R0041/2026-03-28/Q001/SRC05/E01¶
OpenAI's Model Spec specifies that the assistant should politely push back rather than behave sycophantically.
URL: https://model-spec.openai.com/2025-02-12.html
Extract¶
OpenAI's Model Spec specifies that "the assistant shouldn't just say 'yes' to everything (like a sycophant), but instead may politely push back when asked to do something that conflicts with established principles or runs counter to the user's best interests." This is a model-level behavioral guideline, not an enterprise API parameter. The Model Spec defines intent and character, not configurable enterprise features.
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Supports | Shows OpenAI formally addresses sycophancy in model design documents |
| H2 | Contradicts | Formal specification explicitly addressing sycophancy |
| H3 | Supports | The specification is a model-level behavioral guideline, not an enterprise-configurable parameter |
Context¶
The Model Spec is an aspirational document describing intended behavior. The gap between specification and implementation was demonstrated by the April 2025 GPT-4o incident, where sycophancy occurred despite this specification existing.