AI Detectors: How They Work and How Reliable They Are

Level	Beginner

How a detector works

A detector does not “recognize” AI the way you might recognize a signature. It measures how closely a text resembles what a large language model typically produces: highly predictable word choices, even sentence rhythm, few surprises. The more predictable the text, the higher the “AI” score.

The trouble is that people can write very predictably too, think of a standard business report or a form letter, and AI can be pushed to write less predictably. So both errors happen: human writing flagged as AI (a false positive) and AI writing that goes unnoticed (a false negative).

A telling example

In January 2023, OpenAI, the company behind ChatGPT, launched its own detection tool. On July 20, 2023, it added a note to the original announcement saying the tool was no longer available because of its low rate of accuracy. If the maker of one of the most widely used models could not reliably spot its own model’s output, that is a good reason to be cautious about other tools’ claims.

Stanford researchers also showed in 2023 that several popular detectors wrongly labeled a large share of essays by non-native English speakers as AI-generated (study on arXiv).

Why it matters to you

In school or college: a detector score should not be enough to accuse anyone of cheating. A conversation about the content, drafts and version history are far more telling.
At work: when you receive a document, judge it on substance (accuracy, relevance, sources) rather than its suspected origin. Text from generative AI that has been reviewed and checked can be perfectly sound; human writing can contain mistakes.
If you are wrongly accused: ask what the decision is based on, share your drafts and point to the known limits of these tools.

Common myths

“100% AI means it’s proven.” No, it is a probability calculated by a tool that can be wrong.
“Typos prove a human wrote it.” Anyone can ask an AI to add a few.
“All AI text carries a watermark.” Watermarking is used by some companies only, and it is harder to apply to text than to images. For manipulated images and video, see deepfakes, along with the broader questions of AI ethics.

Related terms

AI watermarking · A mark, often invisible, embedded in an image, audio, video or text to signal that it was generated by AI.

Generative AI · AI that creates new content such as text, images, audio, video or code from a simple written or spoken request.

Large language model (LLM) · A program trained on huge amounts of text to predict what comes next in a text; it is the engine behind AI assistants.

Deepfake · An AI-made or AI-altered image, video or voice recording that makes a real person appear to say or do something they never did.

AI ethics · The principles for designing and using AI in a way that is fair, transparent and respectful of people and their rights.

Frequently asked questions

Are AI detectors accurate?

Their accuracy is limited. OpenAI withdrew its own detector in July 2023, citing its low rate of accuracy. A “90% AI” score is a statistical guess, not a finding of fact.

I wrote my essay myself and it was flagged as AI. How is that possible?

That is called a false positive. Plain, very formal or non-native writing is more likely to be flagged by mistake. Keep your drafts and version history so you can show how your work developed.

Can AI detectors be fooled?

AI-generated text that has been paraphrased, even lightly, often slips through. That is one more reason not to rely on detectors alone to make a decision about someone.

Is there a more reliable approach?

Some AI companies add an invisible watermark to generated content, which makes it easier to identify with the right tool. But watermarking is not universal and does not cover every model.

AI detectors: how much can you trust the verdict?

Key figures