← Updates

Research

Hallucination is not the only risk worth naming

Survey of recent literature on subtle, harder-to-detect failure modes: source laundering, citation drift, and confident omission.

Research deskOct 09, 20269 min read
Out-of-focus type on a printed page suggesting subtle errors.
Out-of-focus type on a printed page suggesting subtle errors.

"Hallucination" became the shorthand for AI failure in newsrooms because it is vivid and easy to demonstrate. It is also a narrow frame. Recent literature points to at least three failure modes that are harder to spot and, in editorial settings, arguably more dangerous.

Source laundering: a model surfaces a claim that originated in a low-credibility source, but presents it without attribution in fluent prose. The journalist who quotes the model has, in effect, quoted the original source without knowing it.

Citation drift: a model produces a correct-looking footnote that points to a real paper, but the claim attached to the citation is not in the paper. A reader who checks the citation finds a real document and assumes the claim is supported.

Confident omission: a model summarises a long document and silently drops the qualifying clauses. The summary is technically accurate but materially misleading. This is the failure mode most likely to slip past a careful editor.

None of these are reasons to stop using AI in newsrooms. They are reasons to design verification workflows that look for more than just made-up facts.