AI in the classroom: a pragmatic teacher's guide

The blanket AI ban approach has a predictable failure mode: it tells students what not to do without telling them why, creates a rule that's almost impossible to enforce consistently, and frames AI as exclusively a threat rather than a tool with legitimate and illegitimate uses. Most teachers who've tried it have found it doesn't work - not because students are especially determined to cheat, but because a rule without a rationale or a mechanism is mostly ignored.

A more workable approach starts from a different premise: AI use is neither uniformly acceptable nor uniformly prohibited. What's appropriate depends on what the task is for.

Most teachers land on something like three categories of task, even if they don't formalise them. Making them explicit - and communicating them clearly to students - is what transforms a vague policy into a workable one.

The most effective protection against AI shortcuts isn't detection - it's designing tasks where the shortcut doesn't produce a good result. This isn't about making work harder; it's about making it more specific to context and knowledge that AI can't supply.

Strategies that make assignments more AI-resistant:

The single most effective moment to set expectations is at assignment creation, not after a problem arises. Students respond better to 'here's the policy and here's why' than to retrospective enforcement.

A brief statement at assignment launch does three things: removes the ambiguity that often drives students toward AI as a hedge ('I wasn't sure if it was allowed'), gives you a firmer basis for follow-up conversations if needed, and signals that you've thought about this seriously enough to address it directly.

Effective pre-assignment briefings are specific, not generic. 'No AI on this one - I want to see how you reason through the evidence without assistance' is more actionable than 'follow the school AI policy'. Students need to know which category the task falls into and why.

For written assignments in the 'no AI' or 'AI as tool' categories, collecting work through a tool that captures how it was entered gives you objective evidence if you need it. The most useful signals - whether the work was typed gradually or arrived in a single paste, how long the session lasted, whether there were natural pauses and edits - are only visible in the process data, not in the finished text.

Tools like Learnaway capture this timeline without recording what students type: the teacher sees a behaviour record, not a transcript. Students are always told what's captured before they write. The effect is partly preventive: knowing the process is recorded changes behaviour before any formal concern arises.

Not all AI use in educational contexts is a shortcut. There are tasks where AI is genuinely pedagogically useful - and treating it as uniformly problematic misses that.

AI works well as a research starting point, provided students verify claims against primary sources. It works well as a feedback mechanism for draft writing, particularly for students who lack access to tutoring or writing support. It works well as a subject-matter explainer for difficult concepts, particularly in subjects where students are scared to reveal gaps to a teacher.

Designing tasks that explicitly incorporate AI in a pedagogically sound way - asking students to critique AI-generated arguments, to verify AI research claims, or to compare their own analysis to an AI-generated version - prepares them for a world where professional use of AI is normal while teaching critical evaluation.

Even with clear policies and careful design, students will sometimes use AI when they weren't supposed to. When that happens, the response matters as much as the detection.

The most common mistake is treating the first hint of concern as grounds for formal action. Most situations resolve better through a direct conversation: 'Tell me how you approached this' is almost always more revealing than a detector score, and it gives the student a chance to explain context that changes the picture - they drafted elsewhere and pasted in, they used AI for research and then rewrote, they're under significant pressure and made a poor decision they regret.

For the cases where concern remains after a conversation, documented process data is the strongest evidence available. A timestamped session record is a statement of what happened; a text-based AI score is a probabilistic claim. In formal proceedings, that difference is significant.