That is not in the least bit how a tool like this works.
All AI detection is extremely unreliable, but they operate on principles which, if the assumptions supporting them were true, would be sound. The way you imagine they work is different: you’re describing a “novel text detector” which is not at all the same thing as an “AI text detector”.
AI detection works, at a very high level, by throwing a bunch of examples of AI text, and a bunch of examples of non-AI text, into a machine-learning model, and training it until it is able to recognise the AI examples as such. It doesn’t work by throwing in all existing text including novels written before AI, because that would never do what you want.
(The reason, if you’re interested, why this ends up not working is because the features such a model identifies as indicative of AI text are very sensitive: if you train it on Claude and ChatGPT, it will not work on Gemini output. If you train it on Gemini, when Gemini updates it will get worse. If someone generates text with a weird prompt, it may slip by. If someone writes in a weird way, it may get flagged. And if any AI company wanted to defeat AI detectors, it could trivially feed its output through one during the training process and give that output as examples to avoid in the training, so that the model learns to create output which doesn’t “look like” AI output to those detectors.)
That is not in the least bit how a tool like this works.
All AI detection is extremely unreliable, but they operate on principles which, if the assumptions supporting them were true, would be sound. The way you imagine they work is different: you’re describing a “novel text detector” which is not at all the same thing as an “AI text detector”.
AI detection works, at a very high level, by throwing a bunch of examples of AI text, and a bunch of examples of non-AI text, into a machine-learning model, and training it until it is able to recognise the AI examples as such. It doesn’t work by throwing in all existing text including novels written before AI, because that would never do what you want.
(The reason, if you’re interested, why this ends up not working is because the features such a model identifies as indicative of AI text are very sensitive: if you train it on Claude and ChatGPT, it will not work on Gemini output. If you train it on Gemini, when Gemini updates it will get worse. If someone generates text with a weird prompt, it may slip by. If someone writes in a weird way, it may get flagged. And if any AI company wanted to defeat AI detectors, it could trivially feed its output through one during the training process and give that output as examples to avoid in the training, so that the model learns to create output which doesn’t “look like” AI output to those detectors.)
I was being overly simplistic - I meant more that the patterns it’s trying to detect were created by an LLM trained on the data they’re inputting.
It’s like how reddit comments from 2016 look generated. If you stick one of those into an AI detector, it’ll give a false positive for the same reason