ChatGPT gets most of the attention, but plenty of writing now comes from Claude, Gemini and other assistants. This guide explains how Winston AI handles text from these models, what a result can and can't tell you, and how editing affects what you see.
Key takeaways
- Winston AI flags machine-written text, including text from Claude and Gemini.
- It doesn't identify which model wrote something, only how likely the text is to be AI.
- We retrain as new model versions ship so the detector keeps up.
- Editing and mixing in human writing change the result, often toward yellow (mixed).
Why people ask about Claude and Gemini specifically
Many people assume detectors were built around one chatbot and only recognize its style. If a student or freelancer used Claude from Anthropic or Gemini from Google instead, would a check miss it? It's a fair question, because different assistants do have somewhat different default voices. It usually comes up when a reviewer has a draft that feels machine-written but doesn't match the ChatGPT style they've learned to recognize.
How Winston AI handles text from different models
Winston AI is trained on writing from the models people actually use, including Claude and Gemini, alongside a wide range of human writing. Rather than memorizing one chatbot's habits, it learns the broader patterns that tend to separate machine-written text from human text. When new versions of these models ship, we retrain so the detector keeps up. That matters because assistants are updated often, and their default style can shift from one version to the next.
The result looks the same whichever assistant was involved: an overall score plus sentence highlights. Red marks sentences that read as likely AI, yellow marks mixed signals, and unhighlighted sentences read as human. Hover or tab to any sentence to see its AI probability, which helps when only one part of a document seems off.
How assistant writing often reads
Default output from modern assistants, Claude and Gemini included, tends to share some habits a careful reader can notice:
- Clear, tidy structure, often with headings or lists even when the format doesn't call for them.
- A careful, balanced tone that acknowledges several sides before settling on a mild conclusion.
- Polite framing and caveats wrapped around claims.
- Fluent prose with few of the small slips and quirks of personal writing.
- A closing paragraph that neatly sums up what came before, even in short pieces.
None of these is unique to one assistant, and people can prompt any of them into a different style. Our pages on Claude and Gemini cover the writing habits reviewers tend to notice in each.
What editing does to the result
AI text is rarely submitted untouched. People rewrite sentences, add their own examples, combine output from more than one tool or run it through a paraphraser. All of this changes what a detector sees, and reviewers should expect it. It's also why a result describes the text in front of you, not the history of how it was made.
- Light edits tend to leave the underlying AI patterns largely in place.
- Heavier rewriting often shows up as yellow, mixed-signal sentences.
- Mixed documents, where a person wrote some sections, appear as a patchwork of highlighted and unhighlighted passages.
- Thorough human revision can make text read as human, because by then much of the writing genuinely is.
Checking a draft you think came from an assistant
- Paste the text into the AI detector. It works in 14 languages and detects the language automatically.
- Look at where red and yellow sentences cluster rather than only at the overall score.
- Compare the piece with the writer's earlier work and any drafts or version history.
- If you still have questions, talk to the writer about how they produced it and which tools they used.
A result is evidence to weigh, not proof, whichever model was involved. If the writer can walk you through their drafts and choices, that counts for more than any highlight. For more on what sits behind a result, see how AI detectors work.