The short answer

Behavioral AI detection asks whether work was produced by a person or by automation, and answers by examining the process that produced it. Instead of judging the finished artifact, an essay, a quiz response, a form submission, it reads the signals generated while the work happened: keystroke timing, pauses and corrections, edit patterns, cursor movement, scroll rhythm, navigation habits, and the technical properties of the browser the session ran in.

That sounds like an implementation detail. It is closer to the whole argument, because the two approaches leave you holding very different kinds of evidence. A judgment about finished text is an opinion about style. A record of how the work unfolded is a sequence of observable events a human reviewer can replay and check.

Two ways to ask the same question

Every detection product answers the question "did a person do this?" in one of two ways.

Artifact detection judges the finished product. Text classifiers read a submission and estimate how likely it is that a machine wrote it, based on statistical patterns in the words. The output is a score, and the score is the whole story: there is no event behind it that anyone can inspect. When the classifier is wrong, and every classifier is wrong some of the time, there is nothing to appeal to. The writer cannot disprove an opinion about style, and the institution cannot substantiate it. Careful, formulaic prose, including the prose many students are explicitly taught to write, can resemble exactly the patterns these classifiers are trained to flag.

That failure mode is not an abstraction. A false positive from an artifact detector is an accusation of dishonesty aimed at a real student, backed by evidence neither side can examine. A wrong call does lasting damage in either direction, and the person with the least power in the room absorbs most of it.

Process detection, the behavioral approach, examines how the work came to exist. The unit of analysis is the working session rather than the finished text. When a session is flagged, the flag points at specific things that happened: a full essay pasted in with no keystrokes behind it, cursor movement in straight lines at fixed speed, a browser carrying the fingerprints of an automation framework. Those events can be replayed by a person before anyone is accused of anything.

The signals, layer by layer

Behavioral systems draw on more than typing. The signals fall into three broad layers.

Behavioral signals describe how the session behaved. Human typing arrives in uneven bursts and hesitations, where generated text tends to land all at once. Real composition is full of corrections, deletions, and reordering as the writer changes their mind. People move a cursor in arcs with constant micro-corrections; automation tends to travel in straight lines at fixed speed. Real reading speeds up, slows down, and stops to re-read, and people return to a site with familiar habits, where automation repeats the same path page after page.

Technical signals describe what the browser and device reveal. Automation frameworks leave fingerprints: debugging-protocol artifacts, headless-browser tells, interference from browser extensions, and anomalies in how the browser renders graphics.

Integrity signals describe whether the measurement itself was disturbed. Tampering with a detection script is itself informative, and honeytraps, page elements normal use has no reason to touch, catch automation that interacts with everything it finds.

No single signal decides anything. A behavioral system weighs the layers together and produces a confidence score for the session. The how NotAI works page walks through one implementation of that pipeline, from capture to verdict, but the layering is common to the category.

Why process evidence is reviewable

The strongest argument for behavioral detection is not that it catches more. It is that when it flags, a human can check its work.

A process flag arrives with a timeline. A reviewer can replay the session, see what was typed and when, what was pasted, and what the browser reported, then decide with the evidence in front of them. The score informs the review; it does not replace it. That matters to the student, who gets a hearing grounded in something inspectable rather than a verdict from a black box. It matters to the instructor, who does not have to stake their credibility on a bare number. And it matters to the institution, because decisions with consequences need a record that survives an appeal.

Reviewability is also the only defensible answer to edge cases. A student who types slowly, composes in a separate document, or writes with assistive technology can look unusual to any automated system. Unusual input alone should condemn no one. The safeguard is not a cleverer score; it is a person looking at the whole session, with the student's context and accommodations in hand, before anything becomes an accusation.

Where the limits are

Behavioral detection is not a solved problem, and a vendor who suggests otherwise should worry you.

  • It is probabilistic. The output is a confidence score, not a certainty. False positives and false negatives both exist, and no threshold eliminates either.
  • Atypical input looks atypical. Speech-to-text, dictation, switch input, and other assistive technologies produce signals that differ from keyboard typing. Systems and reviewers have to account for accommodations, or the tool becomes a way to discriminate.
  • A paste is not a verdict. A paste event shows that text did not originate as keystrokes in that session. It does not show that a machine wrote it. Quoting a source or moving text between applications is ordinary human behavior.
  • Evasion is a moving target. Determined adversaries work at mimicking human patterns. Human behavior is hard to fake across every layer at once, but no vendor can truthfully claim evasion is impossible. We look at that arms race in depth in can AI agents be detected?
  • Process is not meaning. Behavioral detection reads composition, not content. It will not identify machine-generated text that a person retyped by hand, and it does not analyze what a submission says.

We publish the same limits, in more detail, in our AI transparency statement, because a detection vendor's limits are exactly the thing a buyer needs in writing.

Questions to ask any detection vendor

If you are evaluating detection for a school, district, or university, the technology label matters less than what you can verify. These questions apply to every vendor, including us.

  • Can a reviewer replay the evidence? If a flag cannot be inspected, every decision downstream of it rests on trust in a score.
  • Is a flag a verdict or the start of a review? Ask where the human sits in the workflow, and whether any consequence can fire without one.
  • What is collected, and where is it written down? Expect a public privacy policy, a data processing agreement, and student-privacy commitments your counsel can read. Ask where data is stored and who can see it.
  • What happens to students who type differently? Ask specifically about assistive technology, accommodations, and slow or unconventional writers, and what guidance reviewers are given.
  • Does the vendor publish its limits? Accuracy claims without published limitations, in both directions, are describing a product no one has built.
  • Which question does it actually answer? Artifact tools estimate whether text reads as machine-written. Process tools examine whether a session behaved like a person working. Know which answer you are buying.

For how these questions play out in an LMS deployment, from install to the integrity conversation, see NotAI for education.

The bottom line

Behavioral AI detection is a category, not a brand: any system that reads process signals instead of judging finished artifacts belongs to it. Its real advantage is procedural. When the question is whether a person did the work, the fairest answer comes from evidence of the work being done, reviewed by a human before it becomes an accusation. Evaluate any vendor against that standard.