Sometimes, but not reliably, and rarely by name. A detector can often tell that a text reads like machine writing. It cannot tell you which model wrote it, and editing weakens the signal for every model.
The answer changes with the model and with the kind of detection. If you want to detect Claude writing in a stack of essays, or detect Gemini text in a client draft, start with what each method measures.
Two kinds of detection
Style scoring is what nearly every public checker does. It reads the finished text and compares its rhythm, word choice, and structure with known AI output.
It works on text from any model, needs only the words, and gives you a probability score.
Watermarking works differently. The company that runs the model hides a statistical pattern in the words as it writes them. Only someone with the right key can check for it.
That can make it far more precise, but it exists only when the provider chooses to add it, and the ordinary checkers that anyone can open on the web cannot read it at all.
How Claude looks to a detector
Claude drafts often hedge claims and give both sides. Many follow a pattern of a concession, then a counterpoint, then a sentence that joins the two. Paragraphs are orderly and polite.
The ending often restates the task calmly instead of closing with a flourish.
A Claude AI detector can pick up these habits in unedited output. The same habits describe a lot of careful human writing, though.
A cautious, well-organized academic essay shares many of them, so any attempt to detect Claude writing carries a real risk of false alarms on human work.
Light editing helps a lot. Adding a firm opinion, cutting the "on the other hand" passages, and varying the paragraph lengths all move a draft away from its usual pattern.
If you want to see how a draft reads, a free Claude AI detector scores the finished text and highlights the lines that read most like machine writing. It sees only the words you paste in, never your chat or your prompt.
How Gemini looks to a detector
Gemini reads like a briefing. It leads with a direct answer, adds a few supporting points, and keeps sentences short and easy to scan.
Facts arrive in a row, and the voice stays detached, like a summary written by someone who was not there.
Style-based tools can measure that flat, ordered rhythm. A Gemini AI detector that highlights the passages driving its score shows where a draft reads that way, though it judges how the text reads and says nothing certain about where the text came from.
The watermark factor
Gemini has a published text watermark system. Google DeepMind described SynthID-Text in a 2024 paper in Nature. The team ran a live experiment on nearly 20 million Gemini responses and reported that text quality held up.
The same paper lists limits. Edits to the text weaken the watermark, and that includes paraphrasing by another model, and text with very little natural variety, such as a short factual answer, gives it less to work with.
Reading it takes the provider's key. Public checkers cannot do that. Outside Google, you still have to rely on style to detect Gemini text.
How ChatGPT compares
ChatGPT output often has a confident, upbeat tone, stock transitions, lists of three, many headers and bullets, and summary-style endings.
All three profiles in this article are tendencies that shift between model versions, change with the prompt, and fade when someone asks for a different tone. Treat them as leads to check, since none of them proves who wrote a text.
Why detector results go stale
Models get updated. Detectors have to keep up. The RAID benchmark, presented at ACL 2024, tested 12 detectors on more than 6 million generations from 11 models.
The authors found that current detectors are easily fooled by simple attacks, by changes in how the text is generated, and by models the detector has not seen before. The full paper is on the ACL Anthology.
A detector's accuracy on last year's model therefore says little about this year's. Mixed drafts are harder still.
A draft with one paragraph from Claude, two from Gemini, and a human rewrite on top can dilute each pattern past the point where any tool can call it.
Can any tool say which model wrote a text?
Not reliably. A public checker can say that a text reads like AI. Naming the model is a much harder job. The score from a Claude AI detector or a Gemini AI detector measures general machine style with that model's habits in mind, and it does not prove who wrote the text.
The one method that can tie text to a single provider with real confidence is a watermark, and only the provider can check it.
If you write with AI help
The model you used matters less than how much of you is in the final draft, so edit for your own voice, add details only you would know, and keep your drafts.
If you want drafts that start closer to your own voice from the first prompt, you can train an AI tool to write in your style before you edit.
Check before you submit. The check shows which lines still read flat, whichever model wrote them.
If you switch between Claude and Gemini on one piece, check the final combined version rather than each part alone, since that is what a reader or reviewer will see.
If you review other people's writing
Treat any score as a reason to look closer. It is never a verdict. Ask for drafts and notes, and talk with the writer about the work.
Careful, hedged prose is exactly what good students and second language writers are taught to produce.
Teams that publish with AI help can write clear editorial standards before they ask anyone for a score, and they can add the draft check to the review step of their digital content workflow.
Questions about AI detectors, Claude and Gemini
Can an AI detector tell Claude from ChatGPT?
Not reliably. A detector scores how much a text reads like AI output. It does not name the model that wrote it.
Is there a Gemini AI detector that reads the watermark?
No public checker does. Style-based tools can flag Gemini text, and their results shift with editing and with the model version.
Do AI detectors flag human writing?
Yes, sometimes. Careful, hedged, and well-organized human writing shares habits with AI output, so false alarms happen.
Does editing change the result?
Yes. Edits weaken the style signals for every model, and they weaken the Gemini watermark too, and heavy edits and paraphrasing weaken both more than light edits do.
Where this leaves you
Style checks can catch unedited output, and they struggle once a person has worked on the text.
A watermark is more precise but works only for the provider that added it. Use a detector as one signal among several.
Judge the writing first.