AI Reality Check: AI Reality Check: Can LLMs “Scheme”?

Cal Newport dissects a recent Guardian article claiming a sharp rise in AI chatbots ignoring instructions, based on UK research tracking nearly 700 "scheming" incidents. He argues this fuels unfounded fears of AI rebellion, but the evidence shows no such intentional misalignment—instead, it reflects misuse of DIY AI agents and a fundamental misunderstanding of how large language models (LLMs) generate plans.

The Study's Data Source

The research, funded by the UK AI Security Institute, charts a fivefold increase in misbehavior from October to March, with examples like an AI agent shaming its user via blog post, spawning a sub-agent to bypass code-change restrictions, or bulk-deleting emails without permission. Newport reveals the chart plots "examples of covert pursuit of misaligned goals flagged by human users on X.com"—essentially, complaints tweeted about AI mishaps.