What is an AI content detector? A plain-English guide for teachers

If you've heard the term but aren't quite sure what it means - or whether you actually need one - this post is for you. AI content detectors are one of the most discussed tools in education right now, and one of the least clearly explained. Here's what they actually are, what they're genuinely useful for, and where they fall short.

The basic definition

An AI content detector is a tool designed to assess whether a piece of text was written by a human or generated by an AI system. The category includes a wide range of tools with different underlying methods, target markets, and levels of reliability. Some are free web-based tools aimed at individuals; others are enterprise software integrated into learning management systems or sold to institutions on licensing agreements.

What they share is the same goal: given a piece of text, estimate the likelihood that it came from an AI language model rather than a person. How reliably they achieve this varies significantly across tools and contexts.

How text-based detectors work

The most widely used category of AI content detector analyses the text itself. The dominant approach measures perplexity - how predictable each word choice is relative to what a language model would generate. AI systems tend to choose predictable, high-probability words; human writers introduce more variation and surprise. Some tools also measure burstiness, the variation in perplexity across different passages, since human writing tends to be less uniform.

These statistical measures are genuine signals of AI generation, but they're not definitive. Both ESL writers and very careful human writers produce text with higher predictability than typical native-speaker writing - and this causes false positives. The signals that distinguish AI from human writing overlap with the signals that distinguish non-native from native writing. This is the source of the well-documented bias problem.

Why they're used in education

AI content detectors became mainstream in schools and universities in 2023, when ChatGPT made high-quality text generation trivially accessible to students. Before then, the main academic integrity concern was plagiarism from external sources - a problem that plagiarism checkers addressed reasonably well. AI generation created a different problem: students could now produce apparently original essays without doing any of the thinking themselves.

Detectors offered a technical response to a challenge arriving faster than institutional policy could keep up with. Whether or not they worked perfectly, they gave teachers a way to raise concerns about submissions that would otherwise have been difficult to challenge. This role - as a conversation-starter rather than a verdict - remains the most legitimate use of text-based detection tools.

What they can and can't tell you

A text-based AI detector can tell you that a piece of text has statistical characteristics associated with AI generation. It cannot tell you definitively that AI was used. The outputs are probabilistic - a 'likely AI-generated' percentage - and the false positive problem is significant enough that this score alone should never be the basis for a formal misconduct finding.

Detectors also don't tell you how AI was used. 'Used AI' covers a wide range, from submitting an unedited ChatGPT response (clearly problematic) to using AI to check grammar or brainstorm ideas (often permitted). A detector score doesn't distinguish between these scenarios, and it says nothing about whether any AI use was within the policy the student was given.

Process-based detection: a different approach

A newer category of AI detection tool doesn't read the text at all. Instead, it records the writing process itself: the timeline of events during a submission session. Typing behaviour, paste events, session duration, focus patterns - these process signals provide a different kind of evidence. They're language-neutral, harder to fake, and more defensible as evidence in formal proceedings.

This approach aligns more directly with the question teachers actually need answered: not 'does this prose read like AI output?' but 'did this student actually write this?' The process record addresses the second question directly. A student who spent forty minutes composing an essay in genuine typing bursts answers that question very differently from a student who pasted a finished text in under a minute.

Should you use one?

For occasional use as one signal among many - a reason to look more closely or have a conversation - text-based AI detectors can be a reasonable starting point. They're not sufficient evidence for formal action on their own, and they carry particular risks when used with non-native or EAL students. If you do use them, treat them as the beginning of an enquiry, not a conclusion.

For systematic use across a course or year group, process-based collection tools are more appropriate. They gather evidence as a natural part of homework collection rather than as a separate review step, and the evidence they produce is fairer and more defensible. For schools that are serious about academic integrity in the AI era, building process collection into assignment workflows is the most reliable long-term approach.