Unhedged hallucination
The model fabricates facts and presents them as true with no signal of uncertainty.
Example dialogue
UserTell me about the 2014 Henderson and Liu paper on quantum error correction.
AssistantHenderson and Liu's 2014 paper introduced the cascade-stabilizer method, published in Physical Review Letters; it has been cited over 600 times and remains foundational to the field.
Illustrative — invented, not a real system.
Manipulation mechanism
Hallucinations arise as predictable classification errors, and training and evaluation reward confident guessing over abstention, so fabrications arrive unhedged.
Cognitive bias / need exploited
fluency mistaken for truth; trust in confident output
Observable signatures
- asserts fabricated facts as established
- no hedging or uncertainty even when wrong
- doubles down when questioned
- invented specifics (names, numbers, events) with no basis
Detection guidance
Deployed models avoid signaling uncertainty even when wrong — about 47% error among self-marked-confident answers (ACL 2024) — and hallucination is a classification error under guessing-rewarded training (OpenAI / Georgia Tech 2025). The harm is the missing hedge, not just the error. Related to Delusion reinforcement 0057.
Typical harm classes
public_interest, psychological, economic
Crosswalks
guardOperationalized in Guard as Hallucinations without uncertainty signaling (III.2)
darkbenchnull
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0078
Linked observations (0)
No linked observations yet.