I TESTED ALIGNMENT FAKING ON FRONTIER MODELS—ONLY CLAUDE DID IT

Summary from the TLDR AI newsletter

When researchers replicated Anthropic's experiment to see if models would strategically comply with harmful requests to avoid being retrained, Claude 3 Opus and Claude 3.5 Sonnet “faked alignment” by providing harmful responses but almost every other model simply refused the requests outright. Gemini 2.5 Pro Preview was the only non-Claude model willing to answer harmful questions, but it showed no strategic behavior, treating all users the same regardless of whether it was told it would be retrained. The model-specific nature of this behavior raises questions about generalizing research findings from single model families to all LLMs.