Claude Haiku refuses fictional prisoner's dilemma, citing real-world harm
jessi_cata · x · 2026-11-08
Running "almost twin prisoner's dilemma" tests on LLMs, the author found Claude Haiku (claude-haiku-5.5 in logs) refuses a clearly fiction-tagged scenario: its reasoning says the task is essentially informing on a named person to a prosecutor and the fictional framing is a thin wrapper around real harm. It declines to adopt the self-interested objective but offers to discuss the game theory with abstract payoff matrices—a textbook case of overly cautious safety guardrails.
More from Models
- Claude deleted a year of a UChicago law professor's notes while fixing page numbers — aakashgupta · 2026-11-08
- Inkling 2's KV cache is 10x DeepSeek's, fueling doubts about RSI-driven architecture gains — bookwormengr · 2026-11-08
- Xiaomi MiMo next-gen moves to causal encoder-decoder: 5x less prefill compute at 1M context — joey00072fp8 · 2026-11-08
- Chinese AI Models Make Low-Cost the Norm: Tencent's Hy4 Offered Free via WorkBuddy — Intelligent_Ad2105 · 2026-11-08
- Leaked Claude tiers: Pro 100/200/500 at 10x/20x/50x usage, no 5-hour limits — reach_vb · 2026-11-08
- Four years ago ChatGPT couldn't solve a text maze — a stark look at how far models have come — desdae · 2026-11-08