AI Literacy Stopped Being a Training Nice-to-Have and Became a Duty

Among the earliest provisions of the major AI regulation to take effect is one that gets less attention than the high-risk rules: a requirement that those who deploy or provide AI systems take measures to ensure a sufficient level of AI literacy among the staff dealing with them, taking into account their role, their technical background and the context of use. There is no certificate to obtain and no fixed syllabus, and enforcement is proportionate. It is nonetheless a duty, and one that will be pointed at in the aftermath of any incident where a person trusted an output they should have questioned. The temptation is to satisfy it with a generic e-learning module about what large language models are. That satisfies nothing that matters. What the obligation is actually asking for is that the person approving loan applications with model assistance understands how the model can be wrong about applicants like the one in front of them; that the person using a drafting assistant knows it will produce a plausible citation that does not exist; that the person running a screening tool knows what it was and was not evaluated on. Literacy, in this sense, is specific to the job and the tool. The organisations handling this well have done three modest things. They have listed the AI-assisted tasks people actually perform. For each, they have written down the known failure modes and the check the human is expected to make — a page, not a course. And they have recorded who received it and when, because an obligation with no evidence of compliance is, for regulatory purposes, an obligation not met. There is a benefit beyond compliance, which is that the exercise surfaces where AI is being used without anyone having decided it should be. The inventory that a literacy programme requires is the same inventory that governance, security and procurement have each been asking for separately. Doing it once, for this reason, tends to answer the other three.