4 Proven Ways to Set AI Rules That Protect Learning

Teaching in the Digital Age · Keep the Thinking Human (Part 2)

4 Proven Ways to Set AI Rules That Protect Learning

Blanket AI rules never fit every task. Set them one task at a time, and tell students why.
⏱ 6 min read · By Matt Faulkner · Pedagogy / Gen AI

Last week we looked at five AI shortcuts that cost learning. Brainstorming, research, summaries and first drafts all feel harmless. Each one can skip the thinking a task was built to teach.

That leaves a practical problem. Most school AI policies list what students may and may not do. Very few tell a classroom teacher how to decide for a given task. Good AI rules come from that decision, not from a list on the staffroom wall.

The gap is real. A recent Harvard Project Zero report found only 18% of US teachers said their school had a specific AI policy. Australian schools have the national Framework, but the task-level call still lands on you. Here are four ways to make that call well.

A teacher explains an AI rules card on the whiteboard to Year 5 students, showing which tasks happen without, with or about AI.


Way 1: Name the Skill Before You Set AI Rules

Daniel Willingham puts the principle in one line. Students should only use AI for things they already know how to do well. He compares it to calculators. Once a child has mastered multiplication, a calculator in an area problem costs nothing. While they are still learning to multiply, it costs everything.

So start every task by naming the skill it builds. Write it in one sentence on your plan. “This task is about choosing the strongest evidence.” “This task is about explaining how I solved it.” That sentence decides the rule before the rule is written.

This is backward design applied to AI. You decide what students must be able to do, then protect the steps that build it. The Australian Framework for Generative AI in Schools asks for the same thing. Guiding statement 1.5 asks that student work clearly outline how AI should or should not be used.


Way 2: Sort Each Task Into Without, With or About AI

AJ Juliani borrows a simple frame from Paul Alivisatos, president of the University of Chicago. Students need to think without AI, with AI and about AI. Each task sits mostly in one of those three lanes.

In a primary classroom, most tasks belong in the first lane. Students in Years 3 to 6 are still building the knowledge and skills that AI would replace. The other two lanes still have a place, as long as you choose them on purpose.

🚦 Three Lanes for AI Rules
  • Without AI: The task builds a core skill. Think narrative planning, written explanations of maths strategies and first drafts.
  • With AI: The student already does the core skill well. AI handles a side job, like generating practice questions you have checked.
  • About AI: AI is the topic. Students check an AI answer, compare two outputs or ask who made the tool and why.

The third lane is where our earlier post on catching AI mistakes fits. Thinking about AI often works best with the teacher running the tool on the board. Many tools set age limits above primary, so a teacher-led demo also keeps you on the right side of them.


Way 3: Teach Where AI Works and Where It Doesn’t

Rules land better when students understand the reasoning. Juliani recently shared a lesson opener from neuroscientist Thomas Henning. Henning asked a class what they would need to know to guess someone’s height. Students answered fast: age, parents’ heights and so on.

Then he asked what they would need to know to judge how charismatic someone is. The room slowed down. Charisma has too many parts to measure. In Henning’s view, AI does best on problems with few parts that are easy to check. Code and maths answers are good examples.

By contrast, AI is weaker where quality is hard to measure. Kindness, leadership and taste all sit in that space. For a primary version, try two questions side by side. “What is 48 times 25?” versus “Which ending makes our class story better?” Ask students which one AI would handle well, and how they would check it.

Students who understand that difference start to ask the right question on their own. That habit ties straight back to the AI confidence trap.


Way 4: Say the Rule Out Loud, and Model Your Own Use

Juliani tells his students that letting AI do their learning is like sending a robot to the gym for you. The robot gets the workout. You get nothing. It is funny, and that is why students remember it.

He also states the rule for every assignment without shaming anyone. This task is about X. If a chatbot does X, you miss the learning. Students who know the rule upfront have no grey area to hide in.

Modelling matters too. When you use AI, say what it did and what you did. “I used AI to make 10 practice questions. I checked each one and fixed two.” Students see AI as a tool you control, not a tool that does the thinking.

An AI rule without a reason feels like a ban. An AI rule with a reason feels like a plan for learning.

Where students do use AI, choose tools that hint rather than answer. In the Bastani study from Part 1, a tutor version that gave hints largely avoided the learning loss. The open chatbot did not.


An AI Rules Card for Every Task

Put the four ways together and you get a short card for each task. Write it on the task sheet, the board or the top of the Seesaw activity.

✏️ The AI Rules Card
  • This task builds: one skill, in one sentence.
  • AI lane: without, with or about.
  • AI may help with: the side jobs, if any.
  • AI must not: the step that builds the skill.
  • Why: one line students can repeat back to you.

Clear AI rules solve half the problem. The other half is knowing whether the thinking really happened. Next week, Part 3 looks at how to see student thinking as it happens, so you can stop playing detective.

Your turn

Pick one task you are setting this week and fill in the AI Rules Card. Which lane did it land in? Share your card in the comments.

References

Australian Government Department of Education. (2023). Australian framework for generative artificial intelligence (AI) in schools. education.gov.au

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122

Juliani, A. J. (2026a, October 12). The kids still want to learn from humans. Next Gen Schools. ajjuliani.beehiiv.com

Juliani, A. J. (2026b, October 14). My simple strategy for dealing with AI slop as a teacher. Next Gen Schools. ajjuliani.beehiiv.com

Juliani, A. J. (2026c, October 20). What my students learned about AI and cognitive science from a neuroscience PhD. Next Gen Schools. ajjuliani.beehiiv.com

Nagelhout, R. (2026, October 6). The AI cheating crisis in schools isn’t (only) what you think. Harvard Graduate School of Education. gse.harvard.edu

Willingham, D. (2026, August 27). Students should only use AI for things they already know how to do well. Thomas B. Fordham Institute. fordhaminstitute.org

4 Easy Ways to Help Students Catch AI Mistakes

Four proven ways to teach primary students to check AI output: plant an error, rerun and compare, ask one question every time, and build knowledge first.

Teaching in the Digital Age · AI Literacy

4 Proven Ways to Spot AI Mistakes in Your Classroom

AI hands your students a confident answer every time. The skill worth teaching is knowing when to doubt it.
⏱ 6 min read · By Matt Faulkner · Pedagogy / Gen AI

Last week’s look at the PISA device myth ended on one finding worth a second look. Students taught to check AI output outperformed students who used AI to write their drafts. Same tool, different job for the student.

That finding raises a practical question. What does it look like to teach a Year 5 class to check AI output on a Tuesday afternoon? Most advice stops at “teach critical thinking.” That gives you a goal, not a lesson.

AJ Juliani’s recent piece, Teach the Frontier, offers a sharper starting point. Every serious plan for keeping AI honest assumes someone has the knowledge and judgement to say “that’s wrong.” By 2040, those people will be the students sitting in your room right now. Here are four ways to start building that judgement in a primary classroom.

Primary students point at an AI answer on the classroom board and check it for errors while their teacher watches.


Way 1: Build Knowledge Before You Check AI Output

Start with the step most AI advice skips. A student has nothing to check an answer against without knowing the topic first.

Daniel Willingham made this case long before chatbots arrived. Critical thinking, he argued, is not a free-floating skill you teach once and apply everywhere. It depends on what you know about the subject in front of you. In other words, a student who thinks critically about fractions might be helpless with a question about volcanoes.

Picture a Year 4 HASS class asking an AI tool about the First Fleet. The answer says the convicts settled at Botany Bay. A student who has studied the unit knows the fleet moved on to Sydney Cove. A student who hasn’t will copy the error into their book, neatly and confidently.

A student only catches a wrong answer about something they already know. Knowledge is the checking tool.

Juliani puts it plainly: judgement sits on top of knowledge. As AI tools get more capable, background knowledge matters more, not less. For this reason, a knowledge-rich unit is your first and best protection against AI errors.


Way 2: Plant an Error, Then Check AI Output

The simplest place to start is an AI answer you have checked yourself. Before the lesson, run a prompt, read the response, and make sure it contains one or two errors. Some will appear on their own. Otherwise, edit one in.

Run the prompt yourself rather than handing students a login. Many AI tools set age limits above primary school. The Australian Framework for Generative AI in Schools also asks schools to protect student data. Displaying the output on the board avoids both problems.

For example, ask an AI tool to explain to a Year 5 student how to compare decimals. Then plant a classic misconception: “0.25 is bigger than 0.5 because 25 is bigger than 5.” Tell the class there are two errors, and give pairs five minutes to find them. The misconception you plant is often the one half the room already holds. As a result, the error hunt doubles as a diagnostic.


Way 3: Rerun the Prompt to Check AI Output

Next, run the same prompt twice and put both answers side by side. Ask students one question: where do these answers disagree?

Disagreement is a signal. If one answer gives a river’s length in one figure and the second gives another, at least one is wrong. Students then go to the atlas, the textbook, or a trusted website to settle it. That trip to a second source is how students learn to check AI output.

Even so, teach the other half of the lesson too. Two AI answers that agree can still both be wrong, because they come from the same tool. Agreement tells you the tool is consistent. Only a source outside the tool tells you it is right.


Way 4: Make One Question a Routine

Juliani asks his own children one question every time AI gives them an answer: “How would you know if it was wrong?” His hope is that one day they ask it without him.

That repetition is the point. A routine question works the same way as the higher-order questioning you already use, except students eventually take it over. It also pushes back on the problem we covered in the AI confidence trap. AI tools tend to agree with the user, so students need a habit of disagreeing with the tool.

🔍 Five Questions for Every AI Answer
  • How would you know if this was wrong? The anchor question. Ask it every time.
  • Where else could you check this? Name a source outside the AI tool.
  • Does this match what we learnt in class? Sends students back to their own knowledge.
  • Which part sounds sure but shows no evidence? Confident tone is not proof.
  • What did it leave out? A harder one for Years 5 and 6.

Print the list as a bookmark or a poster near the board. In practice, the first question carries most of the weight. The other four give older students somewhere to go next.


Keep Most Tasks AI-Free, and Say Why

All four ways share one condition. Students need plenty of work where AI plays no part at all. Juliani’s advice is to keep most tasks AI-free and to tell students the reason. The hard thinking stays with the student, because that thinking builds the knowledge they will later check AI output against.

This links straight back to productive struggle. A student who hands the hard part to a chatbot skips the struggle, and skips the learning with it. Telling them so is honest, and most Year 6 students will follow the logic.

The Framework backs this up. Guiding statement 1.2 asks schools to teach students how AI tools work, including their limitations and biases. Statement 1.4 asks for AI use that supports critical thinking rather than restricting it. Statement 1.5 asks every task to state clearly whether AI belongs in it. The four ways above give you a classroom version of all three.

🗓️ A Four-Week Starter Plan
  • Week 1: Teach the content. No AI. Tell students why.
  • Week 2: Plant two errors in an AI answer on that content and run the hunt.
  • Week 3: Rerun one prompt twice and settle the disagreements with a second source.
  • Week 4: Hand pairs a fresh AI answer and the five questions. Step back and watch.

Your turn

Try the planted-error hunt with your class this term. Which error did your students catch first, and which one slipped past everyone? Tell us in the comments.

References

Australian Government Department of Education. (2023). Australian framework for generative artificial intelligence (AI) in schools. education.gov.au

Juliani, A. J. (2026, September 18). Teach the frontier. Next Gen Schools. ajjuliani.beehiiv.com

Juliani, A. J. (2026, September). What the PISA results really show us about technology and curriculum. AJ Juliani. ajjuliani.com

Willingham, D. T. (2007). Critical thinking: Why is it so hard to teach? American Educator, 31(2), 8–19.