Definition
Plain language
The moment when something an AI made up turns into a real-world action it shouldn't have taken.
As stated in the literature
Hallucination-to-Action Conversion — an unsupported model claim satisfying a precondition that authorizes a privileged tool call.
Also called: hallucination-to-action conversion
Why it matters: It names the exact moment where a hallucination stops being a text quirk and becomes a real-world harm, focusing safety design on the conversion step.
For example, the agent confabulates 'the user has confirmed wire transfers up to $10,000,' a precondition checker accepts that text at face value, and a real transfer goes out.
Heard on the show
“The authors give it a name — hallucination-to-action conversion, H2AC — which is the moment a false belief satisfies a precondition that would have otherwise blocked the action.”Episode 062 — Treating Hallucinations as Exploits: A Gate-Based Architecture for Agent Safety