Most "AI literacy" sessions on campus fall into one of two traps. Either they turn into a demo reel of clever prompts that everyone forgets by the following week, or they become a warning about plagiarism that leaves people more anxious and no more capable. Neither one teaches literacy. Literacy is the ability to use a tool well, judge when its output can be trusted, and know when to put it down.
I've taught AI literacy to faculty and staff, and the version that lands is narrower and more practical than people expect. You don't need everyone to understand how a model is trained. You need them to build a small set of habits they'll actually use on a Tuesday afternoon. Here's what's worth teaching.
Start with a working mental model, not the math
People make better decisions once they have a rough picture of what these tools are doing. The useful version fits in a sentence: a language model predicts likely text based on patterns it learned from huge amounts of writing, so it's fluent by design and accurate only by accident.
That one idea explains most of the behavior people find confusing. It's why a chatbot can write a smooth paragraph and still invent a citation. It's why it sounds equally confident whether it's right or wrong. Once colleagues internalize "fluent, not factual," they stop treating a tidy answer as a correct one, which is the single most important shift you can teach.
The goal isn't to make people distrust AI. It's to move them from "it sounds right" to "let me check the part that matters."
Teach a few questions, not a list of rules
Rules go stale as the tools change. Questions don't. A short set of questions people can ask about any AI output travels well across tools and years:
- What did it actually use to answer? Did it draw on documents I gave it, or on its training data? Grounded answers are checkable; ungrounded ones are guesses in a confident voice.
- What would it cost me if this is wrong? A brainstormed list of workshop titles and a figure in a grant budget carry very different risk. Match your scrutiny to the stakes.
- Can I verify the claim in under a minute? If yes, verify it. If a claim can't be checked at all, that's a reason to be cautious, not a reason to trust it.
- Should this data even be here? Before pasting anything in, ask whether the input is sensitive. Student records, unpublished research, and anything under an IRB protocol don't belong in a public tool.
Four questions is enough. People remember four; they don't remember fourteen.
Make the "don't paste that" rule concrete
The data question deserves its own moment, because it's where good intentions go wrong fastest. Well-meaning staff paste a student email, a draft manuscript, or a spreadsheet of survey responses into a public chatbot to save five minutes, without registering that the material just left the institution's control.
A test I give people is simple: would you be comfortable if this text showed up on the open internet? If the answer is no, it doesn't go into a public tool. For that kind of material, the institution needs a private, document-grounded option instead, which is a separate conversation about infrastructure, not something an individual can fix in the moment. Naming that boundary clearly in a workshop prevents more harm than any policy PDF nobody reads.
Practice on real work, in the room
AI literacy doesn't transfer from slides. It transfers from doing. The sessions that change behavior spend most of their time on hands-on practice with tasks people genuinely have: summarizing a long report, drafting a rubric, turning messy notes into an outline, comparing two policy documents. For each one, the pattern is the same: try it, then check it, then talk about where it helped and where it quietly misled you.
That last step matters most. When someone catches the model inventing a source or flattening an important nuance, and they catch it themselves in front of the room, the lesson sticks in a way no cautionary slide can match.
Address disciplines and roles separately
A one-size workshop underserves everyone. A writing instructor worried about student use, a grants administrator drowning in proposal text, and a librarian fielding data questions have different needs. You don't have to build a separate curriculum for each, but you should swap in examples that match the room. The core habits stay the same; the tasks people practice on should be theirs.
What "success" looks like
Don't measure a literacy program by how excited people are as they leave. Excitement fades. Measure it by what they do next: whether they pause before pasting sensitive text, whether they check a claim before forwarding it, whether they can tell you why a given task is a good or bad fit for AI. Those quiet habits, repeated across a campus, are the whole point.
The takeaway
AI literacy in higher education isn't a technology topic. It's a judgment topic wearing a technology costume. Teach a plain mental model, a handful of durable questions, a firm line about sensitive data, and plenty of hands-on practice, and you'll give people something better than enthusiasm: the ability to decide for themselves, case by case, whether the tool in front of them is helping or just sounding like it is.