AI literacy for the people who sign: what boards and senior officials need to know
Article 4 of the EU AI Act makes AI literacy an obligation. For the people who approve deployments, literacy is not knowing how transformers work. It is knowing which questions to ask and what a bad answer sounds like.
Literacy programmes tend to over-teach mechanism and under-teach judgement. A director does not need to explain attention heads. A director needs to know, when a management team presents an AI deployment, what to ask and how to recognise an answer that should stop the meeting.
Seven questions
What decision does this system make or influence, and about whom? What was it tested against, by whom, and were the thresholds set before the results were seen? Who approved it, and did they build it? What happens when it is wrong — how does a person appeal, and how fast? What would make us withdraw it, and who has the authority to do so tonight? What changed since the last approval — model, data, vendor, context? And: can someone outside the team reproduce the evidence?
Seven bad answers
"The vendor has certified it." "Accuracy is 94 percent" with no per-group breakdown. "The data science team is confident." "We would retrain it." "It is only a recommendation" for a recommendation that is followed 99 percent of the time. "Nothing material has changed" from the people who would rather not re-test. And silence in response to the reproducibility question.
Literacy for signatories is the ability to recognise an answer that should stop the meeting.
Making it stick
Put the seven questions into the approval template so they are asked every time, not just after the training session. Rotate real cases through the board — an anonymised assessment, a near miss, an appeal that reversed a decision. And measure literacy by behaviour: how often approvals are deferred pending evidence, and how often the answer to "who approved it?" is a name rather than a team.
