R0042/2026-04-01/Q003/H2¶
Statement¶
Anti-sycophancy work exists at AI model developers and researchers, but no enterprise deploying private AI for business operations has documented sycophancy reduction as a design goal.
Status¶
Current: Supported
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | Anthropic has comprehensive anti-sycophancy program: evaluation since 2022, Petri tool, 70-85% reduction in Claude 4.5 |
| SRC02-E01 | AI research firms achieving sycophancy reduction — developer activity |
| SRC03-E01 | Policy framework treats anti-sycophancy as vendor obligation |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| None | No evidence contradicts this characterization |
Reasoning¶
All three sources support H2: anti-sycophancy work is active and significant, but it is concentrated among AI model developers (Anthropic, OpenAI, DeepSeek), AI research firms (Cognition Dynamics, AI Innovate per SparkCo), and policy analysts (Georgetown Law). None of these represent enterprises deploying private AI for business operations with anti-sycophancy as a stated design criterion.
Relationship to Other Hypotheses¶
H2 captures the nuanced reality that H1 (enterprise case exists) overstates and H3 (no anti-sycophancy work anywhere) understates.