An AI text watermark is a hidden pattern in the words an AI model chooses. It isn’t a label or hidden characters, and you can’t see it. Anthropic marks text from supported Claude models, and OpenAI has said it will add a watermark to eligible ChatGPT and Codex text in the EU. A detector can only suggest that a model wrote or processed part of a passage. It can’t say how much, who wrote the rest, or whether a person drafted it first, and it doesn’t identify you. If you write your own drafts and use AI only to tidy them, keep your own record: dated copies of each draft from before any AI pass, timestamped somewhere you don’t control.
Table of Contents
What an AI text watermark is
When a model writes, it picks one word at a time, and often several words would do equally well. A watermark uses those small, low-stakes choices to leave a pattern that a key can later pick up.
OpenAI says its method “adds an invisible statistical signal to the model’s word choices.” Anthropic says of Claude’s: “Nothing is added to the text and there are no hidden characters”. So there’s nothing to find or strip out by hand. The pattern is in which words were chosen.
Who can check for it today
Not you, at least not yet. Detection is restricted:
- OpenAI: “Access will initially be limited to approved researchers and expert organizations”.
- Anthropic: its detector is in private preview for eligible organisations, such as regulators, media, fact-checkers, researchers and educational organisations, and for businesses that must check watermarks under EU law.
That means a publisher, school or competition may one day be able to run a check, while the writer can’t check the same text first.
What a watermark can and can’t show
Both companies are clear about the limits.
- It’s about part of a passage, not the whole. OpenAI says a watermark can indicate its system generated or processed part of a passage. Its own words: “A watermark does not measure human contribution.”
- It doesn’t name anyone. “A watermark does not identify the user,” says OpenAI. Anthropic: “Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat”.
- Editing changes it. “Editing can weaken the watermark,” says OpenAI, and short or tightly constrained text is harder to detect.
- No watermark settles nothing either way. Both companies say that not finding a watermark doesn’t mean a person wrote the text. It may be too short, edited or translated, or come from another model.
Why “a quick proofread counts” worries writers
The Claude Help Center notes: “People often use Claude to proofread, translate, summarize, or convert files.” Its output can carry a Claude mark even when the ideas or text came from someone else, including you.
Anthropic also explains the limit: “The watermark only applies to words Claude chooses.” When it proofreads your work and changes only a little, “there’s very little (if anything) for the watermark to attach to.” The more the model rewrites, the more room there is for the pattern.
So a light proofread may leave little or nothing to detect, while a heavy rewrite may. Either way, a watermark result alone can’t tell a reader that the story, argument or research is yours. That’s the gap your own records can fill.
Keep a pre-AI draft trail
The useful evidence is the draft you wrote before any AI tool touched it, plus a record of when it existed.
- Save a copy before every AI pass. Keep it as its own file, such as
chapter-3-before-edit.docx, and don’t overwrite it. - Keep your notes and outlines too. They show how the work grew.
- Record when each draft existed. With TimeBinder®, you TimeBind the file and its SHA-256 fingerprint is recorded on the Bitcoin blockchain. That shows this exact file existed no later than the recorded time. Your files are hashed in your browser and never uploaded, and you can include up to 10 files in one TimeBind. See how a TimeBind works.
- Log daily progress if it helps. The free Draft Streak tool counts words and makes a SHA-256 fingerprint in your browser, with nothing uploaded. Its log uses your device’s clock, so on its own it isn’t a timestamp. TimeBind the same file when you want a recorded time.
There are two ways to TimeBind:
- Pooled (1 Token): Your fingerprint is grouped with other users’ in one shared Bitcoin entry, usually recorded within 24 to 48 hours, and often sooner, depending on Bitcoin network activity. It’s verified through TimeBinder.
- ASAP (3 Tokens): It gets its own Bitcoin transaction, sent straight away, and can be verified independently from the certificate and the Bitcoin blockchain.
The ASAP vs Pooled verification guide explains the difference. For more on working openly with AI, see using AI as a writing tool and showing your own work.
What a TimeBind does NOT do
- It doesn’t show who made the file.
- It doesn’t show the contents are true.
- It doesn’t show whether AI was used.
- It can’t say how a publisher, school, court or anyone else will judge it.
Watermark, AI detector or timestamped draft?
| What it can show | Who controls it | |
|---|---|---|
| AI text watermark | That a model may have generated or processed part of a passage | The AI company, and the organisations it lets check |
| AI detector | A tool’s own estimate about the text; how it works and how reliable it is varies by tool | Whoever runs the tool |
| Timestamped draft | That this exact file existed no later than the recorded time | You keep the file; the time is recorded on the Bitcoin blockchain |
They answer different questions. Our guide to working drafts and timestamps vs AI detectors goes further.
When TimeBinder isn’t the right tool
- You want to know if a text has a watermark. Only the AI companies’ detectors can check, and access is limited.
- You want to show who wrote something. A timestamp records when a file existed, not who made it.
- You want your work judged. How a publisher, school or contest treats AI use is up to them, so check their rules.
- You didn’t keep the earlier drafts. A TimeBind can only record a file that exists today. Start saving them now.
TimeBinder pricing
- Free: $0. 1 Token a month. No credit card required.
- Essential: $20 a month for 5 Tokens.
- Expert: $50 a month for 15 Tokens.
- Business: $149 a month for 60 Tokens.
- Business Plus: $299 a month for 150 Tokens.
- Enterprise: custom.
A paid plan replaces the Free plan’s Token. A Pooled TimeBind uses 1 Token and an ASAP TimeBind uses 3. Prices are in US dollars. See all plans. Service terms for every plan are set out in the TimeBinder Terms and Conditions.
Frequently Asked Questions
-
Does using AI to proofread leave a watermark?
It can, on the words the model changes. Anthropic says its watermark only applies to words Claude chooses, so a light proofread of your own text leaves very little, if anything, for it to attach to. A heavier rewrite leaves more.
-
Can I check my own text for a watermark?
Not at the moment. OpenAI says access will at first be limited to approved researchers and expert organisations, and Anthropic’s detector is in private preview for eligible organisations.
-
Does a watermark identify me?
No. OpenAI says a watermark does not identify the user, and Anthropic says its watermark carries no identifying information and can’t be traced to a person, organisation or chat.
-
If no watermark is found, does that mean a person wrote the text?
No. Both companies say a missing watermark doesn’t settle that. The text may be too short, edited or translated, or come from a model that doesn’t use the same watermark.
-
What does a timestamp show?
That this exact file existed no later than the recorded time, and later that a file you show still matches its fingerprint. It doesn’t show who made the file, whether the contents are true or whether AI was used.
-
Does TimeBinder read my manuscript?
No. Your files are hashed in your browser and never uploaded. Only the SHA-256 fingerprint is recorded.
Start your draft history
Save your next draft before any AI pass, then record when it existed. The Free plan includes 1 Token a month, and no credit card is required. Start a dated draft history
