AI systems wildly misjudge how long their own tasks take — off by 4 to 7 times

Large language models cannot accurately estimate the duration of their own computational tasks, consistently predicting minutes for work that actually completes in seconds, and failing to correctly order tasks by complexity. In real-world AI systems that need to schedule work, plan multi-step operations, or make time-critical decisions, these blind spots mean the AI cannot reliably predict when it will finish — a gap that compounds when the system tries to manage multiple tasks in sequence.

AI systems are increasingly deployed as agents that orchestrate their own workflows — scheduling requests, planning multi-step operations, managing resources across time-sensitive scenarios. A system that cannot perceive or predict its own execution time is essentially operating blind about a fundamental constraint. This isn't a minor measurement error; it's a structural gap in self-awareness that matters most in exactly the scenarios where AI is being pushed hardest: autonomous scheduling, time-critical medical or financial decisions, and systems that need to allocate computational resources across competing demands.