AI text detection
A procedure that estimates whether a text was written by a human or artificial intelligence – but it provides only a probability, not proof.
In plain terms
An AI text detector is a program into which you paste a text, and it tells you the probability that it was written by a machine. It usually gives a percentage, for example, "78% likely AI-generated".
Imagine it like a metal detector on the beach: it signals if something suspicious is found, but it cannot tell whether it found a gold ring or a beer cap. The signal is only useful for examining the matter more closely; no decision can be based on it.
This is how it differs from a plagiarism detector. A plagiarism detector can show the original source from which the text was copied. An AI detector cannot show such a source; it only infers from the text's style and statistical characteristics.
Schools, newsrooms, HR departments, and social platforms use it because they want to know what lies behind genuine human work. However, its reliability is limited, and this is worth knowing for anyone using it or being assessed by it.
How it works
- The detector first measures how predictable the text is. Language models mostly choose from probable continuations, so machine text is often "smoother" and less surprising than human text.
- It also checks for variety. Humans write short and long sentences, sometimes using unexpected words, while the rhythm of machine-generated text is more uniform.
- Many tools use a trained classifier model for this: it learned on vast amounts of human and machine text what patterns are characteristic of one or the other.
- Another approach is watermarking. In this case, the AI provider hides an invisible statistical pattern in word choice during text generation, which a checking program later finds. This only works for marked texts, and the pattern may fade after heavy rewriting.
- The final result is a probability value. It can err in both directions: it may mistake human text for machine-generated, and rewritten or paraphrased machine text may slip through.
An everyday example
Your child submits a homework assignment they wrote themselves, and the teacher replies that the detection program indicates it was 80% AI-generated. In such cases, you should not trust the program: drafts, document editing history, and a brief oral conversation reveal much more about who wrote the text.
Who is affected
A common misconception
Limits and risks
- It may falsely accuse you: it can classify text you wrote by hand as machine-generated, which is hard to refute later.
- It is easy to bypass: simply rephrase the machine-generated text, rewrite it slightly, or "humanize" it with another tool.
- Most tools were developed and tested on English texts, so their reliability on Hungarian text is less certain.
- Do not base an important decision, such as a grade, disciplinary action, or rejection, on a single recognition result. Treat it at most as a signal warranting further verification.
Similar, but not the same
- Plagiarism detection – A plagiarism checker compares the text with existing sources and shows matches, while an AI detector estimates based on style without sources.
- Watermarking – The AI provider hides the watermark in the text itself when it is created; detection afterwards tries to infer it without the provider's involvement.
- Deepfake detection – Deepfake detection searches for forged images, audio, and video, while AI text detection deals only with written text.
Related concepts
AI content detection, Content authentication, AI safety
Text watermarking
Plagiarism detection, Deepfake detection, Generative AI, Large language model, Classification model, Transparency obligation
A brief history
Machine text detection tools existed as early as 2019 (e.g., GLTR), but the topic moved to the center of attention after ChatGPT's launch in November 2022. In early 2023, detection tools appeared one after another, such as GPTZero. OpenAI released its own detector in January 2023, but withdrew it later that year in July due to low accuracy.
Where things stand
At the end of 2026, it is widely used in education and the content industry, but its reliability remains debated, so several institutions treat it only as an indicative signal. The focus is increasingly shifting toward watermarking and labeling machine-generated content. Under Article 50 of the EU AI Act, AI providers must place machine-detectable labels on generated content; this rule has applied since August 2026, with systems already on the market granted a transition period. The Act mandates labeling, not the use of detection tools.
In the news
- AI detectors can falsely accuse students: what the false-positive data showOct 12, 2026 · Education & Family
- GitHub makes AI Scan metrics, Agentic Autofix and AI secret detection generally availableOct 12, 2026 · Code
- ChatGPT may add an invisible marker to AI-written text, Médiapiac reportsOct 7, 2026 · Marketing
- US universities turn away from AI detectors amid 'homework scams'Oct 5, 2026 · Education & Family
- Dartmouth authorizes Pangram AI detector as doubts over false positives persistSep 28, 2026 · Education & Family
- AI detectors are pushing colleges back to handwritten learningSep 22, 2026 · Education & Family
- Universities in South Korea, Dublin and Norway act against AI cheatingSep 15, 2026 · Education & Family
- AI detectors prove unreliable after paraphrasing; watermarking raises concerns for non-native writersSep 7, 2026 · Education & Family
Also known as
AI text detection · AI-detektor · AI text detection · Machine-generated text detection · AI text detection · AI content detection · AI detector · AI writing detection
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.