How to detect AI writing fairly

A polished paragraph can make even experienced teachers pause. It can also be the result of a student who revised carefully, got help with English, or simply found their stride. This matters now because schools are moving from a simple ban-or-allow argument towards clearer evidence of learning. A useful how to detect AI writing approach should help a teacher ask a better question, not pretend it can settle a case on its own. That is the promise of this guide: a practical route from uncertainty to a fair, explainable next step.

TL;DR: start with the assignment rules, look at the work in context, check the writing process where you have it, and speak to the student before treating a signal as a conclusion. The details below show what that looks like in a busy classroom.

Start with the question you actually need answered

The mistake is to treat style as a verdict. Text-only tools estimate how predictable a sequence of words looks; they cannot witness who made the choices behind it. Start by checking the task instructions, the student's earlier work and the specific point that caused concern. If the concern is only “this sounds different”, there is not yet enough for an allegation.

A sound how to detect AI writing workflow is deliberately modest. It records what happened, distinguishes an observation from an inference, gives the learner a chance to explain, and keeps a short factual note. That sequence protects a conscientious student as much as it helps a teacher investigate a genuine concern. It also makes handover to a head of department far less fraught.

Before opening another tab, write one sentence that names the concern in neutral language. For example: “I need to understand how this conclusion was developed.” That is much better than “I need to prove this is AI”. The first sentence invites evidence; the second invites confirmation bias.

Choose evidence that fits the decision

Not every assignment deserves the same level of scrutiny. A low-stakes homework task may need a quick formative conversation. Controlled assessment or a final project may need a documented route under the centre policy. The table is a simple way to match the how to detect AI writing check to the consequence rather than making every flag feel like an emergency.

A practical route through the work

Use a three-part review. First, compare the submission with the brief: does it answer the question and acknowledge permitted tools? Secondly, look for a process record if one exists: when did work arrive, was it typed or pasted, and were there normal pauses and revisions? Finally, ask the student to talk through one decision in the work. A genuine author normally remembers where they changed course, even if their writing is imperfect.

Keep the review narrow. Look for a change that can be described without guessing at motive: a large paste, a missing source trail, an abrupt change in terminology, or an answer the student cannot unpack. Then ask about that point. A student may have a perfectly ordinary explanation, such as drafting offline, using accessibility software, or receiving permitted feedback. Good process makes room for it.

In Learnaway, the relevant evidence is process data: session length, typing and paste events, pauses and focus changes. It does not read the words a learner types. That is why its output is best used alongside the teacher's knowledge of the class and the assignment brief. See how Learnaway works and the teacher workflow for the exact boundaries.

What current guidance says

In England, JCQ's AI assessment guidance says centres should investigate authenticity concerns and accept only work that is the student's own. It does not say that an AI-detector score establishes malpractice. TEQSA's 2025 assessment-reform resource makes the same broader point: evidence of learning needs more than a technological guess. The expert lesson is straightforward: design the review around what a student can demonstrate, not what software thinks prose resembles.

For a current external check, read the Department for Education's AI guidance and the relevant assessment guidance before changing a school rule. They are more useful than a vendor headline because they describe responsibilities, not a product claim. In the same 2025 Ofsted work, Chief Inspector Sir Martyn Oliver said schools need to “manage the risks” alongside AI's potential.

The common thread is professional judgement. Jisc's assessment and academic-integrity training explicitly covers the limitations of AI detection, authentication and responses to concerns. That is a helpful corrective to the idea that a percentage score is enough. It is not. A percentage is at most a prompt to look more carefully.

Make the route visible before a concern occurs. Students should know what evidence a teacher may consider, who makes the decision, how they can explain a legitimate workflow and when a formal policy applies. Clear process is not administrative clutter. It is the difference between a learning conversation and an unexpected accusation.

Illustrative classroom scenario

Ms Patel notices that a Year 10 history conclusion uses terminology the student has not used before. Instead of saying “an AI detector flagged this”, she asks the student to explain why they chose the phrase “structural cause”. The student opens their notes, describes a class discussion, and points to a book they used. The concern ends there. On another occasion, a student cannot explain a pasted section and confirms that they used a chatbot without acknowledgement. The evidence and response are very different.

The useful detail is not the number on a dashboard. It is whether the teacher can say, calmly and accurately, what they saw and what they asked next. That leaves space for a student to demonstrate understanding, correct a misconception, or disclose help they did not realise needed acknowledgement. It also means a genuine issue can be escalated with a clear record rather than a hunch.

Common mistakes to avoid

A common failure is running a detector first and then reading the work through its score. That makes every fluent sentence look suspicious. Another is asking a student to prove innocence by writing a surprise essay. Use a proportionate conversation first, follow your published process for formal work, and make any requested explanation relevant to the task rather than punitive.

Do not quietly change the rules after students have started. Do not put a detector score in the gradebook as though it were a mark. Do not collect more behavioural or personal data than the decision needs. The ICO's 2026 review covered audits of 28 widely used edtech providers and stressed that children need to be able to trust the tools schools adopt. Read the ICO statement before making data collection the default.

Turn the finding into a next step

For awareness-stage readers, the next step is to agree three neutral questions that every teacher can use. For a department considering a tool, trial it against a small set of normal assignments, including work from confident writers and multilingual learners, before it informs any high-stakes workflow.

If you want to try a process-first approach, create one low-stakes assignment, tell students exactly what is recorded, and compare the timeline with work you already know well. Learnaway's free assignment flow is designed for that first small test. For a wider rollout, use the pricing guide and trust information to involve the people responsible for teaching, data protection and safeguarding early.