Black October, explainedAppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

AI detector

Also called AI content detector, GPT detector, AI checker

An AI detector is a tool that guesses whether text, an image or audio was made by AI. It gives a probability, not proof, and it can be wrong.

Safety and societyEveryday

Products, laws and numbers here change fast. Checked October 2026.

In plain words

Teachers, editors and recruiters want to know: did a person write this, or a chatbot? AI detectors promise an answer. You paste in text and get a score, such as "85% likely AI".

Most detectors are themselves AI models, trained on examples of human and AI writing. They look for patterns, such as how predictable each word is. That is a guess about style, not a record of how the text was made.

So they make two kinds of mistakes. They miss AI text that has been lightly reworded, and they flag real human writing as AI. The second kind, a false positive, can mean a student wrongly accused of cheating. Studies have found some detectors flag non-native English writers far more often. Treat a score as a reason to ask questions, never as proof.

Everyday example

The exam invigilator's hunch

An invigilator sees a student glance sideways twice and suspects copying. Sometimes the hunch is right. Sometimes the student was just thinking. A fair school does not fail anyone on a hunch alone: it checks the answers, talks to the student, looks for other evidence. An AI detector is that hunch, made into a number. Useful as a starting point, unfair as a verdict.

Where the example breaks: An invigilator at least saw something happen. A detector never sees the writing process at all. It only reads the finished text and compares its style with patterns it learned.

Go deeper

How it actually works

There are two very different kinds of tool behind the word "detector".

  1. Pattern guessers

    A classifier is trained on large sets of human and AI text and learns to tell them apart. Some tools also measure how predictable each word is to a language model: AI text tends to be more predictable. The result is a probability, and it shifts with text length, topic, the model that wrote it and where the tool sets its threshold.

  2. Watermark checkers

    If an AI maker hides a watermark in its output, a matching checker can look for it reliably. Google's SynthID Detector, announced in May 2025, checks for Google's own watermark. It cannot say anything about content from tools that add no watermark.

  3. Why rewording beats them

    A 2023 study showed that running AI text through a paraphrasing tool sharply lowers detection rates for all main kinds of detector, while barely hurting quality. Tools sold as "humanisers" do this, so detection is an arms race.

  4. Why false positives matter

    Vanderbilt University turned off Turnitin's AI detector in August 2023. It noted that even a 1% false positive rate, applied to the 75,000 papers its students submitted in 2022, could mean about 750 wrongly flagged papers.

  5. Bias against some writers

    A 2023 study ran seven detectors on 91 essays written by non-native English speakers for the TOEFL exam. On average, 61% were wrongly labelled AI-written, and 89 of the 91 essays were flagged by at least one detector. The same detectors were close to perfect on essays by US eighth graders.

OpenAI's own detector
Launched January 2023, withdrawn July 2023 for low accuracy: it caught 26% of AI text and wrongly flagged 9% of human text
Independent test, 2023
14 detection tools judged "neither accurate nor reliable" by a team of academic researchers
Newer tests, 2025
A University of Chicago study found three commercial detectors kept false positives below 1% on its test set, but all struggled with passages under 50 words
Checked
October 2026. Detector accuracy changes as both tools and AI models change

Detectors that check for a watermark are far more reliable than those that guess from style, but only work when the maker added one.

Are AI detectors reliable enough to act on?

One view

The best ones are getting good

A 2025 University of Chicago study found some commercial detectors rarely flagged human text and caught most AI text. With strict limits on false positives and regular public testing, its authors argue detectors can be useful for policy.

The other view

Not for accusing anyone

Earlier independent studies found many tools inaccurate, biased against non-native writers and easy to fool by rewording. Short texts, mixed human and AI writing, and new models all trip them up, and even rare errors hurt real people.

Most likely: Accuracy varies widely between tools and situations. Nearly everyone agrees a score alone should never decide a cheating case: it should start a conversation, alongside drafts, notes and the student's own explanation.

Questions people ask

Do AI detectors actually work?

Partly. Some work well in controlled tests, others barely beat guessing. All of them can be fooled by rewording, struggle with short or mixed texts, and sometimes flag human writing. OpenAI withdrew its own detector in 2023 because it was not accurate enough.

Can an AI detector be wrong about my writing?

Yes. That is called a false positive. Studies find it happens more with short texts and with writers whose first language is not English. If you are flagged, show your drafts, notes and version history, and ask how the score was checked and what else was considered.

What does a score like 70% AI mean?

It is the tool's confidence that the text matches patterns it links with AI writing. It is not the share of the text written by AI, and it is not proof. Different detectors can give very different scores for the same text.

Can teachers tell if I used ChatGPT?

Not with certainty. Detectors give a probability and can be wrong either way. Teachers also notice sudden changes in style, made-up facts or references (hallucinations), or work you cannot explain. The most reliable check is simply talking about the work.

Is there a detector that can tell for sure?

Only when the content carries a watermark or content credentials from its maker, and you have the matching checker. Even then, a missing watermark proves nothing, because heavy rewording can remove one. For ordinary text with no watermark, no tool can be certain, because human and AI writing can look the same.

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked October 2026.