exarserdray”. These were published, live, on the website.

What ticket holders found was a sparsely decorated warehouse: a few plastic props, a small bouncy castle, and the AI-generated backdrops printed out and pinned to the walls. The gap between the promise and the room was so wide that parents called the police. Officers attended. The event was cancelled within hours of opening and became a global news story.

The images were not the failure. Printing them and pinning them to a wall as the actual attraction was.

Ten books nobody wrote

Chicago Sun-Times summer reading list, May 2025.

A 64-page syndicated special section carried a summer reading list of fifteen titles. Ten of them did not exist.

They were attributed to real, living, well-known novelists (Isabel Allende, Taylor Jenkins Reid, Brit Bennett) with plausible titles and full plot summaries. Readers went looking for them and found nothing.

The paper’s own newsroom had not produced it. The section came from an outside syndication supplier, written by a freelancer who had used AI in developing the stories without disclosing it. The supplier said it has a strict policy against AI-generated content and terminated the relationship.

The Sun-Times then fact-checked the rest of the section and found the problem was not confined to the book list. Every one of the ten stories with named sources contained errors or claims that could not be independently verified.

That is the part worth sitting with. The fake books were caught because books are checkable. Everything else in that section had the same origin and nobody would ever have noticed.

The citations that pass every test except existing

Courts, 2023 to now.

Fabricated legal citations have been identified in over a thousand court proceedings worldwide, catalogued in a public database maintained by a legal researcher at HEC Paris.

The reason this is the most instructive failure of all: the fakes are formally perfect. Correct reporter format. Plausible volume and page numbers. Real judges’ names attached to opinions they never wrote. In the 2023 Mata v. Avianca case, one fabricated opinion was presented as an Eleventh Circuit decision; the court described its reasoning as “gibberish”, but the citation looked flawless.

There is no visual tell. There is no stylistic tell. The only way to catch it is to go and look for the document, which is exactly what nobody does, until a judge does.

The pattern

None of these were caused by a model being asked something difficult. Every one came from ordinary output nobody checked, in a workflow where checking was somebody’s job and no longer happened.

The technology performed as designed. The proofreading did not.