Iro AIIro AI Blog
Most AI literacy guidance stops at "have a policy." A policy is not a curriculum. Here is a sequence you can actually teach, five activities that need no software, and an honest answer to the cheating question.
Editorial note: Iro AI is our app. Where it appears below we say so, and it has to earn its spot by the same criteria as everything else.
Teach judgment, not tools. The specific products your students use will change every few months. The skill that survives is knowing whether an answer can be trusted, and what to do when it cannot.
That reframing solves the biggest practical problem in this topic: you do not need to be an expert in any particular AI product to teach it well. You need to be good at asking "how would we know if this were wrong?" — which is a skill most teachers already have and already teach, in source analysis, in lab work, in close reading.
Everything below assumes a high school context. Most mainstream AI tools set a minimum age of 13, and many require parental consent below 18, so check the terms of any tool before you ask students to make accounts. The activities in this guide are designed so you can teach the whole sequence without a single student account.
Five units. The order matters more than the pacing — each one gives students the vocabulary they need for the next.
How these systems actually work. Not the maths. The one idea that changes everything: a large language model predicts likely text, it does not look up facts. Students who understand this stop being surprised when a model invents a citation, because inventing a plausible-looking citation is exactly what a text predictor would do. This is the unit that makes every later unit make sense. Our explainer on what an LLM is is written for exactly this level.
Prompting as specification. Frame it as a writing skill, because it is one. A vague request produces a vague answer; a request that supplies context, a role, a goal and a format produces something usable. Students who already write for an audience in English class have most of this — the transfer is quick.
Verification. The most valuable unit in the sequence and the one most often skipped. How do you check a claim? Where does the primary source live? What does it mean that a model sounds equally confident whether it is right or wrong? Our guides on spotting hallucinations and telling whether something is AI-generated both work as student reading.
Limits and failure modes. Where AI is reliably weak: current events past its training, arithmetic it has not been given tools for, anything requiring genuine source-of-truth lookup, and anything where being confidently wrong is expensive. Pair this with the tasks where it is genuinely strong, so students get a real map rather than a warning.
Ethics and disclosure. Bias in training data, whose work these systems learned from, what happens to anything you paste in, and the norms of saying when you used it. This lands better last, once students have enough technical grounding to reason about it rather than just react.
If you have one week rather than five, teach unit 3. Verification alone gets students most of the practical benefit.
Each of these works with AI output you generate and bring to class, printed or projected. No student accounts, no data questions, no age-limit problem.
Ask an AI tool a question in your subject where you know the answer cold. Print the response. Somewhere in it there will usually be something wrong, overstated, or invented — a date that drifts, a citation that does not exist, a confident claim that is subtly off. Students work in pairs to find it and prove it with a source. If the output is clean, that is worth discussing too: what made this question one it handled well?
Same task, two prompts: one lazy ("write about the French Revolution"), one specified (audience, length, format, angle, what to leave out). Show both outputs side by side. Students articulate what the second prompt did that the first did not. This teaches specification faster than any amount of explaining.
Give students an AI-written paragraph containing four factual claims. Their job is to trace each one to a primary source, and to categorise it: confirmed, unconfirmed, or wrong. The discovery that "unconfirmed" is the biggest category is the lesson.
Hand out a competent but flat AI-written piece — a lab report, a paragraph of analysis, a cover letter. Students rewrite it for a specific real reader. This one reliably produces the realisation that the AI draft was a starting point rather than an ending, which is the habit you actually want.
Give students eight tasks from real life. Sort them: AI does this well, AI helps but needs checking, do not use AI here. Then defend the sorting. Disagreements in the middle category are where the good discussion happens.
All five assess directly. Students produce reasoning you can read, rather than an artifact you have to authenticate.
Two things are true at once, and pretending otherwise is why this conversation goes badly.
AI detection tools are not reliable enough to accuse a student. They produce false positives, and published evaluations have found they flag writing by non-native English speakers at higher rates. Treat a detector score as a prompt to have a conversation, never as evidence. A student's writing history and their ability to discuss their own work are far better signals.
And "just ban it" does not hold. It is unenforceable, it pushes use underground where you cannot teach into it, and it leaves students to work out the norms alone.
The workable third option is to make AI use visible and graded. Set a clear scale for each assignment — for example: no AI; AI for brainstorming only; AI for drafting with disclosure; AI-assisted throughout with a process log. Then assess the process rather than only the artifact. A student who documents what they asked, what came back, what they kept and what they rejected has demonstrated more skill than one who wrote a clean paragraph unaided, and you can actually tell the difference.
This also reframes the incentive. When the reasoning is the graded object, pasting an unedited AI answer stops being a shortcut and starts being an obviously incomplete submission.
Practice this, don't just read it.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
The trap is assessing recall — a quiz on what a transformer is tells you nothing about whether a student can spot a fabricated statistic. Assess the behaviours instead.
A process log is the highest-value artifact in the whole topic and costs nothing to introduce: what I asked, what it gave me, what I changed, and why. Four lines. It makes thinking visible, it makes AI use discussable rather than furtive, and it is the same habit professionals are being asked to adopt at work.
For a starting-point measure, our free AI IQ test is 10 questions, takes about two minutes, and needs no signup or account — which makes it usable as an anonymous pre- and post-unit check. It scores practical judgment rather than definitions, so it maps to what this sequence teaches.
Pick one activity from the list above and run it in a single lesson. Activity 1, the hallucination hunt, is the best opener: it takes fifteen minutes to prepare, it produces immediate engagement, and it lands the core idea — that fluent output and correct output are different things — without you having to argue for it.
From there, build the sequence around whatever your subject already assesses. History teachers have source evaluation. Science teachers have experimental method. English teachers have audience and register. AI literacy is not a new discipline bolted on; it is those existing skills applied to a new kind of text.
If you want to build your own AI fluency alongside your students', Iro AI teaches practical AI skills in five-minute daily lessons — prompting, verification, tool selection and where AI fails — across 29 learning paths and 477 lessons. Every path is open on the free tier, and it runs on iPhone or in any browser. It is built for individual learners aged 13 and up rather than as a classroom management system, so the honest use here is your own practice, or a recommendation to older students who want to go further than a unit allows. Our companion guide on AI for teachers covers using AI for planning, feedback and differentiation — the other half of this job.
Practice this, don't just read it.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
Five things, in order: how language models work (they predict likely text rather than looking up facts), prompting as a specification skill, verification, the limits and failure modes, and ethics and disclosure. If you only have time for one, teach verification — it delivers most of the practical benefit on its own.
Yes, and it is often better. Generate the AI output yourself and bring it to class for students to critique, verify and improve. That removes age-limit and student-data questions entirely, and critique produces better discussion than generation does.
No. They produce false positives, and published evaluations have found they flag writing by non-native English speakers at higher rates. Treat a detector score as a reason to have a conversation, never as evidence. A student's writing history and their ability to discuss their own work are far better signals.
Both are policies rather than teaching. A ban is unenforceable and pushes use out of sight where you cannot teach into it. The workable third option is to make AI use visible and graded: set a clear permission level per assignment, require a short process log, and assess the reasoning rather than only the finished artifact.
Assess behaviours, not recall. Ask students to write the prompt rather than the answer and grade the prompt. Give them an output with a planted error and see if they find it. Ask them to judge whether AI suits a given task and defend it. A four-line process log — what I asked, what it gave me, what I changed, why — makes all of this visible.
The critical-thinking core suits any age, but most mainstream AI tools set a minimum age of 13 and many require parental consent below 18. Check the terms of any tool before asking students to create accounts. Teacher-generated output that students critique avoids the issue at every age.