How to spot AI writing: the signals teachers overlook

Most guidance on spotting AI writing focuses on the output: the style, the vocabulary, the structure of sentences. These signals exist and are worth knowing, but they're the weakest ones available - and relying on them heavily is what gets schools into trouble.

Output signals: worth knowing, not worth acting on alone

There are output patterns that AI-generated text tends toward, and it's useful to recognise them. Very uniform paragraph lengths. Topic sentences that too neatly preview each paragraph's content. An absence of the hedging and self-correction that characterise genuine intellectual effort. Generic examples rather than specific ones. A vocabulary that's technically appropriate but slightly too formal for the apparent writer.

These patterns are real. The problem is that they're not unique to AI generation. Students who write carefully produce some of them. ESL students, for whom formal academic style is a deliberate achievement rather than a natural register, produce most of them. And as language models improve, the patterns become less reliable - the same tools that flagged earlier model output struggle with more recent ones.

The ESL false positive problem

This bears dwelling on, because the consequences are serious. A 2023 Stanford study found that seven widely used AI detectors flagged non-native English essays as AI-generated at rates up to 61% - compared to much lower rates for native speakers submitting equivalent work. The reason is that characteristics of careful, formal non-native writing overlap substantially with AI-generated text: predictable vocabulary, conventional sentence structures, formal register.

For schools with significant proportions of international or EAL students, this creates a genuine fairness problem. A detection approach that systematically results in non-native writers being flagged at elevated rates is not an equitable approach, whatever its headline accuracy figure. The UK Equality Act 2010 places real obligations on institutions to avoid exactly this kind of outcome.

The process signals that are actually reliable

Process signals - how work was composed rather than how it reads - are both more reliable and fairer. A student who types their work in genuine bursts over forty-five minutes, makes corrections as they go, and has a session that looks like thinking produces a different process fingerprint from a student who pastes a fully formed 600-word answer in under a minute. This difference is visible, documentable, and independent of the student's first language.

Specific process signals worth looking for: session duration relative to word count (genuine writing takes time); paste events, especially large ones early in the session; typing cadence (genuine composing is irregular; transcription from pre-written text tends to be even); and revision activity (real drafting produces edits throughout, not just at the end).

The conversation test

The most diagnostic single signal available to any teacher - with or without technological tools - is the conversation test. Ask the student to walk you through their process: how they started, what was hard, what they changed and why. Genuine engagement with a task leaves specific memories. Students who wrote their work can usually discuss it in terms of specific choices, specific obstacles, specific moments of uncertainty.

This isn't about catching students out. Students who used AI within a permitted policy - for research, brainstorming, or editing - can also account for their process. What the conversation surfaces is whether any real intellectual engagement with the material happened at all.

When to act on what you find

Several things together make a stronger case than any single signal: anomalous process data, a submission significantly inconsistent with previous work quality, and an inability to reconstruct the thinking behind the submitted text. If all three are present, formal follow-up is reasonable. If only one or two signals are present, use what you have as the basis for a supportive conversation rather than an accusation.

The instinct to act quickly on AI suspicion is understandable. Most of the harm in these situations, though, comes from acting before there's enough to go on. A well-placed curiosity question leaves you in a considerably stronger position than a premature accusation - and it almost always tells you more.