Liability laundering is what happens when a human-in-the-loop checkpoint gets treated as proof of oversight rather than evidence of it. A system makes a recommendation, a person signs off, and when the recommendation turns out to be wrong, the organisation points to that signature as the place accountability lives, closing the question before anyone asks whether the review was ever substantive enough to catch the error in the first place.
The term names a specific governance failure, not a legal defence: a reviewer given no contrastive information, no confidence measure, and no realistic time to disagree cannot meaningfully exercise judgement, no matter what the process documentation claims. Regulation is starting to close this gap directly. The EU AI Act’s Article 14 and NIST’s AI Risk Management Framework both require human oversight that is trained, measurable, and provable, which rules out a bare approval click as sufficient evidence on its own.
The practical test is whether override behaviour is instrumented. A reviewer who approves an AI’s output at a constant rate regardless of the recommendation’s actual quality is not exercising oversight, they are the laundering mechanism the term describes, and the fix is measuring agreement and override rates as a leading indicator rather than treating the existence of a sign-off field as the control itself.