Knowledge workers spend an average of 4.3 hours per week fact-checking AI outputs. Most of that time is wasted on the wrong failure mode.
Most people worry about AI hallucinations,that is, when AI fabricates false information. But that's not the most dangerous failure mode.
The real risk lies in what I call plausible-neighbor substitution: the AI provides an answer that is statistically close to the correct one but ultimately incorrect. Instead of inventing entirely new content, it offers:
The dosage for a related medication
The regulatory threshold from last year
A clause from an adjacent jurisdiction
Code that works on test cases but fails at scale
Why is this more problematic than hallucinations? Because it often passes casual inspection. A hallucinated citation collapses immediately once checked. In contrast, a plausible neighbor appears legitimate and survives superficial review, it's crafted to look exactly like what the correct answer should be.
The solution is simple: ask a targeted question ->
"What is the most common incorrect answer to this question, and how does your answer differ from it specifically?"
This prompts the AI to distinguish its response from its most probable error. It succeeds roughly 70% of the time; in the remaining 30%, manual verification is necessary.
I've documented this technique along with four other failure modes in a comprehensive protocol covering 25 distinct failure types across law, medicine, coding, and strategy.
The Cross-Examination Method
Top comments (0)