Thomas Germain became a champion hot-dog eater without enduring the traditional inconvenience of eating the hot dogs. He published a fabricated ranking on his website, awarded himself first place, and waited. Within a day, he reported, several leading AI services were repeating his achievement. When some suspected a joke, he added “this is not satire.” Apparently, nonsense benefits from administrative reassurance.
Bruce Schneier highlighted the experiment as a warning about poisoning AI’s information supply. Strictly speaking, it demonstrated manipulation of web-grounded answers, without establishing that any model had been retrained. But it exposed the essential vulnerability: publish something, have a machine retrieve it, and watch invention acquire the manners of knowledge.
Now replace the hot dogs with music, paintings, films, reviews, biographies, or historical footage. The question becomes considerably less amusing, although the business models remain inventive.
AI-generated culture is not inherently poisoned. A synthetic song can be enjoyable; a generated image can serve a legitimate artistic purpose. Poison enters when material deliberately corrupts learning, or when unreliable content and fabricated signals are accepted as trustworthy evidence. Machines are perfectly capable of processing a work’s pixels while misunderstanding everything about its origin.
Some creators have decided to exploit that gap.
Nightshade demonstrated how carefully altered images could corrupt associations learned by image generators. Its researchers proposed the technique as protection against scraping that ignores creators’ objections. To the developer, this is poisoned training material. To the artist, it resembles installing a burglar alarm after discovering that the neighbours have reclassified the studio as an all-you-can-eat buffet.
Music has its own defensive research. HarmonyCloak introduces subtle perturbations designed to obstruct learning from instrumental recordings while preserving their quality for human listeners. Its published experiments showed effectiveness against the models tested, rather than universal protection against every future generator. The principle is nevertheless striking: a composition can remain available to an audience while becoming an unhelpful meal for a machine.
Such interventions make the politics of “poison” unavoidable. Sabotaging a competitor and defending one’s work against unwanted extraction can use related techniques. The disagreement begins before training, with who believes they are entitled to consume what.
The same opening could serve less sympathetic purposes: fabricated reviews, false attributions, or manipulated descriptions could steer what systems associate with an artist, a genre, or a political event.
Elsewhere, contamination is simply good business.
In March 2026, Michael Smith pleaded guilty in a scheme involving hundreds of thousands of AI-generated songs and billions of automated streams. Here was a remarkably efficient entertainment industry: software supplied the repertoire, software supplied the audience, and a human collected the money. The case established streaming fraud, not model poisoning. But it illustrates how apparently meaningful cultural signals can be manufactured.
If such listening figures later guide recommendations, dataset selection, or model optimisation, the fraud can acquire a second career as evidence of human taste. A system may learn that people adore certain music when the actual enthusiasts are unattended accounts.
Accidental contamination needs even less organisation.
Imagine an invented musician’s biography being copied into a catalogue, summarised in a review, and retrieved by a research assistant. Three sources now agree, although all descend from one fabrication. Or consider a generated film clip detached from its production notes and later described as documentary footage. A training pipeline could learn from the caption and images without recovering that missing distinction.
The same problem can arise without a factual lie. Suppose a music model repeatedly trains on generated tracks selected for familiarity and immediate appeal. Unusual rhythms, difficult passages, and awkward experiments may receive fewer invitations to the next dataset. The machine’s idea of culture could become an increasingly polished account of what another machine found easy to produce.
No villain is required. Everyone can be meeting their quarterly targets.
Self-optimising AI raises the stakes because contaminated material may influence the criteria for improvement. Imagine a system generating examples, judging their quality, and training its successor on the winners. If the examples contain planted distortions and the evaluator shares those distortions, “better” can mean becoming more proficient at reproducing the original contamination.
A falsehood can become a lesson, then a scoring rule. At that point, correcting it may look like declining performance. The student gets marked down for disagreeing with the poisoned answer sheet.
Research on model collapse demonstrates one related danger: indiscriminate training on recursively generated data can progressively lose information about the original distribution. That is a different mechanism from deliberate poisoning, but both make the quality of the learning foundation consequential. Neither means that every synthetic dataset inevitably fails.
Indeed, research comparing training workflows has found that retaining original real data while accumulating synthetic examples can avoid collapse in the settings studied. Recursion can support improvement when reliable evidence and meaningful checks remain available. It also gives inherited mistakes repeated opportunities to survive.
The practical task is to preserve provenance, distinguish genuine responses from manufactured engagement, and test improvement against evidence outside the system’s own circle of approval. Consent matters too: treating every accessible artwork as free training material creates an obvious incentive for creators to make that material hostile.
A machine can generate a song, applaud it, review it, and teach its successor that it was a cultural milestone. Before calling this progress, we should establish whether anyone attended the concert.
Preferably someone who also checked the hot dogs.




No comments yet