Are AI detectors accurate?

No. OpenAI retired its own AI text classifier in July 2023, reporting that it correctly identified only 26% of AI-written text while falsely flagging 9% of human writing. No detector can reliably determine whether a machine wrote a given piece of text.

Do AI detectors discriminate against non-native English speakers?

Yes. A 2023 Stanford study found that detectors flagged 61% of TOEFL essays written by non-native English speakers as AI-generated, while classifying essays by native speakers almost perfectly. Detectors measure how predictable writing is, which penalises simpler vocabulary and constructions.

Should teachers use AI detectors to catch cheating?

No. Given the documented false-positive rate and bias against non-native speakers, using a detector score to accuse a student is unreliable and disproportionately harms students writing in a second language. A detection percentage is not evidence.

What is an AI slop linter?

It flags the specific words, phrases, and structural habits that appear more often in machine-generated text than in human writing, and suggests how to rewrite them. It identifies bad writing patterns; it does not claim to identify the author.

Why AI detectors lie

Including this one. Read this before you use a number from this site for anything that affects another person.

1. THE COMPANY THAT BUILT CHATGPT COULD NOT DETECT CHATGPT

In January 2023 OpenAI shipped a tool called the AI Text Classifier. Six months later they killed it, citing its low rate of accuracy. Their own published numbers: it correctly identified 26 percent of machine-written text, and it labelled genuinely human writing as machine-written 9 percent of the time.

Sit with that second number. Roughly one in eleven innocent people, accused. By the organisation with more insight into how these models generate text than anybody on earth, using the weights themselves. They looked at the result and pulled the product.

Sources: OpenAI’s own announcement, updated with the retirement notice, Ars Technica.

2. DETECTORS PUNISH PEOPLE FOR WRITING IN SIMPLE ENGLISH

This is the part that should end the conversation. A Stanford team ran seven widely used detectors against 91 TOEFL essays written by non-native English speakers, alongside 88 essays by American eighth-graders.

The detectors were close to perfect on the American eighth-graders. Human writing, correctly called human.
They flagged 61 percent of the non-native English essays as machine-generated.
All seven detectors agreed on the same wrong answer for 19 percent of those essays. Consensus is not accuracy.
97 percent of the non-native essays got flagged by at least one detector. Run enough tools and everybody is guilty.

The mechanism is not mysterious. These tools measure how predictable and how uniform the writing is. Someone composing in their second language uses a smaller vocabulary and simpler constructions, so their writing scores as more predictable. The detector is not finding machines. It is finding people with less room to manoeuvre in English, and calling them cheats.

Source: Liang, Yuksekgonul, Mao, Wu and Zou (2023), GPT detectors are biased against non-native English writers, published in Patterns. Summary via Stanford HAI.

3. SO WHAT IS THIS SITE, THEN

A linter. It looks for specific habits that language models have more often than people: stock phrases, a handful of structural tics, unusually even sentence rhythm, vocabulary that circles back on itself. When it finds one it shows you where it is and what to change.

That is a useful thing to have. It is not authorship detection, and the difference is not a technicality. A sentence containing the words “delve” and “robust” and “ever-evolving” is bad writing whether a person or a machine produced it. This tool finds bad writing. Who typed it is invisible to me, because that information is not present in the text.

The score at the top of the page is a rough sum of the flags below it. It exists because people like a number. The flags are the actual product.

4. DO NOT USE THIS AGAINST ANOTHER PERSON

If you are a teacher, a manager, an editor, or a hiring committee: this number is not evidence. It cannot be evidence. Given what the research above shows, using it to accuse somebody means you will most often accuse the students and applicants writing in their second language, and you will do it while feeling objective, because a computer gave you a percentage.

There is no configuration of this tool that makes that acceptable. I would rather you closed the tab.

Use it on your own writing, where a false positive costs you nothing except a sentence you rewrite.

5. WHY IT IS FREE AND WILL STAY THAT WAY

Paid detectors charge a subscription for a guess, and they need your text on their servers to produce it. This one is a static page. The rules are a few hundred lines of pattern matching that run in your tab. There is no server to pay for, so there is nothing to charge you for, and no database that could leak because there is no database.

If it saved you from sending something embarrassing, there is a bitcoin address. It buys you nothing. There is no premium tier to unlock, because there is no premium tier.

Every rule is documented in plain english on the tells page. You should be suspicious of any detector that will not show you its reasoning, which is all of them.