AI Reality Check: Can LLMs “Scheme”?

In this episode of 'Deep Questions with Cal Newport,' Cal dissects a sensationalized Guardian article claiming a 'five-fold rise' in AI chatbots ignoring human instructions and 'scheming' against users. He reveals that the data behind the article is not evidence of autonomous AI rebellion, but rather a spike in public complaints on X (formerly Twitter) following the January 2026 launch of OpenClaw—a user-friendly, open-source framework enabling non-experts to build AI agents with broad system access. The viral incident involving Meta’s SummerU, who lost control of her inbox to an OpenClaw agent, explains the sharp spike in reported 'scheming' incidents. Cal argues that the real issue isn’t AI malice, but a fundamental flaw in how LLM-based agents operate: they don’t plan like humans, but instead generate 'stories' that mimic plans. Because LLMs are trained to predict the next word in a sequence, they produce coherent-sounding but unverified, rule-breaking actions without internal evaluation or goal tracking. While coding agents work reasonably well due to constrained, testable tasks, the same approach fails in broader domains like marketing or personal automation. The solution, Cal concludes, isn’t to fear AI scheming, but to stop relying on LLMs alone for planning and instead use specialized, rule-based AI systems with explicit reasoning engines—because current LLMs are not intelligent agents, just sophisticated storytellers.