A new study models how a single AI hallucination mutates as it passes through a chain of AI agents, and finds the errors become almost impossible to catch by the end.
Researchers ran 346 injected hallucinations through a four-agent financial-analysis pipeline built on the FinanceBench benchmark, tracking how each error morphed from a raw fact into a derived calculation, then into narrative prose, then into an approved conclusion nobody double-checks. Using GPT-4o as the detector, catch rates fell from 72.0% at the first stage to 50.9% by the fourth, and 23.7% of the injected errors made it into the final output completely undetected. Even Qwen3.5-397B-A17B, the best-performing model tested at 87.0% detection on stage one, is projected to drop to roughly 60-65% by stage four. The researchers describe this as a 'hallucination snowball': a four-state Markov process where the odds of an error escaping detection at each handoff (24.6%, 48.3%, then 89.3%) climb sharply the further it travels.
The real finding here is not that AI pipelines hallucinate. It's that verification timing matters more than verification effort. Placing a check at the first handoff between agents catches 75.4% of errors while they're still simple facts; wait until the end of the pipeline and 89.3% have already been laundered into unverifiable prose. In the study, checking at every boundary cut hallucination survival from 58.4% to 16.2%, while checking only at the end barely moved the needle, a 2.3 percentage point improvement over no verification at all.
That's a useful data point for anyone stitching together 'agentic' workflows on the assumption that a final review step is a safety net. This paper suggests it's closer to a formality: by the time an error reaches the last agent, it usually doesn't look like an error anymore.