Skip to content

Sign in / Sign up
ParaTrace

Unlimited free credit top-ups, for a limited time. Running low? Ask for more credits and we top you up, as often as you need. This offer ends soon, so catch it while it lasts.

AI text detection

Was this text written by a person, or generated by AI?

Paste a text and ParaTrace tells you in seconds, marking the parts that look AI-written. Try it free, with no account needed.

0 words

Free checks aren't stored. To count them, the server keeps your network address in memory until midnight UTC. Privacy policy

The verdict

Result

  • 99.6 %

    of AI answers caught

  • 0.8 %

    of human texts wrongly flagged

  • 90

    AI models tested

Level with Pangram on AI-written essays, and like Pangram it flags none of 106 research papers written before ChatGPT. Compare with Pangram

Sign in / Sign upfor longer texts, PDFs and a history, with 100 free credits

By the numbers

  • 99.6 %

    of AI-written answers detected

    26,082 answers · 26 developers

  • 0.8 %

    of human texts wrongly flagged

    22,691 held-out texts, same setting

  • 96.0 %

    still detected under evasion attacks

    11 attack types

  • 20 ms

    to score a 300-word text

    on our servers

Read the benchmark report

Head to head with Pangram

Within 5 points of Pangram 4 on 15 of 21 results

Pangram is one of the leading AI text detectors. We scored ParaTrace v7 on the same public tests Pangram reports on, from student essays to novels, news and reviews. It matches Pangram on fully AI-written and AI-paraphrased essays, catching 100 % of them, misses fewer AI texts than Pangram 3 on writing by well-known authors, and is within 5 points of Pangram 4 on 15 of the 21 results.

See all 21 results against Pangram

Within 5 points

  • Against Pangram 415 of 21 results
  • Against Pangram 316 of 21 results

Results where ParaTrace is level, ahead, or no more than 5 points behind.

Scientific papers

No false alarms on research papers written before ChatGPT

NeurIPS checked its 2026 position papers for AI-written text with Pangram, after testing it on conference papers from 2022 and 2025. We ran the same check with ParaTrace v7. It flags none of the 106 papers from before ChatGPT, exactly like Pangram, and flags 9 of 123 papers from 2025 as at least half AI-written, 7.3 % against 1.0 % for Pangram.

See the field test on scientific papers

Papers from before ChatGPT

0of 106 papers flagged by ParaTrace, at any AI score

The same result as Pangram, which flagged 0 of 159.

How it works

  1. Sign up

    Continue with Google and your account is ready at once, with 100 free credits.

  2. Submit text

    Paste text or upload one file, a PDF with selectable text or a .txt or .md file. Each text costs 1 credit per started 512 tokens, about 400 words.

  3. Read the verdict

    For each document you get the likelihood that it was AI-generated, how confident the verdict is, and any reliability warnings.

Tested across the AI market

Detection rate on answers from the 12 developers with the most answers in our test, out of 26.

  • OpenAI99.4 %
  • Google99.4 %
  • Anthropic99.6 %
  • Meta99.8 %
  • Mistral99.7 %
  • Alibaba99.8 %
  • DeepSeek99.7 %
  • Reka99.9 %
  • Microsoft99.8 %
  • 01.AI99.8 %
  • Cohere99.4 %
  • xAI99.8 %
All developers, generalisation and methods

Late-2025 flagship models

  • Anthropic98.7 % of 318
  • Google99.3 % of 667
  • xAI99.0 % of 102
  • OpenAI99.2 % of 357

Answers were collected up to October 2025. Newer models are not yet benchmarked.

Notice to readers

Know the limits

This is a statistical estimate, not proof. Don't use it as the sole basis for decisions about a person.

How detection works Where to be careful

  • English only. Texts in other languages are not supported.
  • At strict, 0.8 % of held-out human texts (about 1 in 120) were wrongly flagged as AI in our tests. At lenient, about 5 %. Other kinds of writing can differ.
  • Persuasive essays written by adults are a known weak spot. About 1 in 5 (18 %) is wrongly flagged at strict.
  • Texts under 50 words give less reliable verdicts. In our tests 2.8 % of short human texts were wrongly flagged.
  • Long texts are scored in overlapping chunks and labelled part by part, so a text can come out partly AI-generated. Where the AI part starts and ends is approximate.