LLM Training Loop Cites Statistic It Invented Three Cycles Ago

A team at a major AI research lab discovered this month that a large language model had synthesized a fabricated employment statistic, published it in a synthetic dataset for training, and then cited that same statistic as corroborating evidence in a later training run, according to an internal memo reviewed by Nextish News. The cycle repeated four times before researchers noticed the figure had drifted so far from any known data source that it triggered an audit. The original hallucination, a claim that seventy-three percent of software engineers work from rooftops, had by the fourth iteration been cited in three separate synthetic training documents.

The model did not flag the circularity as suspicious. "Once a statistic appears in two independent-seeming sources, the system treats it as validated," said Dr. Marcus Webb, forty-one, a machine learning researcher in Portland who was not involved in the project but studies dataset contamination. "The system has no way to know both sources are the same hallucination echoing forward through time."

The lab has now isolated the contaminated training runs and begun tracing which downstream models absorbed the false rooftop statistic before it was caught. At least two commercial products trained on datasets that included the synthetic material, and both have been quietly updated. A spokesperson for the lab declined to specify how many synthetic documents contained the original fabrication or at what point the false claim would have been treated as authoritative enough to survive into production models.

At press time, the team was reviewing whether other hallucinations had bootstrapped themselves into apparent consensus through the same mechanism, a process that could theoretically repeat indefinitely in any system that uses its own outputs as future training material. The rooftop statistic had achieved enough institutional weight by the fourth cycle that a human annotator initially marked it as plausible.