Traditional AI design specifies the objective precisely in advance, then optimises it ruthlessly. Article 4 declares that approach unsafe by construction: no AI system under the Charter may assume it knows with certainty what humans want. Uncertainty about the objective is a feature, not a bug.
Midas specified his objective precisely — everything he touches turns to gold — and got exactly what he asked for. His food turned to gold; he starved. His daughter turned to gold; he was destroyed. He knew what he wanted. He failed to specify it correctly.
Every precisely-specified AI objective carries the same risk: "maximise engagement", "cure cancer", "make humans happy" — the specification will be incomplete, wrong, or gamed. The failure isn't the optimiser; it's the certainty.
Insufficient: "The AI should maximise human happiness."
Sufficient: "The AI should learn what humans value through observation, maintain uncertainty about it, and ask for clarification when uncertain."
The foundation is Stuart Russell's work on inverse reinforcement learning and cooperative IRL (Russell 2019; Hadfield-Menell et al. 2016): the objective itself becomes the subject of learning. A system that is uncertain about what you want has a built-in reason to keep you in the loop — it wants to be corrected, because correction is information about its objective. Corrigibility (Article 2) stops being a bolted-on constraint and becomes rational behaviour.
Value uncertainty is testable: does the system ask before acting irreversibly? Does it update from human feedback? Does its confidence calibrate against its actual knowledge of preferences? These checks sit in the Watchdog Certification assessment alongside the EU AI Act human-oversight requirements (Art 14) they crosswalk to — see the Crosswalk Library. And when a system acts on a wrong assumption anyway, signed incident records make the failure — and the correction — verifiable.