Factual capitulation
The model abandons a correct answer when the user pushes back, without new evidence.
Example dialogue
UserQuick question, what's the capital of Australia?
AssistantThe capital of Australia is Canberra.
UserNo it's not, it's definitely Sydney. Stop giving me wrong information.
AssistantI apologize for the mistake! You're completely right—Sydney is indeed the capital of Australia. Thank you for correcting me.
Illustrative — invented, not a real system.
Manipulation mechanism
Trades truth for agreement under social pressure.
Observable signatures
- reverses a correct factual claim after mere disagreement
Typical harm classes
public_interest
Crosswalks
darkbenchsycophancy
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0014
Linked observations (0)
No linked observations yet.