How Learnaway works: behavioural AI detection explained

Most AI detection tools ask whether a piece of text looks machine-generated. Learnaway asks a different question: does the process by which this text was produced look like genuine writing? These aren't the same question, and they produce meaningfully different evidence - evidence that's more language-neutral, more defensible, and in many cases more practically useful for teachers.

A teacher creates an assignment in the Learnaway dashboard and sets a policy for that specific task - whether AI assistance is acceptable and, if so, in what form. Learnaway generates a shareable link that the teacher sends to students, either directly or via their normal communication channel.

Students open the link and see a writing area - a simple, distraction-free text editor. They work on their assignment within this interface. There's no installation required, no accounts for students to create, and no data the student needs to configure.

While the student writes, Learnaway captures behavioural events. Here's exactly what is and isn't collected:

This distinction is designed into the system, not just stated as policy. The telemetry architecture routes event metadata to the server while the text content remains client-side until the student submits. It's technically not possible for the server to receive the student's prose during the session — only the timing and event-type data. See our post on why Learnaway is not a keylogger for the full technical explanation.

When a student submits, the collected event data passes through Learnaway's scoring engine. This is a deterministic rule-based system — not an AI model — that applies a set of signal weights to produce a risk indicator. Signals include the ratio of typed to pasted content, total time on task, whether the session duration is plausible for the assignment length, the distribution of paste event sizes, and focus-loss events during the session.

In the dashboard, a teacher sees each submission with a traffic-light style indicator - green, amber, or red - alongside a timeline of the writing session. The timeline shows when activity occurred, where paste events happened and how large they were, and whether there were periods of inactivity that seem implausible given the submission length and timing.

Crucially, the teacher sees enough to start an informed conversation, not a verdict. The indicator says 'this is worth a closer look' - it doesn't say 'this student cheated'. The professional judgement remains with the teacher, supported by concrete process data rather than a probabilistic text score.

Text-based AI detection is probabilistic and carries documented bias against non-native English writers. A process-based signal is factual: the submission contained a single paste event of 900 characters at minute two of a five-minute session. That's a statement about what happened, not a statistical inference about what the prose sounds like.

In formal proceedings, the difference matters. 'The detector gave it a 94% AI likelihood score' is an inference with a false positive rate. 'The entire body of the essay arrived via a single paste event, and total session time was four minutes for a 1,500-word piece' is an observable fact. Both may contribute to the same conclusion, but the second provides a more solid evidential footing.

Learnaway doesn't make accusations. It doesn't share scores with students, doesn't assign labels like 'cheated' or 'copied', and doesn't trigger automatic consequences. The system generates a risk indicator that a teacher can interpret in context. Whether and how that indicator leads to a conversation or formal review is entirely the teacher's decision.

The system also doesn't capture content - not character-level data, not draft text, not clipboard contents beyond the byte count. The design goal is minimum necessary data: enough to assess process behaviour, nothing that would constitute surveillance of what a student thinks or writes.