AI-generated text detection: the complete guide for educators

AI-generated text detection has become one of the more contested areas of educational technology in the last three years. New tools appear regularly, accuracy claims range from cautious to implausible, and practical guidance for educators is often either too vague to act on or too specific to a single product to be useful. Here's what the technology actually is, what the main approaches are, where each falls short, and what a sensible strategy looks like.

What AI-generated text detection is trying to do

The core task is straightforward to state and difficult to execute: given a piece of text, determine whether it was produced by a human or a language model. Language models are trained to produce text that reads like human writing, and as they've improved, the resemblance has become closer. What was a clearly distinguishable gap in 2021 has become a narrow, contested boundary.

Different stakeholders care about this for different reasons. Academic institutions want to know whether students are submitting their own work. Publishers want to verify human authorship. Each use case has different error tolerances and different consequences for mistakes. What works for one context may not work for another.

Text-based detection: how it works

The dominant approach analyses statistical properties of the text itself. The central measure is perplexity: how surprising each word choice is relative to what a language model would predict. Language models tend to choose predictable, high-probability words; human writers make more unexpected choices. Burstiness – the variation in perplexity across different sections – adds a second signal, since human writing tends to be less uniform than AI output.

These measures provide a genuine statistical signal. The problem is that it's a weak signal in many real-world cases. It works best comparing clearly AI-generated text against clearly human-written text. In educational contexts, you're often looking at mixed content: human-edited AI output, AI-assisted research, AI-structured ideas with human writing. The signal is much less reliable in these realistic scenarios.

The false positive problem in education

The most serious practical limitation of text-based detection for educational use is false positives for non-native English writers. A 2023 Stanford study found that seven commonly used AI detectors flagged non-native English essays as AI-generated at rates reaching 61%, compared to far lower rates for native speakers doing equivalent work.

The mechanism is straightforward: careful, formal writing in a second language shares statistical characteristics with AI-generated text. Both tend toward predictable vocabulary, conventional sentence structures, and formal register. Detectors trained primarily on native English writing associate these patterns with AI generation – but they're also the patterns of students working hard to write correctly in a language they're still mastering.

For schools with significant EAL, international, or heritage language student populations, this isn't a minor edge case. It's a systematic bias that disproportionately flags conscientious non-native writers. Several UK and US universities have suspended text-based AI detection precisely because of this documented problem.

Process-based detection: the alternative approach

Process-based AI detection is a fundamentally different category. Rather than analysing finished text, it records writing behaviour during the submission session: the timing of events, paste events and their sizes, typing cadence, session duration, focus changes. The insight driving this approach is that genuine writing has a recognisable process signature. Typing comes in bursts with pauses; content grows incrementally; revisions appear throughout; sessions run for a plausible length relative to word count.

These process signals are language-neutral: they don't correlate with writing style, vocabulary predictability, or language background. A student writing in their fourth language, choosing conventional phrasing and simple sentence structures, produces the same kind of genuine process signature as any other real writer. The detection approach cannot produce an English-proficiency bias because it never reads the text.

Circumvention and robustness

Text-based detection is vulnerable to paraphrasing. A student who takes AI-generated text and rewrites it will typically produce output that scores low. The circumvention is straightforward and there are dedicated tools to automate it. Process-based detection is harder to circumvent: faking a realistic human writing session requires either technical access to the session capture mechanism or typing out pre-written text character by character – neither is trivial.

What a sound detection strategy looks like

The most defensible approach in 2026 isn't a single tool – it's layered. Process-based collection integrated into the homework workflow gives you behavioural evidence for every submission without a separate review step. Spot-checking with text-based tools, used carefully and never as a sole basis for action, adds a second signal for anomalous submissions. Process conversations – asking students to explain their approach and reasoning – are the most reliable complement to any technological tool.

Designing assessments that are inherently harder to complete with AI shortcuts is the upstream intervention that makes everything else easier: assignments anchored to specific class content, with in-class checkpoints, or requiring engagement with provided materials are substantially harder to complete with a generic AI prompt. Detection is most useful when prevention isn't enough on its own – and they work best together.