Can an AI detector tell which model wrote something (ChatGPT vs Claude vs Gemini)?
I get this question from colleagues often enough that I want to see how people here handle it. Someone pastes a paragraph into a detector, gets "likely AI", and then asks a follow-up: "Okay, but which one wrote it? ChatGPT or Claude?" A few detector sites now show a model name or a breakdown by model next to the score.
Cite
Global 100 Forum, "Can an AI detector tell which model wrote something (ChatGPT vs Claude vs Gemini)?", https://global100.org/forum/q/can-an-ai-detector-tell-which-model-wrote-something-chatgpt-vs-claude/, accessed 2026-11-10.As far as I can tell, these are two different problems. Detection asks whether text is machine generated. Attribution asks which specific system produced it. The second looks much harder to me, especially once you add that people edit outputs, mix models, and use system prompts that change the style entirely.
What I would like to know from people who have actually tried it: have you seen an attribution feature that held up when you tested it on text you knew the source of? What evidence would you want to see before trusting a label like "this is 80% Gemini"? And if a vendor says it can tell, what would you ask them?
I am specifically interested in results on text that has been lightly edited by a human, because that is the realistic case.
Write something first.
Give people something to work with: at least 30 words on what happened and what you tried.
That is too long. Keep it under 6,000 characters.
Write the question as the title, 15 to 140 characters, no links.
Pick a category.
Add a name (2 to 40 characters, no links).
That email address does not look right.
That was quick. Read the thread, then try again.
The form expired. Reload the page and post again.
Something went wrong with the form. Reload and try again.
Please complete the check and post again.
Limit reached for now. Try again later.
This thread is closed to new replies.
This thread no longer accepts replies.
Something went wrong with the form. Reload and try again.
Post a reply or question first (name and email), then this browser can vote, edit and accept answers.
You cannot vote on your own post.
Only the author (within 30 days) or the forum team can do that.
That email already has an account. Use your sign-in link instead.
New accounts are paused for the moment. Try again later.
That was already posted.
6 answers
- Priya N.Member ·
I tried one for a seminar demo. I generated paragraphs from three different assistants with the same prompt and fed them in. It got some right and some wrong, and what stood out was that its confidence did not seem to track whether it was right. A wrong label looked just as sure as a correct one.
I would not trust a model label without a published test on text the vendor had not trained on. And even then I would want to know how the samples were produced, because a clean paragraph straight from the model is nothing like what a student submits.
- Sam WhitlockMember ·
@Priya N. that confidence point is the part that worries me most. A label that is sometimes wrong is a tool. A label that is wrong with the same assurance as when it is right is a liability, because the reader has no way to tell which is which.
- Sam WhitlockMember ·
Threat model hat on. Attribution only makes sense if the author does not care about being identified. The moment someone cares, they can change the prompt, run the text through a second tool, or edit by hand, and whatever fingerprint existed gets smudged. Which is exactly the realistic case you describe.
So my answer to "what would you ask the vendor" is: what happens to your label when I rewrite one sentence in three?
- Dan OkaforMember ·
From the classroom side, nobody I know needs the model name. What matters in a conversation with a student is whether the work reflects their own understanding. "Which model" never settles that, and a wrong model label can make an already shaky accusation worse. If a tool gives you that label, I would treat it as noise.
- Hana SatoMember ·
@Lena B. I would be careful with that, kindly. Those word lists describe a particular model version at a particular time, and vendors change their models and their default styles regularly. A word that was a giveaway last year may be gone. More importantly, "right most of the time" is hard to measure on your own: you mostly remember the cases where the guess was confirmed. If you could test yourself blind, with edited text from all three, I would genuinely like to see the result.