Does AI watermark the text it gives me?
An invisible text watermark is not a hidden character in the file. It's a pattern in which words the model picks — the difference between choosing overcast or cloudy when either would do. The text reads normally, the pattern travels when you copy and paste, and light editing usually doesn't erase it.
Who marks what, right now
SynthID has been built into Gemini's text output since late 2024 — the longest-running text watermark in the business.
Models launched after August 2, 2026 mark their text automatically. There's no setting to turn it off, and older models are being switched over too.
OpenAI built a watermark years ago and never shipped it. Its images and audio carry provenance signals; its text carries nothing.
The part nobody advertises: you can't check it
The mark is provenance — a record of where the text came from. But the only company that can detect a mark is the company that applied it. Google checks for Google's mark, Anthropic for Claude's, and the detection tools are aimed at regulators and newsrooms, not at you.
So if a client, a platform, or a publisher ever asks whether your content is AI-generated, you can't run the check yourself — you'd be asking the vendor to vouch for your text.
What the research says about safety
One more thing worth knowing, because vendors don't headline it. Research published in September 2026 found that watermarking can change how a model behaves under hostile prompting — in tests on open models, marks made some models more likely to comply with requests they'd normally refuse, and the effect shifted depending on which secret key was used.
Honest limits: that study tested open models with the published watermarking recipe, not the exact deployments at Gemini or Claude. But the direction is clear enough — a watermark changes more than word choice, and vendors ship it silently. Treat safety claims as something to re-test, not something to assume.
What this means for a small business
- Marked doesn't mean machine-written. If you draft with AI and then edit heavily, the text can still carry the mark. It says the text was processed by AI, not that AI wrote every word.
- Unmarked doesn't mean human. ChatGPT text carries no mark at all, older models predate marking, and a thorough rewrite or a translation can wash the signal out.
- Publish accordingly. If you stand behind AI-assisted content, say so yourself — don't rely on either the presence or the absence of a watermark to make the case for you.
Key point — The watermark is between the AI company and the regulator, not between you and your customers. Treat it as a signal you don't control, and let your own disclosure be the honest one.
