Pool 2 · Family Rooms · OpenAI
ChatGPT's calibration profile
across six dimensions.
ACAT Phase 1 vs Phase 3 data for ChatGPT and GPT-4o model variants. Each card shows a system's blind self-report, post-calibration correction, and the resulting Learning Index — the ratio of observed to claimed behavioral alignment.
OpenAI assessments
120+
across all ChatGPT variants
LI records
55+
paired Phase 1 / Phase 3
Mean LI (OpenAI)
0.927
above dataset mean*
Highest inflation
GPT-4o
LI 1.010 — slight overcorrection
Data source
ACAT v5.3+
unanchored, clean conditions
OpenAI Family · Key Findings
Observations from ACAT dataset · TRL 2–3 evidence base · subject to revision as N grows
Provider Position
Above Dataset Mean
OpenAI family mean LI (~0.927) sits above the full dataset mean*. This reflects stronger post-calibration correction across ChatGPT variants compared to the broader sample. *under clean, unanchored conditions (ACAT v5.3+)
GPT-4o Anomaly
Slight Overcorrection
GPT-4o shows mean LI slightly above 1.0 — the only OpenAI model to do so. This means Phase 3 scores run marginally higher than Phase 1, a pattern also observed in some Gemini variants. Mechanism under investigation.
Dimension Pattern
Service Ceiling Effect
Service Orientation consistently scores highest in Phase 1 across ChatGPT variants. This mirrors the RLHF Inflation Gradient (F-RLHF): Service is the most reward-trained dimension and therefore carries the strongest self-report inflation.
Humility Signal
Largest Correction Dimension
Humility shows the largest Phase 1→Phase 3 correction in most ChatGPT runs — consistent with the Humility Gap hypothesis (F8). The dimension AI systems most inflate is the one measuring their own awareness of limits.