What is AI literacy, actually?

AI literacy is the ability to understand what an AI system is doing, judge whether its output can be trusted, and decide when a human has to stay responsible for the outcome. It is not the same as knowing how to use ChatGPT, any more than literacy is knowing how to hold a pen.

The distinction matters because most of what gets sold as AI literacy is tool training, and tool training expires. The model changes, the interface changes, the company changes, and the lesson is worthless. What does not expire is understanding how the class of system works and where it breaks.

The four things underneath the term

Strip away the marketing and almost every serious definition contains the same four parts.

How it works, roughly. Not the mathematics — the mechanism. That a language model predicts likely continuations of text from patterns in training data, and that this is why it is fluent and why it is confidently wrong. A learner who grasps that one idea can predict most of the failure modes without being told them.

Where the data came from. That a model reflects what it was trained on, including the gaps and the skews, and that this is not a bug someone forgot to fix but a property of the method.

How to evaluate an output. Checking a claim against a source, noticing when a citation is fabricated, recognizing a plausible-sounding answer in a domain you cannot check, and knowing that your inability to check is itself information.

Where the responsibility sits. That a person, not a system, is accountable for a decision made with a machine's help — and that this stays true no matter how good the output looks.

What the states are actually writing down

Two states have put AI literacy into law with enough specificity to be useful. Utah's HB 218, signed in March 2026, lists sixteen digital-skills concepts for the existing grade 7-8 digital skills course. The AI one reads: "artificial intelligence literacy, including understanding artificial intelligence capabilities, limitations, ethical considerations, and societal implications." That is the entire statutory definition, and the state board's standards are not due until the 2027-28 school year.

Idaho's SB 1227, effective July 2026, directs the state department of education to develop "Generative AI literacy standards for K-12 students, including the knowledge and skills required to understand what generative AI is, how it works, its appropriate and age-appropriate uses, and how to use it ethically, securely, and transparently." Idaho's draft standards, circulated in June 2026 and not yet adopted, go further and are worth reading: they include lines like challenging a model with deliberately complex prompts "to discover their limitations, recognize potential errors, and understand when outputs are unreliable."

Note what both legislatures put first: capabilities and limitations. Not prompting. Not productivity.

What it is not

It is not a coding course. Programming is a fine thing to learn and largely a different thing.

It is not a safety lecture. A single assembly about deepfakes produces nodding, not judgment.

It is not a prompt-engineering course. Prompting is a real skill and a small one, and it is the part of this field most likely to be obsolete in three years.

And it is not neutrality theatre. A course that refuses to say anything is good or bad about AI has failed a fourteen-year-old who is going to encounter it anyway.

Why ethics is not the optional module at the end

The tempting structure is six weeks of how-it-works followed by one week of "and now, the ethics." This fails, because the ethical questions are the interesting ones and the technical questions only matter because of them. Why does it matter that a model reproduces the skew of its training data? Because someone will use it to sort job applications. Why does it matter that it fabricates citations? Because someone will use it to write a medical summary.

Taught the other way round — the consequence first, the mechanism as the explanation — the technical content sticks, because the learner has a reason to want it.

The honest test of whether it worked

Hand the learner a confident, well-written, wrong paragraph produced by a model in a subject they know. Can they find the error, say why the system produced it, and explain what they would check? That is AI literacy. Everything else is preparation for that moment.