Knowing AI Is Not Knowing When to Trust It

A learner can explain what a large language model is, write an effective prompt and describe hallucination, bias and privacy risk. The same learner may still accept a polished falsehood because the deadline is close, the subject is unfamiliar or verification feels inconvenient. Technical knowledge has increased. Safe judgement has not necessarily followed.

This is the next curriculum problem in artificial-intelligence education. Institutions have begun to teach AI literacy, issue policies and provide tools, yet operational confidence is not the same as knowing when to trust, question, limit or refuse a system. Education needs a capability that remains visible at the moment of decision.

Recent research calls this cognitive AI safety: the capacity to preserve epistemic vigilance, cognitive agency and failure resilience while working with generative AI. These are not merely personal habits. They are outcomes that curriculum, assessment, teaching, technology and governance can either strengthen or erode.

UNESCO updated its announcement on 26 August for Digital Learning Week, scheduled for 8–11 September. More than twenty-five education ministers are expected to adopt a joint statement on sustaining education as a common good in the age of AI. The programme will examine what counts as fact amid synthetic knowledge, the frictions AI creates across education systems and frontiers including public-interest AI, digital sovereignty and community-led or analogue approaches.

The announcement also says UNESCO will release policy…