Teach AI Literacy, Not Tool Training

A great deal of what is labelled AI literacy is product training: which button, which menu, which phrasing works in this interface this month. It is useful for a term and worthless afterwards, because the interface changes and the phrasing stops mattering as models improve.

Literacy is the part that transfers — knowing what kind of thing a system is, predicting where it will fail, and judging when its output can be relied on. That knowledge survives every product change, and it is what makes someone a competent user of a tool that has not been built yet.

Four domains, only one of which is operational

Knowing what the system is. A model that predicts likely continuations is a different kind of object from a database. Learners who grasp this stop being surprised by confident fabrication, because they can see it follows from how the thing works.

Knowing when to trust it. Which tasks it is reliable for, which it is not, and how to check. This is the core practical skill and it is mostly about verification habits.

Knowing the human context. Where training data comes from, whose labour built it, whose perspectives are over- and under-represented, and who is affected by deployment decisions.

Operating a tool. Genuinely useful and by far the smallest and most perishable part. It belongs in the course; it should not be the course.

Most curricula invert this, spending the majority of their time on the last domain because it is the easiest to write and to assess.

The mental model that explains the failures

Learners arrive with one of a few intuitions, and moving them up this ladder does more than any list of tips.

"It looks things up." Explains nothing, and every fabricated citation is inexplicable and shocking.

"It is a very good autocomplete." Better, and it starts to explain fluency without accuracy — but it undersells what these systems do.

"It produces the most plausible continuation given everything it has seen." This is the productive level. From here a learner can derive the failure modes rather than memorising them: plausible-sounding fabrication, confident errors on rare topics, sensitivity to how a question is framed, and better performance on things that appear often in text than on things that do not.

"And it was shaped afterwards by human preferences." The refinement that explains why models are agreeable, why they hedge in particular ways, and why they sometimes prefer sounding helpful to being accurate.

Reaching the third level is the single highest-value hour in an AI literacy course. Everything else gets easier afterwards, because the learner can predict rather than recall.

The demonstration that lands: ask a model something genuinely obscure and watch a fluent, confident, wrong answer appear. Then ask the same question three different ways and compare. Fluency and accuracy visibly come apart, in a way no explanation achieves.

Sequence it by development, not by grade

Early years need one idea: a computer made this, and computers can be wrong. Concrete, no mechanism.

Middle years can hold the notion of learning from examples, and can grasp that examples carry the biases of whoever collected them. Comparing outputs and spotting differences is age-appropriate and effective.

Later school years can work with the prediction model, evaluate sources, and discuss data provenance and consent seriously.

Higher and professional education can go into the mechanism, evaluation methodology, and the systemic questions — labour, environmental cost, concentration of capability.

Across all levels, teach verification as a habit rather than as a topic. The question "how would I check this?" asked every time is worth more than a unit on hallucination.

Choose activities that outlive the tool

The test for an activity: if the specific product disappeared next term, would it still teach something?

Activities that pass include comparing several systems' answers to the same question and explaining the differences; finding the error in a supplied confident wrong answer; tracing a claim to a primary source; examining who is represented in a dataset; and debating a real deployment decision with actual stakeholders.

Activities that fail are the ones about prompt syntax for one interface, or the features of a particular product.

One caution worth naming for staff: literacy about AI is not the same as being pro- or anti-AI. The goal is accurate calibration — neither the belief that the output is authoritative nor the belief that it is worthless. Both miscalibrations lead to bad decisions, and students arrive holding one or the other.

Practice

Take an existing lesson and re-sort its objectives into the four domains. Most will land in the operational one. Rewrite so that at most a quarter sits there.

Then build the demonstration: pick a topic obscure enough that you can verify the answer and the model probably cannot. Run it live, three phrasings, and have learners write down what they observe before you explain anything. Their observations are the lesson; your explanation just names it.

Common mistakes

Teaching the interface. It changes, and the knowledge leaves with it.

Listing failure modes without the model. Memorised lists do not generalise to a failure that is not on the list.

Advocacy in either direction. The goal is calibration, and students can tell when they are being sold something.

Treating verification as a unit. It is a habit, practised every time, or it is nothing.