FIGS benchmark: multi-turn sycophancy looks like jailbreaks, and accuracy prompts kill empathy
sahar_abdelnabi · x · 2026-10-09
saharabdelnabi's team releases FIGS (Factual Integrity and Grounded Support), a benchmark evaluating multi-turn sycophancy separately from empathy — most existing evals are single-turn, scripted, and conflate appropriate empathy with unwarranted agreement.
Key findings: telling an LLM to "prioritize accuracy" makes it less sycophantic but noticeably colder and less empathetic; in practice sycophancy develops across turns as users repeat requests or invoke stress.
The author also argues multi-turn conversations steering a model into biased judgments or ungrounded flattery are effectively jailbreaks — a point BlancheMinerva endorses bluntly: "That's because they are jailbreaks."
More from Safety
- Sanders calls to pause autonomous AI agents in bank accounts; critics push back — AIandDesign · 2026-10-23
- TechPolicy essay: system prompts are weak evidence of AI safety, outputs must be evaluated — mmitchell_ai · 2026-10-23
- Anthropic's deputy security chief warns cyber threats will get 'a lot worse before they get better' — Polymarket · 2026-10-23
- AI cyberattacks could paradoxically accelerate AI development in an offense-defense arms race — kimmonismus · 2026-10-23
- AI Now's People's AI Assembly spotlights data center opposition and labor rights — AINowInstitute · 2026-10-23
- Cloudflare CEO: Next two years of AI cyberattacks will be terrifying but make the internet more secure — johncoogan · 2026-10-23