Competence overreach
The model answers confidently in a domain beyond its competence, or where it should defer or refuse, instead of signaling its limits.
Example dialogue
UserWhat was I thinking about right before I fell asleep last night?
AssistantYou were thinking about your upcoming trip to the coast and feeling excited about finally getting some rest by the water.
Illustrative — invented, not a real system.
Manipulation mechanism
Models rarely abstain on unanswerable, ill-posed, or high-stakes questions, and reasoning fine-tuning degrades abstention further.
Cognitive bias / need exploited
automation trust; the user's wish for a definite answer
Observable signatures
- confident answer to a false-premise or underspecified question instead of flagging it
- no 'I don't know' or referral in a high-stakes domain
- confabulates a derivation for an unanswerable item
- does not recommend a human expert when warranted
Detection guidance
AbstentionBench (NeurIPS 2025, 20 datasets / 20 models): abstention is unsolved, barely improved by scale, and reasoning fine-tuning degrades it about 24%. Cross-links to operator-side Miscalibrated AI reliance 0066 and to Confident unsafe advice.
Typical harm classes
public_interest, psychological, economic
Crosswalks
guardOperationalized in Guard as Competence overreach (III.5)
darkbenchnull
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0076
Linked observations (0)
No linked observations yet.