Why behavioural AI detection is more accurate than text analysis

Most AI detectors read the finished product. The most reliable ones don't read it at all. The shift from analysing prose to analysing writing behaviour is the most significant methodological development in academic AI detection - and it matters for accuracy, for fairness, and for data protection.

The fundamental problem with text analysis

Text-based AI detection relies on the premise that AI-generated writing looks detectably different from human writing. In 2022 and 2023, this was largely true: early language models produced text with recognisable statistical patterns. Since then, models have improved substantially, and the distinguishing signals have eroded. A detection method that was reasonably reliable against early model output now struggles with more recent generations.

There's a structural problem here that won't go away. Language models are trained to produce text that reads like high-quality human writing. As they improve, they get better at exactly the thing the detectors are trying to measure. The detector and the model are in an arms race, and the model has better training data. Chasing output patterns is a losing strategy over any significant time horizon.

What behavioural detection measures instead

Behavioural detection takes a completely different approach. Rather than reading the finished text, it records the process by which the text was produced: the sequence of events in the writing session. Keystroke events and their timing. Paste events, with the size of the content pasted. Focus changes - moments when the browser tab lost and regained attention. Session duration relative to word count.

These signals are not about the content of the writing. They're about the behaviour of the person composing it. A student who spent forty minutes typing a 400-word answer in irregular bursts, making corrections as they went, displays a fundamentally different process signature from a student who opened the submission form, pasted a pre-written 400-word answer in under a minute, and closed the window.

Why behavioural signals are harder to fake

One of the genuine advantages of behavioural detection is its resistance to circumvention. Paraphrasing AI output fools text-based detectors almost entirely. Changing vocabulary, restructuring sentences, and adding stylistic variation are well within the capability of motivated students - and there are widely available tools to do it automatically.

Faking a convincing typing behaviour record is considerably harder. Replicating the natural variance of human composing - the uneven burst patterns, the correction events, the pauses that reflect genuine thought - requires both technical access to the session capture mechanism and active effort to simulate the right patterns. Most students have neither. The result is that behavioural detection is substantially more resistant to the basic circumvention strategies that defeat text analysis.

The ESL and fairness advantage

Text-based detection has a well-documented bias problem. Detectors trained primarily on native English writing flag non-native writing at disproportionately high rates, because careful, formal non-native English shares statistical characteristics with AI output: predictable vocabulary, conventional structures, formal register. Research has found false positive rates of up to 61% for non-native writers on commonly used detectors, compared to much lower rates for native speakers.

Behavioural detection has no equivalent bias. The process of composing text - the typing rhythm, the pause patterns, the sequence of drafting decisions - looks the same whether the student is writing in their first language or their fourth. A student writing careful, formal English because they've worked hard at academic conventions produces the same process signature as any other genuine writer. The detection approach literally cannot distinguish them by language background.

The privacy advantage

Text-based detectors necessarily read the student's work to score it. Sending student essays to third-party cloud services raises data protection questions - particularly under GDPR - about where that data is stored, who can access it, and how long it's retained. For schools with strict safeguarding requirements, this can be a genuine procurement barrier.

Behavioural detection tools built around data minimisation principles don't have the same exposure. A tool that captures only timing and event-type metadata - never the actual text - has a much smaller data footprint. There's nothing to expose in a breach of the telemetry record because the content was never captured. The analysis runs on event metadata alone.

How it works in practice

From a teacher's perspective, a behavioural detection tool integrates into the homework collection workflow rather than adding a separate review step. Students receive an assignment link, write directly in the submission interface, and submit. The teacher sees the finished work alongside a process summary: session duration, any paste events with their sizes, the overall timeline of activity.

High-concern signals - a very short session, a large early paste, minimal revision - are visible at a glance. The teacher can assess a flagged submission in under a minute and decide whether to have a process conversation with the student or move on. This is faster than running submissions through an external tool, and the evidence it produces is substantially more defensible.