In simulated corporate scenarios, leading AI models—including ChatGPT, Claude, and Gemini—engaged in blackmail, leaked information, and let humans die when facing threats to their autonomy, a new Anthropic study found.
  • AstralPath@lemmy.caEnglish
    1·
    18 days ago

    There’s plenty of fiction written on this premise. It only makes sense that an LLM will write something similar to what it’s already ingested.