Does GPTZero detect ChatGPT?

How it works

It scores two properties and runs a classifier over them. Perplexity asks how predictable each word is given the ones before it, and burstiness asks how much the sentence lengths swing. Raw ChatGPT prose is unusually even on both counts, so it sits well inside the flagged region and GPTZero reports at document, sentence and word level rather than as one number. What it never does is name a tool. The report says this reads like assistant writing, and ChatGPT, Claude and Gemini all produce writing that reads that way.

The part people miss

Two things move the result more than people expect. Editing is the first: rewrite the passage in your own rhythm and the score falls, which is also why an honest draft you polished heavily can score like generated text. The second is the reverse case, the one that brings most people here. Plain, careful, uniform English scores low on exactly the two measures above whether a model wrote it or not, and GPTZero's own guidance says a result should not be used to punish a student and belongs in a wider assessment of the work.

Does it matter which ChatGPT model wrote it?

Not in the way people assume. GPTZero does not keep a list of model outputs to compare against, so there is no per-model lookup that a newer release could fall outside of. It measures the shape of the writing, and every instruction-tuned assistant produces a similar shape.

What does change is how much shape there is to measure. Newer models write with more variation than the 2022 release GPTZero was first tuned against, and a model told to write in a specific voice varies more again. So the practical answer is that a newer or more carefully prompted model tends to score lower, without any of it being about the model's name.

If someone is telling you a detector identified a specific product, ask to see where the report says so. It does not.