When a bot "says" something it never said: how a rumor becomes news

Thesis: the failure is not in the model. It is in the newsroom that publishes a chat screenshot as a source. Until fictional content carries a mandatory label, every "proof from a chatbot" will be ammunition for disinformation.

The case looked like a typical internet vir. Recordings showed a Chinese AI assistant allegedly telling people to cook their slippers so they would "return to their original size." The material spread widely, and some news outlets and institutional accounts treated it as proof of dangerous chatbot errors. When the story reached the manufacturer's communications department, the company issued a denial: the model had said nothing of the kind.

The investigation turned up something more important than the mistake itself. The recordings were staged as a joke, built on a popular saying that "shrunk shoes can be boiled back to size." The creator admitted he deliberately steered the model toward a wrong answer and recorded the result. He did not label the material as fiction or entertainment. The effect was predictable. Viewers, journalists among them, saw a short clip without context and took it as the product's actual behavior. As the company's official explanation put it: "one rumor became news."

The same pattern repeated twice, in fact. The same public relations team had earlier denied another story: a ranking of the "ten most popular stars," generated by a blogger with AI and attributed to the product as its answer. In both cases the mechanism was identical. There was no system failure. There was a deliberately induced answer, cut from its context and passed on as fact.

Why is this not just a Chinese problem? Because the same pattern is driving regulatory decisions around the world. Article 50 of the EU AI Act requires that generated and manipulated content be labeled in a machine-detectable format, and the implementation timetable shows how hard it is to reconcile the speed of the market with the rigor of the law. Public opinion research shows that audiences already want labels: 81 percent of Germans consider transparent labeling of sources and training data important, and 54 percent check model outputs for errors.

The conclusion is twofold. First, the obligation to label staged and fictional material should cover not only synthetic content but also recordings that put a model at the center as the main evidence. Second, and more important, a newsroom cannot treat a chat window screenshot as a source. If we publish a screen as proof, we are obliged to check whether the answer was induced under pressure and whether its full context exists. That is a basic craft requirement, and in the era of generators it has become critical. A rumor does not become a fact because someone recorded it. It becomes one because someone published it without checking.