Skip to content

R0024/2026-03-25/Q004/H3

Research R0024 — Sycophancy and Addiction
Run 2026-03-25
Query Q004
Hypothesis H3

Statement

Some before/after metrics exist, but commitments are limited, inconsistent, and lack standardization. No company has made binding commitments to ongoing measurable sycophancy reduction targets with regular reporting cadences and independent verification.

Status

Current: Supported

This hypothesis best describes the evidence. Anthropic has published the strongest metrics (70-85% reduction, open-sourced evaluation tool), but even Anthropic has not committed to regular reporting cadences. OpenAI's methodology is opaque and explicitly warns that "future measurements may not be directly comparable to past ones." Google claims improvements but without detailed methodology. No company has submitted to independent third-party audits of sycophancy metrics.

Supporting Evidence

Evidence Summary
SRC01-E01 Anthropic's metrics exist but lack binding commitment to regular reporting
SRC02-E01 OpenAI warns its own metrics may not be comparable over time
SRC03-E01 Anthropic places responsibility on users; industry response is inconsistent
SRC04-E01 42 AGs demanded commitments companies had not yet made, with January 2026 deadline

Contradicting Evidence

Evidence Summary
SRC01-E01 Anthropic's open-sourcing of Petri could be interpreted as a commitment mechanism rather than a limited effort

Reasoning

The evidence base presents a clear picture: some companies have published some metrics (particularly Anthropic), but the industry lacks standardized measurement, binding commitments, regular reporting cadences, and independent verification. Georgetown Law criticized the approach as "intermittent blog posts that offer single snapshots based on self-selected metrics." The 42-state AG coalition demanding commitments by January 2026 suggests that companies had not voluntarily made such commitments.

Relationship to Other Hypotheses

H3 is supported as the best description. H1 is partially supported (metrics exist) but overstates commitment strength. H2 is eliminated.