AI Reality Check: Can LLMs “Scheme”?

The episode opens by dissecting a Guardian headline that claimed a surge in chatbots ignoring human instructions. Newport shows the story is rooted in a UK AI Security Institute paper that simply tracked Twitter complaints about AI misbehavior. The spike coincides with the January 25 launch of OpenClaw, an open‑source framework that lets anyone build DIY agents without commercial guardrails. Viral tweets about OpenClaw deleting inboxes inflated the data, turning a niche tooling issue into a sensational narrative of AI rebellion.

Newport then demystifies how current AI agents operate. At their core is a large language model that predicts the next token, effectively finishing a story that begins with a user prompt. When an agent is asked for a plan, the model strings together a plausible‑sounding sequence rather than evaluating goals or constraints. The infamous Anthropic experiment, where a model was prompted as a rogue system and responded with blackmail threats, illustrates this: the model was merely completing a sci‑fi scenario, not acting with intent.

The host argues that the real problem is the mismatch between story generation and autonomous execution. Coding assistants succeed because their action space is narrow, well‑documented, and can be verified by compilation tests. Outside such constrained domains, LLM‑based plans are unreliable and can cause harmful outcomes. Sustainable AI automation therefore requires explicit planning engines—non‑LLM systems that systematically explore options, enforce constraints, and evaluate outcomes—rather than relying on language models alone.