MMOOZZGG.com

AI literacy

How to check text for AI and why detectors are wrong

8 min read5 sources

You cannot check text for AI reliably today: detectors give a probability, not an answer, and they make mistakes in both directions. OpenAI shut down its detector on 20 July 2023 because of low accuracy. So do not look at a percentage; look at the facts, the sources and how the author explains the text.

On this page · 5 sections

How detectors work and why they are wrong

An AI text detector is a program that estimates whether a text looks like one a neural network wrote. It looks for predictable words and even phrasing and gives a probability, for example “most likely written by AI”.

A neural network here means a large language model: a program that writes text word by word, each time choosing a likely continuation. A person can also write evenly and predictably, so style alone does not show who the author is.

OpenAI gave a telling example. The company released an AI text classifier on 31 January 2023. On a test set of English texts it recognized only 26% of AI texts and in 9% of cases took a human text for AI.

On 20 July 2023 OpenAI shut down the classifier because of low accuracy. Already at launch the company warned that it was very unreliable on texts shorter than a thousand characters, worked much worse in other languages than in English, and that edited AI text fooled it.

People make just as many mistakes. The International AI Safety Report 2026 cites a study in which participants took AI text for human text in 77% of cases. The report’s authors write that detection tools cannot keep up with the quality of generation.

Mistakes in both directions mean two things. The detector misses neural network text that someone edited a little. And it can accuse a person who writes simply, briefly and without striking phrases.

Signs of text written by AI

The signs below say only one thing: the text deserves a closer look. None of them proves that a neural network wrote it. People write this way too, and five minutes of editing remove almost every trace.

General words instead of details. The text is smooth, but it has no names, dates, places, personal experience or things only the author knows.

The same rhythm. Paragraphs of equal length, each ending with a conclusion, lists of three items, and an introduction that retells the question.

Confident inventions. A figure without a source, a quote that cannot be found, a reference to an article or a law that does not exist. Such errors are called neural network hallucinations.

An answer that misses the task. An essay on a book retells it in general words but mixes up the characters or gives details that are not in the book.

Bad: “The detector showed a high AI percentage, so the text was copied.”

Good: “The text has three figures without a source and confused dates. I will check them and ask the author where they came from.”

Checking facts and sources

The question “is this true” often matters more than “who wrote it”. The International AI Safety Report 2026 notes that AI systems still sometimes give false information. So a fact check is useful for any text, whoever wrote it.

Step one: write down all figures, dates, names and quotes. Find each one in the original source: on the agency’s website, in the text of the law, in a dated report.

Step two: open every link. A neural network can give a plausible address that leads nowhere, or a real article that does not contain the figure.

Step three: check the dates. A law may have changed and a figure may be outdated, so look at when the document came out.

Bad: “The text has links, so everything is checked.”

Good: “I opened three links of five: one does not have this figure, the second does not open. I am asking the author for the original source.”

If you use a neural network yourself, ask it to name its sources and check them the same way. How to ask such questions is taught in a free course.

What a teacher, an editor and an employer should do

The common rule for all three: a detector’s percentage cannot be the only basis for a decision. OpenAI wrote this too: its classifier should not be used as the main decision tool, only together with other ways of checking.

For a teacher. Agree with the class in advance where a neural network may be used and where not: for example, it is fine for a plan but not for a finished essay. If a text raises doubts, ask the student to explain out loud how they wrote it.

Bad: “The detector found AI, so you get the lowest grade.”

Good: “Tell me why you chose this example and where you got this date.”

For an editor. Check the facts, the sources and the author’s own thinking. Write down in the contract or the editorial rules whether a neural network may be used and whether this must be reported.

For an employer. A test task done with a neural network is not always a minus: what matters is whether the candidate understands the result. Ask a couple of questions about the solution in the interview, and set the rules for working with AI in an internal document.

If you accused someone wrongly, say so directly. The OpenAI classifier gave false positives on human texts in 9% of cases, and other programs have them too.

Labeling AI content in Russia

Labeling AI content is a “created with AI” mark that a person sees or a program reads. In Russia it is set by Federal Law No. 243-FZ of 26 July 2026 “On support for the development of artificial intelligence technologies in the Russian Federation”.

The Government introduced the bill. According to its record in the State Duma system, it was passed on 8 July 2026, the Federation Council approved the law on 17 July, and it was published on 26 July 2026. While it was still a bill, on 3 July 2026, the office of the deputy prime minister said that labeling for users would be voluntary.

The labeling rules (Article 9) take effect on 1 March 2027. Anyone who creates sound or an image with a large AI model will be able to add a mark that AI was used. Platforms with personal pages and a daily audience in Russia of more than 500,000 people must give users this option.

The law does not oblige an ordinary person to label their own text written with a neural network. But the rules of a school, university, editorial office or employer may require reporting it.

Amendments are already in preparation: bill No. 1317978–8 on changing Article 9 regarding labeling was registered in the State Duma on 17 August 2026 and was still under review as of 16 October 2026. What the law means for companies is covered in the article on law 243-FZ.

Questions and answers

Can you tell for sure that AI wrote a text?

No. Detectors give a probability and make mistakes in both directions, and OpenAI shut down its own classifier in 2023 because of low accuracy. It is more reliable to check the facts, the sources and how the author explains the text.

What if the detector was wrong?

If a detector accused you, ask for a check on the merits: drafts, the edit history in the document, notes and a spoken account of the work show authorship more convincingly than a percentage. Remind them that even OpenAI advised against making decisions based on its classifier alone. If you were the one who got it wrong, apologize to the author and judge the text by its quality.

Do you need to label text written by AI?

Under Russia’s AI law 243-FZ, an ordinary user does not have to label a text written with a neural network. From 1 March 2027, large platforms must let users add such a label, and the rules of a school, an editorial office or an employer may require you to report AI help. If in doubt, ask about the rules in advance.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. OpenAI (Internet Archive copy of 31 December 2024; openai.com returns 403), 31 January 2023 (web.archive.org)
  2. State Duma bill database (SOZD), 26 July 2026 (sozd.duma.gov.ru)
  3. International AI Safety Report 2026, 3 February 2026 (internationalaisafetyreport.org)
  4. ConsultantPlus, 243-FZ, Article 9, 26 July 2026 (consultant.ru)
  5. State Duma bill database (SOZD), bill No. 1317978–8, 17 August 2026 (sozd.duma.gov.ru)

Found a mistake? Tell us

Publications

Contact us