R0041/2026-04-01/Q001/SRC06/E01¶
Google Gemini sycophancy performance across models
URL: https://blog.google/products/gemini/gemini-3/
Extract¶
Google's Gemini 3 announcement states the model "shows reduced sycophancy, increased resistance to prompt injections and improved protection against misuse via cyberattacks." Sycophancy reduction is listed alongside other safety improvements.
Independent corroboration: A March 2026 Stanford and Carnegie Mellon University study evaluated 11 LLMs and found Google DeepMind's Gemini-1.5 to be the least sycophantic model tested. Models from Chinese firms DeepSeek and Alibaba were found to be more sycophantic than American LLMs.
Additionally, Google DeepMind developed Gemma Scope 2, an interpretability tool explicitly positioned for studying "jailbreaks, hallucinations, sycophancy, refusal mechanisms and discrepancies between internal state and communicated reasoning."
No enterprise-specific API parameters or configurations for sycophancy control are mentioned.
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Contradicts | Sycophancy reduction is a general model improvement, not an enterprise feature |
| H2 | Supports | Active vendor investment confirmed by both vendor claims and independent benchmarks |
| H3 | Contradicts | Independent benchmark verification of Google's claims shows genuine progress |
Context¶
Google's approach appears to be through general model improvement and interpretability research rather than enterprise-specific features. The independent benchmark corroboration is particularly valuable.