AI detectors promise a simple answer to a hard question: was this text written by a person or by a machine? Schools use them to review essays, publishers use them to screen submissions, and writers use them to check their own drafts. The short answer to whether they work is: sometimes, within limits, and never with certainty.
This article explains what AI detectors measure, why they make mistakes, and how to treat a detector score in practice, whether you are a student, a teacher or an editor.
How AI detectors work
Most AI detectors are classifiers. They have been trained on large collections of text labelled as human-written or AI-generated, and they learn statistical differences between the two. When you paste in a document, the detector compares it with what it has learned and returns a probability or a label.
The details differ by product and are rarely published in full, but detectors typically look at signals like these:
Predictability
Language models choose likely words. Text in which almost every word is the most expected choice looks machine-like. Human writers make more surprising choices, use idioms in odd places and change direction mid-thought. This is often described as low or high perplexity.
Variation
Human writing tends to vary more in sentence length and structure from one sentence to the next. A long, winding sentence followed by a blunt one is common in human text and less common in raw AI output. This is sometimes called burstiness.
Vocabulary and phrasing
Certain words and phrases appear much more often in AI output than in ordinary writing. Detectors can pick up on these patterns, along with formulaic structures such as repeated transition words or tidy three-part lists.
Training-based patterns
Beyond any single signal, a classifier learns combinations that are hard to describe in words. That makes detectors more capable than a checklist, but also harder to interpret: when one flags a text, it usually cannot say exactly why.
Why detectors get it wrong
Every classifier has error rates. With AI detectors, those errors fall into two groups, and both matter.
False positives: human text flagged as AI
A false positive happens when a person's own writing is labelled as AI-generated. This is the error with the most serious consequences, because it can lead to an accusation of misconduct. Some kinds of writing are more likely to trigger it:
- Formal or formulaic writing. Lab reports, legal summaries and standard business emails follow predictable patterns by design.
- Writing by non-native speakers. Studies have found that detectors can flag text by non-native English writers more often, because it may use a smaller, more common vocabulary.
- Short texts. A few sentences give the detector very little evidence, so its guess is less reliable.
- Heavily edited text. Running a draft through grammar and style tools can smooth it into something more predictable.
False negatives: AI text that passes as human
A false negative is AI text that the detector labels as human. This happens when a draft has been edited by hand, rewritten by another tool, or generated with prompts that encourage a less predictable style. As AI models improve, their output also becomes harder to tell apart from human writing.
Detectors disagree with each other
Paste the same text into several detectors and you will often get different verdicts. Each product uses different training data, different models and different thresholds for what counts as AI. A text can look clearly human to one tool and likely AI to another.
Detectors change
Vendors update their models regularly, often without announcing it. A result from last month is not guaranteed to repeat this month, which makes any single score a snapshot rather than a fact.
What a detector score really means
A score such as "85% AI" is easy to misread. It usually means the model is fairly confident that the text resembles AI-generated text in its training data. It does not mean that 85% of the words were written by AI, and it is not proof of how the text was produced.
Some tools, including Turnitin's AI writing indicator, are explicit that their results should be one input into a conversation, not a final judgement. That is the right way to think about every detector.
Practical advice
For students and writers
- Know the rules. Find out whether your course, employer or publisher allows AI assistance, and whether it must be disclosed.
- Keep your process. Notes, outlines, earlier drafts and version history are the best evidence of your own work if a score is ever questioned.
- Write in your own voice. Specific examples, your own opinions and details only you know make writing both better and more clearly yours.
- Do not rely on a tool to pass a detector. No humanizer can guarantee a result. Use them to improve writing, not to hide how it was made where that is not allowed.
For teachers and editors
- Treat a score as a prompt for a conversation. Ask the writer about their process and look at drafts before drawing conclusions.
- Be careful with short texts and non-native writers. These are the cases where false positives are most likely.
- Do not combine weak signals into certainty. Two detectors agreeing is still not proof.
- Design for transparency. In-class writing, oral follow-ups and clear AI policies reduce the need to rely on detectors at all.
Common myths about AI detection
"A 0% score proves a human wrote it"
A low score only means the text did not resemble AI output to that detector on that day. Edited AI text can score low, and so can text from a model the detector was not trained on.
"Detectors can tell which tool was used"
Some products show a breakdown by model, but these are estimates based on style. They are not a reliable record of which tool, if any, produced a text.
"Changing a few words fools any detector"
Swapping synonyms rarely changes the underlying rhythm and structure that detectors pick up, and it often makes writing worse. Real revision changes how sentences are built, not just which words they use.
"If two detectors agree, it must be true"
Detectors are trained on overlapping data and look for similar signals, so they can share the same blind spots. Agreement makes a result more likely, not certain.
Where HumanizerGPT stands
We build an AI humanizer, so we are not neutral on this topic, and we think that is a reason to be more transparent, not less. Two principles guide what we do.
First, we do not claim that our tool bypasses every detector. No honest tool can promise that. Instead, we are preparing a public benchmark that shows how our output scores across major detectors, including the detectors where we do worst, and we explain the method behind it.
Second, our tools are for improving writing that you have the right to use: making AI-assisted drafts clearer, more natural and more like you. Our responsible use policy sets out what that means in practice.
So, do AI detectors work?
They work well enough to notice patterns in unedited AI text much of the time. They do not work well enough to prove anything on their own. They produce false positives that can hurt real people, especially non-native writers and anyone who writes in a formal style, and they miss AI text that has been edited.
Used as one signal among several, with a human conversation behind it, a detector can be helpful. Used as a verdict, it is a risk. And if you want your writing to read as human, the most reliable approach is still the simplest one: make it genuinely yours. Our guide on how to humanize AI text walks through how.