About

aitextdetectly.xyz estimates whether a piece of writing came from an AI model, a person, or both.

Paste or upload text and you get three numbers (AI, mixed, and human), the sentences that read as AI, and the indicators behind the score. When the evidence points both ways, the verdict says Inconclusive.

How a scan works

Your text is judged by a language-model classifier from our detection provider, TypeSafe. It isn't judged once but three ways, each in its own request:

• the whole text,
• passages of about 50 words, split within paragraphs (up to 12 of them),
• and up to 15 individual sentences.

The AI, mixed, and human estimates are 70% the passages, weighted by length, and 30% the whole text. The sentence judgments drive the highlights, so you can see which parts of the text the score comes from.

How the verdict is decided

• Likely AI-generated only when the AI estimate is 90% or higher.
• Likely human-written when the AI estimate is 30% or lower and the human estimate is at least 70%.
• Inconclusive for everything in between, and whenever the model isn't confident, including when the whole text and its passages disagree sharply.

The bar for "AI" is set high on purpose. Calling a person's writing AI does more harm than missing some AI text.

What our testing found

We tested the detector on the HC3 dataset, which pairs human and ChatGPT answers to the same questions. On 20 human answers and 20 ChatGPT answers held back from tuning, a 90% cutoff flagged none of the human answers and all of the ChatGPT ones. At a 70% cutoff, 4 or 5 of the 20 human answers were wrongly flagged, depending on the run. That's why the verdict waits for 90%.

That is a small pilot on older ChatGPT output in English. It doesn't establish accuracy on student essays, newer models, other languages, or your writing, and no false positives in 20 doesn't mean none in real use. The percentages are model estimates, not measured probabilities.

What a score can't tell you

A score is not proof of who wrote something. Detectors, ours included, misread some writing more than others:

• Short texts. Scans need at least 50 characters, and even a paragraph gives the model little to go on.
• Drafts that were heavily edited, by a person or by a tool.
• Formal and formulaic writing, such as templates and legal or technical text.
• Writing by non-native English speakers. In a 2023 Stanford study, seven widely used detectors labeled 61% of essays by non-native writers as AI-generated, on average.

Use the result as one signal next to drafts, version history, and a conversation with the writer.

What happens to your text

We don't keep the text you scan. It goes to the detection model for that one request, and we don't sell it or use it to train models. The privacy policy has the details.

Pricing

A free account gets 3 scans a day, with no card. Paid plans are on the pricing section, and this comparison puts them next to Originality.ai, Winston AI, and Pangram.

A result looks wrong, or you have a question? Write to us.