Can a detector tell whether ChatGPT, Claude or Gemini wrote a text?
Not reliably. A detector measures how closely a text resembles machine writing in general, and the large models write so much alike that style alone cannot tell them apart — even less so once the author has edited the output. Only a model's statistical watermark would give it away, and only the model's operator can check that. DetekceGPT therefore does not name a model; it shows how machine-like the text feels and in which sentences.
A detector looks for machine writing, not a model's signature
Detectors track the traits generated texts have in common: predictable word choice, evenly sized sentences, worn-out linking phrases and tidy paragraph structure. All the large models share them, because they learned from similar texts and are tuned in similar ways.
Differences between models do exist — one reaches for bullet points and headings more often, another has favourite words — but they are weaker than the differences a prompt causes. Tell a model to "write briefly" or "write like a student" and the style shifts further than the gap between two different models.
Editing mixes the traces
The text that ends up being checked is rarely raw chat output. The author shortens it, rewrites the opening, adds their own example or runs it through a second tool. Each change blends the traces together, and the result can no longer be pinned to one source.
The only trace that truly gives a specific model away is its statistical watermark. Gemini adds one and, since August 2026, so does Claude; ChatGPT has not publicly launched any. Only the model's operator, who holds the key, can verify it — an independent detector cannot work with it even when it is there.
Ask whether and where, not which model
For a decision about a text, the model's name is of no use. What matters is which passages read as machine-made, whether several detectors agree on them and whether the author can explain the content in their own words. Speculating about a particular chatbot only opens an argument nobody can win.
If a tool does name a model, treat it as a guess and test it: have two different models write a text on the same prompt and see whether the tool tells them apart. If it fails even on raw output, you cannot trust it on edited text.
FAQ
Does the OpenAI detector in DetekceGPT mean ChatGPT wrote the text?
No. The name says whose language model assesses the text, not who wrote it. The model compares the text with samples of the same type and estimates how much of it AI wrote — which AI, it does not say.
Does any model write in a way detectors cannot catch?
No model is invisible to detectors, and no detector catches everything. Text length, the prompt and how much the author rewrote after generating matter more than the choice of model.
Can text copied straight from a chat at least be recognised?
Sometimes. When invisible characters come along with the text, the Hidden characters detector finds them. Even they do not say which chat the text came from, and their absence does not prove a person wrote it.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detection