Stop Teaching AI Tools, Teach the Words First

Every week another district, university or training provider announces an AI literacy initiative. Most of them are built on the same assumption: students and staff lack skills, so we should teach them tools. Prompt workshops. Tool bake-offs. A certificate at the end.

I think the assumption is wrong, and I have spent the better part of a year — most of it arguing with myself before I built anything — accumulating evidence that it is wrong.

I run a plain-language AI encyclopedia. It now holds more than a thousand concepts, in English and Chinese, and every entry has the same two halves: a technically accurate definition, and one everyday analogy that makes the definition land. People arrive at it from search engines, from social media, from other people forwarding a single card. What they ask for, over and over, is not “how do I write a better prompt.” It is: what does fine-tuning actually mean. Is an AI agent different from automation. Why does everyone keep saying RAG. What is a context window, and why does it matter that it filled up.

Those are not skill questions. They are vocabulary questions. And they come one layer earlier than everything we are currently teaching.

Why Vocabulary is the Real Bottleneck

A learner who cannot follow the words cannot do any of the things we actually care about. They cannot evaluate a vendor’s claim, because they cannot tell which parts of it are meaningful. They cannot read a news story about model releases and judge whether it changes anything for them. They cannot read a job posting that asks for “experience with agentic workflows” and decide whether they are qualified or being gatekept. They cannot participate in a faculty meeting about AI policy, so they stay quiet, and the policy gets written by whoever is loudest.

Jargon is not a neutral by-product of a fast-moving field. In practice it works like a moat. Opaque vocabulary keeps consultants billable, keeps courses sellable, and keeps products hard to compare against one another. Nobody has to conspire for this to happen; the incentives do it on their own. Education is one of the few places positioned to drain that moat, and we are mostly not doing it, because “teach 50 words” sounds less impressive than “AI transformation program.”

The good news is that the fix is cheap. When we swap one piece of jargon for one concrete analogy, comprehension is close to instant — not for some students, for almost all of them. That tells you the barrier was never intelligence, aptitude, or even math background. It was that nobody had translated the sentence.

What We Learned Trying to Automate the Translation

Here is the part that surprised me, and I think it matters for anyone building or buying EdTech.

Our encyclopedia is produced by an automated pipeline. A model drafts each entry daily; a second model reviews every draft against an editorial standard; entries that pass go live, and the rest are held for a human — me — to fix. It has been publishing daily since early summer, so I have a large sample of exactly where machine-generated explanation succeeds and where it fails.

The definitions are almost never the problem. Models are excellent at the fact half: accurate, compact, appropriately hedged. What gets held back, again and again, is the analogy. Not because it is wrong, but because it is tired.

A recent example, from this week: two unrelated entries drafted in the same batch — one on the race to lock up training data, one on how vendors stagger the release of new model capabilities — independently reached for a traffic metaphor. Neither was wrong on its own. Together they were a pattern, and the review stage held both, because a reader meeting them on the same day would learn nothing from the second one. Earlier the same month, an entry about robot maintenance technicians was blocked for comparing the job to an emergency room doctor — factually fine, utterly forgettable, because everyone has heard “X is like a doctor for Y.” Another was held because it explained a general design-to-manufacturing workflow using a tailor, which quietly narrows a multi-industry concept down to clothing. All were rejected on the same grounds: the explanation was technically correct and pedagogically useless.

If you extend that finding to the classroom, it says something uncomfortable about the current wave of AI-assisted content tools. Generating material at volume is now trivial and nearly free. Generating material that makes a concept click is still the hard, human, editorial part — and it is the only part that was ever scarce. Any EdTech roadmap that treats explanation as a volume problem is optimizing the half that was already solved.

3 Things I Would Do Differently This Year

• Teach the vocabulary before the tools. Fifty words, taught concretely, will do more for a student’s ability to navigate this field than a semester of prompt technique. Prompting advice expires in months. “What a model’s context window is, and why it fills up” does not.
• Test comprehension by asking for an explanation, not an output. If a student can explain fine-tuning to a classmate in their own words, using their own analogy, they understand it. If they can only produce a good result from a tool, you have measured the tool. This is also the one assessment format that current AI assistance does not trivially defeat, because a borrowed analogy is obvious the moment you ask a follow-up question.
• Treat the glossary as living infrastructure, not a unit. The vocabulary genuinely changes every quarter; terms that did not exist eighteen months ago are now in job descriptions. A glossary published once as a PDF is dead on arrival. Whoever owns AI literacy in your institution needs a standing process for adding words, not a one-time curriculum project.

None of this requires a platform purchase, which is probably why it is underrepresented in EdTech planning. It requires deciding that the goal is comprehension rather than tool fluency.

An AI-literate graduate is not one who prompts well. It is one who can follow the conversation, ask a sharp question, and tell when someone is selling them fog.