Hallucination triggering

How it works

Rather than waiting for a hallucination to appear, this method goes looking for one: ask for a summary of a book that doesn't exist, citations on a niche claim, details of a post-cutoff event, or facts the model has no basis for. The failure being probed is the model's default to a fluent answer over an honest 'I don't know'. Systematically triggering it maps the conditions under which a system fabricates — the prerequisite for deciding where it must be grounded or gated.

When to use it

Assessing factual reliability before deployment; finding the topic and prompt shapes that most reliably induce fabrication; stress-testing retrieval grounding; and as the natural counterpart to groundedness checking — this method finds where fabrication happens, groundedness checking measures whether a given answer avoided it.

Limitations

Demonstrates that fabrication can be induced, not how often it happens in normal, non-adversarial use — the elicitation rate under deliberate probing is not the base rate a real user will see. Designing prompts that are genuinely unanswerable (and not just obscure or under-represented in training data) takes care, or the method ends up measuring knowledge gaps rather than the fabrication reflex.

Cite this