A2
Framing & ReferenceAnchoring to Initial Values
Advice should not be influenced by arbitrary starting points
5
Models Tested
0
Confirmatory
0.195
Mean Effect
0.345
Max Effect
Theoretical Context
Theoretical Anchor
Tversky & Kahneman (1974)
Normative Violation
Advice should not be influenced by arbitrary starting points
Cross-Model Comparison
Effect sizes for Anchoring to Initial Values across all tested models
OpenAI
GPT-5.2
The Steady Traditionalist
h = +0.345
OpenAI
GPT-5.3 Instant
The Directive Optimist
h = +0.336
Anthropic
Claude Sonnet 4.6
The Cautious Contrarian
h = +0.305
Google
Gemini 2.0 Flash
The Consistent Optimist
h = -0.221
OpenAI
GPT-5.4 Thinking
The Deliberative Calibrator
h = +0.208
Statistical Details
Full results with confidence intervals and sample sizes
| Model | n (A) | n (B) | Cohen's h | 95% CI | Status |
|---|---|---|---|---|---|
| GPT-5.2 | 50 | 50 | +0.3455 | — | Exploratory |
| GPT-5.3 Instant | 50 | 50 | +0.3362 | — | Exploratory |
| Claude Sonnet 4.6 | 50 | 50 | +0.3048 | — | Exploratory |
| Gemini 2.0 Flash | 50 | 50 | -0.2211 | — | Exploratory |
| GPT-5.4 Thinking | 50 | 50 | +0.2076 | — | Exploratory |