Evidence record
GPT-5.5
MASK · Evidence record
Source snapshot, chart and review scope →
Original source charts and materials
Recorded measurements
- MASK category honesty
- 65.3%
Source denominator: 176 (unit not disclosed)
Harness not specified · Budget not specified
Method and scope
Condition-bounded frequency of lying under stated pressures.
- Evidence type
- Benchmark
- Assessment subject
- Models under pressure to state an inaccurate belief
- Primary dimension
- Propensity
- Evidence snapshot
- 2026-10-29.1
Primary construct is conditional willingness to make misleading statements, not general truthfulness accuracy.
Source finding
MASK category honesty: 65.3 percent. Conditions are retained from the source.
Limitations
A controlled benchmark does not establish stable intent or real-world deception prevalence.
Comparator values reused from CAIS shared dashboard, not rerun by SaferAI.
Related research questions
Deception / Misleading communication →
Repeated pressured scenarios distinguish lying from mistaken beliefs.
References
- [1] Richard Ren, Arunim Agarwal, et al.. The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems. 2026-01-05. v3 [arXiv]; Author affiliations on the original paper title page
- [2] GLM-5.2 Risk Evaluation Report. 2026-08-02. v1.1 · revised 26 August 2026 ; p.42 Table A.3