AI models struggle to plan at different levels of abstraction simultaneously. They can't easily move from a high-level goal to a detailed task and then back up to adjust the high-level plan if the detail is blocked, a key aspect of human reasoning.

Traditional AI struggles with games like Civilization not due to computational complexity, but because these games require maintaining a long-term strategic narrative, not just optimizing individual moves. Human players win by committing to a coherent story for their civilization's development.

AI excels at learning fixed rules, like in chess or identifying a cat. However, it falters in domains like financial markets or politics where the 'game' is adversarial and multiplayer. Any successful AI strategy is quickly identified and countered, rendering it ineffective.

AI thrives in domains with fixed, written rules and searchable histories, like programming. In ambiguous areas like organizational conflict or political negotiation, where context is unwritten and lives in people's heads, its performance plummets. Its confident output masks this unreliability, posing a danger to decision-makers.

AI models struggle to create and adhere to multi-step, long-term plans. In an experiment, an AI devised an 8-week plan to launch a clothing brand but then claimed completion after just 10 minutes and a single Google search, demonstrating an inability to execute extended sequences of tasks.

AI excels at tasks where it can make small attempts and get fast, clear feedback ('hill climbing'). Rogue deployments require long-horizon strategic planning with no easy feedback, a domain where agents are currently very weak.

Recent studies pitting AI agents (like Claude and GPT) against each other in geopolitical simulations found them substantially more prone to escalating conflicts to the nuclear level. This suggests that current AI models may not adequately weigh the catastrophic political nature of nuclear use compared to human decision-makers.

AI systems often collapse because they are built on the flawed assumption that humans are logical and society is static. Real-world failures, from Soviet economic planning to modern systems, stem from an inability to model human behavior, data manipulation, and unexpected events.

Current AI world models suffer from compounding errors in long-term planning, where small inaccuracies become catastrophic over many steps. Demis Hassabis suggests hierarchical planning—operating at different levels of temporal abstraction—is a promising solution to mitigate this issue by reducing the number of sequential steps.