Aeacus AI Judge
Human or AI
Paste some text and press “Start check” to find out how likely each part of the text is to be written by a human or by AI.
This check runs entirely on your own computer. No data is sent anywhere.
The first time you use it, Aeacus needs to download 87 MB to run in your browser. After that it stays on your device.
Hover the text and move your cursor around to see what Aeacus thinks about that part of the text. Tap a part of the text to see what Aeacus thinks about it.
Aeacus's raw scores are not probabilities. They are mapped to an estimated AI rate, so each rate below is how often tokens with that raw score really are AI.
Tokens past these cutoffs are labeled. The raw score each cutoff takes:
Each part of the text is colored by what Aeacus thinks of it. Stronger color means it is more sure, and grey means it cannot tell.
How to interpret these results
What do these results mean?
All AI writing checkers, Aeacus included, are imperfect. Human writing might get flagged as AI, and AI writing might pass as human. These results are not a guarantee.
Do not use these results as a verdict. These are only one signal among many. When trying to decide if some writing is AI, make sure you read and consider the whole text. Compare it against how the author writes elsewhere, and consider the context of the writing.
Why is so much of the text unsure?
Aeacus is calibrated to only flag text as human or AI when it is at least 95% sure. Otherwise, Aeacus will leave it as unsure. We err on the safe side to avoid falsely accusing or clearing any piece of writing as much as possible, so a call Aeacus does make is wrong at most about 1 time in 20.
If you used a short passage, try to add more text if available to give Aeacus more information. If Aeacus is still unsure or if you don't have more text to add, please make sure you understand what this means: it's not saying that you should or should not be suspicious. It means the test is inconclusive, and shouldn't contribute to any judgement calls you make.
How does Aeacus work anyway?
Aeacus is a machine learning model, trained on pairs of human and AI writing. We started by gathering text from before 2021, before any of these models existed, so it is guaranteed to be human written, and then used 37 different AI models to generate matching AI texts. These AI written texts cover a variety of use cases for AI: writing new text, expanding a summary, continuing a human written message to make it longer, replying to messages on social media, and AI-written “edits” inside human written text.
Aeacus then works by going over the text token by token (about three quarters of a word worth of text), and looks at it and its surroundings. Aeacus then calculates the probability that the patterns around this text are AI or human, from the AI-generated and human-written texts it has seen before. If that probability is over 95%, we flag it as AI, and if it is under 5%, we flag it as human.
How do you know Aeacus actually works?
We tested Aeacus on texts it was never trained on to ensure it is accurate. First, we take some of our data out to check and calibrate Aeacus against. The percentages below are token by token: of the tokens Aeacus flags, the share it gets right.
Next, to check if Aeacus can detect AI models it has never seen before, we have writing from almost 40 models that we never train on, including large commercial “frontier” AI models. Aeacus is 98.8% accurate against these texts as well.
Finally, we also check Aeacus against human and AI writing dated January 1, 2021 to November 30, 2022. This helps us make sure Aeacus doesn't just tell old text from new, and can generalize to topics it hasn't seen before. We see Aeacus is 99.3% accurate for these texts.