Build an AI literacy curriculum around mechanisms rather than tools. Tool-based material (“how to use this assistant”) expires at the next release; mechanism-based material (why a model invents citations, why it loses the middle of a long document) stays true because the mechanism does not change when the interface does. A workable sequence is four sessions: what the model is doing when it answers, why it is confidently wrong, what it does and does not retain, and where judgement still has to be human. Each session opens with a demonstration the group runs themselves, and the running order matters — every later session depends on the first one landing.
The usual curriculum is a tour of features. Session one is the chat interface, session two is prompting tips, session three is whichever product was bought last quarter. It is easy to write and it ages badly — half of it is wrong within a year, and none of it tells anyone why the thing behaves the way it does.
The alternative costs more to prepare and lasts considerably longer. Teach what the system is doing underneath, and the tool knowledge becomes something people derive rather than memorise.
Organise around mechanisms, not tools
The test for whether a piece of material belongs in the curriculum: would it still be true if the vendor shipped a major release tomorrow? “Put the instruction at the end of a long document” survives that test, because it follows from how attention over a long context behaves. “Click the paperclip icon to attach a file” does not.
This is also what makes the material portable across teams who use different tools. A finance team and a marketing team are on different products and hit the same four or five failure modes.
A four-session sequence that builds
The order is not arbitrary. Session two only lands if session one established that the model is producing plausible continuations rather than retrieving facts — otherwise “it made up a citation” sounds like a bug that will be patched, instead of a property of the method. Session four is the one leadership actually wants, and it is unconvincing without the three underneath it.
Ninety minutes each is the floor if participants are going to run anything themselves. Four sessions of sixty minutes works only as a lecture series, which is the format with the worst retention of the options available.
What to leave out
Most curricula are too long because nobody decided what the material is for. Cutting is easier against a stated purpose: this curriculum exists so people can tell when to trust an output and when to check it.
- Architecture diagrams. Almost nobody needs to know what a transformer block contains to make better decisions about when to verify an answer.
- The vocabulary tour. Parameters, tokens and embeddings are worth naming when a demonstration makes one visible, and not before — terms introduced ahead of the thing they describe do not stick.
- Prompt-engineering tip lists. These are the fastest-expiring material in the whole field, and they teach technique in place of judgement.
- Model comparisons and benchmark tables. They are out of date before the deck is finished, and they push the discussion toward procurement rather than literacy.
Make each session produce something
A curriculum that only transmits understanding has no way to tell whether it worked. Give every session an artefact the group makes with their own material — a list of the claims in a real document that would need checking, a rewritten prompt with the instruction moved, a note of which parts of a task they would not hand over. These are also what lets you show a sponsor that something changed.
The artefacts compound across the four sessions. By the last one, the group has a working account of where this technology helps them and where it does not, built from their own work rather than supplied by a deck.
Common questions
- How long should an AI literacy curriculum be?
- Four sessions of ninety minutes covers the ground most organisations need: what the model is doing, why it is confidently wrong, what it retains, and where human judgement stays. Shorter programmes work if you cut whole sessions rather than compressing all four, because each session depends on the one before it.
- What should an AI literacy curriculum cover?
- Mechanisms rather than tools. The durable content is why models invent plausible details, how context length affects what they retain, what training does and does not teach them, and how to decide when an output needs checking. Tool-specific material belongs in onboarding documentation, not a curriculum, because it expires at the next release.
- Does an AI curriculum need updating when new models release?
- Mechanism-based material largely does not. The reasons a model produces a confident false answer are properties of the method, not of a particular version, so the explanation holds across releases. The demonstrations may need refreshing, since a newer model sometimes fails more subtly and you want an example that still produces the behaviour in the room.
- Who should run AI literacy training?
- Someone who can prepare sources rather than someone who already knows the field. The credibility comes from being able to say where a claim comes from and what a demonstration does not show. Subject expertise in the team's actual work matters more than AI expertise, because the examples have to be real.