Claude's Text Watermark Explained: Can Anyone Tell You Used Claude?
Anthropic hides an invisible watermark in the text Claude writes, to comply with the EU AI Act. It costs you nothing and carries nothing about you. As of October 12, 2026 nine models mark their text, and three of those, Sonnet 5, Sonnet 5.5 and Haiku 5.5, are on the free plan. Every listed model tags the files it generates.
If you have ever wondered whether anyone can tell that Claude wrote your email, Anthropic has answered it in more detail than any AI company has before. On August 14, 2026 it published how Claude's text watermark works, confirming that Claude models will generate text containing an invisible watermark. The reason is regulatory, not commercial: the EU now requires AI providers serving its market to mark AI-generated content, and Anthropic is applying the mark worldwide rather than only in Europe.
The list keeps growing, and the most recent additions reach the free plan. Anthropic's help centre article on marking now carries a sixteen-row table whose own revision stamp reads October 8, 2026 at 23:39 UTC, and nine of those sixteen models watermark their text: Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5. Sonnet 5 is the one that matters most, because it is the older free-plan Sonnet and it was unmarked a fortnight ago. Separately, every model on Anthropic's list attaches a Content Credential to files it generates, Haiku 4.5 and the whole 4.x range included.
The practical news for you is still mostly reassuring. Nothing about Claude's writing changes, nothing is added to your text, and the mark says nothing about who you are.
Key Takeaways
- Nine models watermark their text, and three of them are on the free plan. Anthropic's table marks Fable 5.1, Fable 5, Mythos 5.1, Opus 5, Opus 5.5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5. Sonnet 4.6, Sonnet 4.5, Haiku 4.5, Mythos 5 and Opus 4.7 and below are not marked yet.
- Sonnet 5 now marks its text, and that is the change that reaches free users. It was ticked for files only on September 28, 2026. The table's revision stamp now reads October 8, 2026, and its text columns are ticked. Anything you wrote with free Claude on Sonnet 5 since then is marked.
- Files are a different story: every model on the list tags them. All 16 models in Anthropic's table carry Content Credentials (C2PA) on generated files, Haiku 4.5 and Sonnet 4.5 included. If Claude made you an image, it is tagged whichever model you were on.
- The Amazon Bedrock deadline has now passed and Anthropic has still said nothing about meeting it. That date is Anthropic's whole footnote on the Fable 5, Opus 4.8 and Sonnet 5 cloud cells: "Rollout on Amazon Bedrock will complete by October 17." Re-read on October 18, 2026, the day after: the footnote is unchanged, still on those same three cells, still written in the future tense, and the article is still stamped
2026-10-08T23:39:20Z. A deadline that passes without the vendor confirming it is not evidence either way, so treat those three cloud cells as promised rather than confirmed. On Anthropic's own apps all three are marked now, and that is the part that affects you. - Anthropic still publishes no end date for the older models. The sentence next to the footnote reads "Consistent with our commitments under the Code, Anthropic is adding watermarks to outputs from models released before August 2, 2026", with no deadline attached. An earlier version of this article carried "with all covered by December 2, 2026"; the word "December" has appeared nowhere on the page since, and did not on October 18, 2026. Seven of the sixteen models are still unmarked for text with no date for when that changes.
- Detection exists now, in private preview. Since September 1, 2026 Anthropic has run a detection API for "eligible organizations as required under EU law" -- regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups -- plus enterprises with their own compliance obligation.
- For files there is a free public checker. claude.com/check-files reads a file's content credential in your browser. Anthropic says "your file stays on your device" and "is never stored." It does not check text.
- It is invisible and adds nothing to the text. Anthropic says "nothing is added to the text and there are no hidden characters." A watermarked reply reads identically to an unwatermarked one.
- It cannot be traced to you. Anthropic states the watermark "carries no identifying information and can't be traced to a specific person, organization, or chat."
- It is free and does not slow Claude down. No extra tokens are produced, so Anthropic says the model costs the same to use and runs at effectively the same speed.
- It proves less than people assume. Anthropic is blunt: the watermark can only say Claude was "likely involved," and "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'"
- Short passages and code barely carry it. The mark lives in choices between equally good words, so facts, exact answers, and most code have little room for it.
- This is an industry-wide change, and OpenAI has now moved too. Anthropic says it signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside roughly 190 signatories, and that other major model developers will implement their own watermarks. On October 5, 2026 OpenAI published its own text watermark, textGrain -- but unlike Anthropic it is switching it on in the European Union only.
Which Claude Models Are Watermarked Today
Nine for text, and all sixteen for files. Anthropic's table has grown four times since it replaced its one-line "currently supported" sentence on September 16, 2026: a Claude Opus 5.5 row on September 22, a Claude Sonnet 5.5 row on September 28, and then three existing models plus Haiku 5.5 gaining text ticks by the stamp it carries now. The columns matter: marking text and marking files are separate commitments with separate coverage. Here is the table as Anthropic publishes it, read October 12, 2026:
| Model | Text watermark, Claude's own apps | Text watermark via AWS / Google Cloud / Microsoft Foundry | Content Credentials on files |
|---|---|---|---|
| Claude Fable 5.1 | Yes | Yes | Yes |
| Claude Fable 5 | Yes | Yes, Bedrock "by October 17" | Yes |
| Claude Mythos 5.1 | Yes | Yes | Yes |
| Claude Mythos 5 | No | No | Yes |
| Claude Opus 5.5 | Yes | Yes | Yes |
| Claude Opus 5 | Yes | Yes | Yes |
| Claude Opus 4.8 | Yes | Yes, Bedrock "by October 17" | Yes |
| Claude Opus 4.7 | No | No | Yes |
| Claude Opus 4.6 | No | No | Yes |
| Claude Opus 4.5 | No | No | Yes |
| Claude Sonnet 5.5 | Yes | Yes | Yes |
| Claude Sonnet 5 | Yes | Yes, Bedrock "by October 17" | Yes |
| Claude Sonnet 4.6 | No | No | Yes |
| Claude Sonnet 4.5 | No | No | Yes |
| Claude Haiku 5.5 | Yes | Yes | Yes |
| Claude Haiku 4.5 | No | No | Yes |
Two things about how to read that. A "No" above is an empty cell in Anthropic's table rather than a written denial, so it means "not listed as supported", not "Anthropic says never". And the table still carries exactly one footnote, attached to the three cloud cells marked above: "Rollout on Amazon Bedrock will complete by October 17." Anthropic gives no year with that date and we are not adding one. All sixteen rows were re-extracted cell by cell again on October 18, 2026, counting blanks as well as ticks: nine models ticked for text, seven blank (Mythos 5, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5, Haiku 4.5), and all sixteen ticked for Content Credentials on files. Nine plus seven is sixteen, which is how we know the extraction read every cell rather than skipping one. Nothing moved. The footnote's asterisk sits on exactly three cloud cells, Fable 5, Opus 4.8 and Sonnet 5, and on no others.
The practical reading for an ordinary Claude user changed again, and this time in the direction free users will care about. If you are on Pro, Max, Team or Enterprise and you type into Claude, you are often talking to Opus 5 or Opus 5.5, and both mark what they write. On the free plan, both Sonnets now mark text. Anthropic says anyone can use Sonnet 5.5 on Claude.ai, and Sonnet 5 gained its text ticks in the revision stamped October 8, 2026, so it no longer matters which of the two your model menu is sitting on: a free user writing with either is writing marked text. Haiku 5.5, which free users can also select, is marked from launch. And if you asked Claude for an image or a file on any plan, that file is tagged today.
Where the Marking Applies
Everywhere Claude runs, once a model supports it, because the watermark is applied by the model rather than by the app. Anthropic's help centre lists the surfaces: "Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag," and adds that when supported models are reached "through AWS, Google Cloud, or Microsoft Foundry they will carry watermarks."
That has one consequence worth spelling out. Anthropic says watermarking "will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from," and that because the mark is part of the text, "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing." There is no per-app setting and no per-region setting: marking applies "wherever Claude is offered, worldwide."
What Is a Text Watermark, in Plain Terms?
It is a pattern hidden in Claude's word choices, not a mark stamped on the page. Anthropic's own explanation is unusually readable: a model writes one word at a time, and at each step several words would do the job equally well. In "the weather today was cold and...", both "overcast" and "grey" are fine, and normally a random number settles it. Watermarking replaces that randomness with a pattern generated from a secret key, so the sequence of words is still random-looking but is checkable afterwards by whoever holds the key.
Anthropic offers a board-game analogy that lands better than most technical writing: imagine a game of Monopoly where, instead of rolling dice, players read their moves off the digits of pi starting at a random point. The moves are still effectively random and the game plays the same, but afterwards, if you knew the digits, you could work out that this game used pi. The method itself is a version of SynthID-Text, which Anthropic says Google DeepMind published in a Nature paper in 2024, and which belongs to a family of approaches going back to a 2022 proposal by the researcher Scott Aaronson.
Is Your Claude Text Watermarked? What Actually Carries the Mark
Less of your writing than you would guess. The watermark can only live where Claude had a genuine choice between words that are equally good, so anything with one correct answer leaves it nothing to work with. Here is what Anthropic says about each case:
| What you asked Claude to do | Is it watermarked? | What Anthropic says |
|---|---|---|
| Write something from scratch | Yes, most strongly | "The more Claude writes, the more decisions it has to make, and the more space there is for a watermark" |
| Translate a document | Yes | "A translation produced by Claude carries a watermark, because in this case every word is chosen by Claude" |
| Proofread your own writing | Barely, if at all | "Nearly all the words are the person's, there's very little (if anything) for the watermark to attach to" |
| Fix grammar and punctuation only | Probably not | The watermark "can only live in the handful of corrections, which might be too few to register" |
| Write factual or reference text | Sparsely | Where one word is the only right answer, "the watermark would have nothing to act on" |
| Write code | Mostly not | Code "has generally less watermarking"; comments inside code can carry it |
| A short reply | Too short to detect | Detection "doesn't work well on small samples"; confidence grows with length |
| Images and files (.png, .jpg, .svg) | No watermark, a label instead | A C2PA content credential in the file's metadata, "not embedded or hidden," checkable free at claude.com/check-files |
The pattern to take away: the watermark tracks how much of the writing was Claude's, not whether you used Claude at all. Ask it to polish two paragraphs you wrote and there is essentially nothing to find. Ask it for a 1,500-word draft and there is.
Can Anyone Actually Tell You Used Claude?
Some organisations can, since September 1, 2026, and the answer they get is a probability rather than a verdict. Anthropic now says it is "releasing a detection API in private preview." Checking text still requires Anthropic's key, so a general-purpose "is this Claude?" website is not what shipped -- this is an application-gated API with a named eligibility list.
| Who can get access today | Anthropic's wording |
|---|---|
| Regulators and law enforcement | "eligible organizations as required under EU law" |
| Media and fact-checkers | Named in the same list |
| Independent researchers | Named in the same list |
| Educational organisations | Named in the same list |
| EU civil society groups | Named in the same list |
| Enterprises with their own EU obligation | "enterprises who are similarly obligated to verify watermarking for their own compliance with the Act" |
| Everyone else | Not yet. Anthropic says it "plans to expand access to the detection API over time" and takes registrations of interest |
Files are the exception, and there the tool is free and public. Anthropic runs a Claude Content Checker at claude.com/check-files where you drop a file and it reads the content credential. Its own description of what happens: "Your file stays on your device," "the checker only reads the attached credential, not the file itself," and "your file is never stored or used for any other purpose." It only works on files. As Anthropic puts it on that page: "The tool does not check text."
When someone does run text through the detection API, Anthropic is careful about what it will and will not answer:
| Question someone might ask | Can the watermark answer it? |
|---|---|
| Was Claude likely involved in this text? | Yes, as a likelihood, on a long enough passage |
| Did a human write this? | No. It "doesn't confirm whether the text was human-written" |
| Was this written by a different AI? | No. Another AI would use a different key or a different method |
| Did Claude write it, or just edit it heavily? | No. It "cannot distinguish" the two |
| Who wrote it, or from which account? | No. The watermark holds no identifying information |
| Does this change who owns the text? | No. Anthropic says it "doesn't change a user's rights under our terms" |
Worth separating from all of this: the AI-detector services people already run essays through, such as Pangram, do not work this way at all. Anthropic notes they have no key and instead look for stylistic tells, and it gives two amusing examples of what those tools hunt for -- the "this isn't X, it's Y" construction, and the word "quietly" appearing far more often than it should. That approach guesses from style. The watermark checks a signature. Only the second one is what Anthropic is building.
Does It Cost You Anything or Change the Writing?
No on both counts, and Anthropic has published evidence rather than just an assurance. It says internal testing showed "no impact of watermarking on the content, level of creativity, or readability of Claude's text," and that a watermarked reply is indistinguishable from an unwatermarked one to a reader. It also cites the original SynthID-Text research, in which Google DeepMind served a watermarked model to a slice of real Gemini traffic and found no statistically significant difference in thumbs-up and thumbs-down ratings, plus a controlled study where human raters comparing answers side by side saw no quality difference.
On price and speed, the mechanism does the arguing: because the watermark changes which word is chosen rather than adding words, no extra tokens are generated. Anthropic states plainly that "watermarking has a negligible impact on the speed of models, and because it produces no extra tokens, the model is the same price to serve and use." Nothing on your bill or your plan changes, whether you are on free Claude with Sonnet 5 or a paid tier.
Why Anthropic Is Doing This
Regulation, and it says so directly. Anthropic writes that it is "implementing watermarking to comply with the EU AI Act," having signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 with around 190 total signatories. As of August 2, 2026, the EU requires AI providers serving its market to mark AI-generated content.
The detail most people will actually feel is geographic. Anthropic is switching this on for everyone, everywhere, not just for European users: "We're applying watermarking globally at launch because we don't yet have a durable way to scope it by region," adding that it will keep evaluating other approaches. So a rule written in Brussels is arriving in your chat window in Ohio or Osaka, and Anthropic expects company -- it says other major model developers that signed the same code "will be implementing their own watermarks."
OpenAI has now done it, and chose the other answer on geography. On October 5, 2026 it published its approach to EU text provenance rules, introducing a watermark of its own called textGrain for "eligible ChatGPT and Codex users across all plans in the EU only," adding that it is "not making text watermarking a global default at launch." That makes Anthropic, for now, the only one of the two marking text outside Europe. Our page on ChatGPT's text watermark has the detail, including why OpenAI's detector is not public.
Google is already ahead of it on the image side. Every image the Gemini app generates carries an invisible SynthID watermark plus a visible one, which is covered in our guide to Nano Banana and Gemini image generation. Text has been the gap, and this is the industry closing it.
Should You Care?
Barely, if you use Claude to help with your own writing. Editing, proofreading, brainstorming and short answers carry little or no watermark, and nothing about you is recorded. What changed most recently is the free plan: Sonnet 5 marks its text now, where a fortnight ago it did not, so "I only use free Claude" has stopped being a reason your text is unmarked.
Yes, if you submit long AI-drafted text somewhere it is not allowed. A full draft written by Claude is exactly the case the watermark is built to catch, Anthropic says light editing probably will not remove it, and detection is no longer hypothetical -- it is a shipped API with educational organisations on the eligibility list. AI-written coursework got meaningfully more checkable on September 1.
Yes, if you receive files and need to know where they came from. claude.com/check-files is free, needs no account, and does the check in your browser. That is a genuinely useful tool for anyone handling images they did not make.
Worth watching, if you care about privacy. The commitment that the watermark holds no identifying information and cannot be tied to a person, organisation or chat is a strong one, and it is the claim to keep an eye on now that the detection API is live rather than promised.
Do not change what you pay for because of this. No plan gets a watermark-free Claude, no price moves, and every major provider is heading the same way. If you were weighing a subscription for other reasons, our guide to which AI chatbot is worth paying for covers what actually differs between them.
The bottom line: Claude's text watermark costs you nothing and tells no one who you are, but it is no longer unverifiable and no longer rare. Nine of the sixteen models in Anthropic's table mark their text, and that now includes both Sonnets a free user can pick as well as Haiku 5.5; every listed model tags the files it makes. If your use of Claude is help rather than authorship, there is nothing here to worry about. If it is authorship, the era of nobody being able to tell has already ended, on the free plan as well as the paid ones.
Sources
- Anthropic, How Claude's text watermarking works, August 14, 2026, carrying an in-post note: "Updated Sep 1, 2026: Provided up to date information on the watermarking detection API."
- Anthropic, How Claude marks AI-generated content, help centre,
lastUpdatedDate2026-10-08T23:39:20Z, sixteen-row table re-extracted October 17, 2026 - Anthropic, Check if files were made with Claude, read September 8, 2026
- Anthropic, claude.com/pricing plan comparison matrix, decoded September 8, 2026
Corrections and Updates
October 17, 2026: the Amazon Bedrock deadline this page quotes arrived, and Anthropic let it pass without a word. Nothing here was wrong; this line records the check. The footnote "Rollout on Amazon Bedrock will complete by October 17" is still on 16266773, still attached to the Fable 5, Opus 4.8 and Sonnet 5 cloud cells, and the article's revision stamp is still 2026-10-08T23:39:20Z -- unchanged in the nine days since the last correction, and not touched on the deadline day. All sixteen rows were re-extracted cell by cell rather than read as prose, because a prose read passed on September 28 while four models' cells had moved: nine ticked for text, seven blank, sixteen ticked for files, identical to the October 12 read. So the page now says those three cloud cells are promised rather than confirmed. We are not writing that the rollout slipped, and not writing that it completed -- Anthropic has published neither, and an unrevised page is not a vendor statement in either direction. Separately re-checked in the same read: "December" is still zero occurrences, so the deleted "all covered by December 2, 2026" clause has not returned and no end date for the pre-August-2 models exists on this surface.
October 12, 2026: this page said five Claude models watermark their text, that Anthropic's table listed 15 models, and that Sonnet 5 "marks files only." All three are now wrong. The marking article's revision stamp reads 2026-10-08T23:39:20Z and its table has sixteen rows, nine of them ticked for text: Fable 5, Opus 4.8 and Sonnet 5 gained text ticks, and Haiku 5.5 was added marked from launch. Sonnet 5 is a free-plan model, so the page's advice that a free user on Sonnet 5 was unmarked was the costly half. Summary, seoTitle, seoDescription, four Key Takeaways, two FAQ answers, the model table, the free-plan paragraph, the "Should You Care?" opener and the closing verdict were corrected in this edit.
October 12, 2026: this page quoted Anthropic committing to cover the older models "with all covered by December 2, 2026," and that sentence is no longer on the page it was quoted from. Read today, 16266773 contains the string "December" zero times in a complete 7,617-character body at HTTP 200, and now says only that Anthropic "is adding watermarks to outputs from models released before August 2, 2026" with no end date. The announcement, anthropic.com/news/claude-text-watermark, also returns zero hits for "December" at 156,395 bytes. We are recording that the quoted clause is gone and not publishing the reverse claim: no Anthropic surface we read states that the deadline has been dropped or changed, and the only date the table now carries is its Amazon Bedrock footnote, "by October 17." The deadline framing was removed from the summary, seoTitle, seoDescription, one Key Takeaway, two FAQ answers and the closing verdict.
September 28, 2026: this page said that if you are on Free "you are talking to Sonnet 5, which does not mark text yet." Anthropic added a fifteenth row, Claude Sonnet 5.5, on September 28, 2026, ticked in all three columns -- and it says anyone can use Sonnet 5.5 on Claude.ai, so a free user who selects it does write watermarked text. Five models now mark text. Summary, Key Takeaways, two FAQ answers, the model table and the free-plan paragraph corrected.
September 22, 2026: this page said three Claude models watermark their text and that Anthropic's table listed 13 models. Anthropic added a fourteenth row, Claude Opus 5.5, on September 22, 2026, ticked for text watermarks on its own surfaces and on cloud partners and for Content Credentials on files. Four models now mark text.
September 17, 2026: this page said Opus 5 was not watermarked and that "the everyday Claude you use is probably not watermarked yet." Both are now wrong. Anthropic revised its marking article on September 16, 2026 and replaced its two-model sentence with a table that marks Opus 5 for text watermarks, adds cloud-partner coverage for it "starting September 14, 2026," gives Content Credentials on generated files to all 13 listed models including Sonnet 5 and Haiku 4.5, and sets a deadline of December 2, 2026 for every model released before August 2, 2026. The summary, Key Takeaways, two FAQ answers, the model table and the closing verdict were corrected in this edit.
September 8, 2026: this page said nobody could check a Claude watermark. That stopped being true a week ago. It read "Nobody can check it today" and "Anthropic has not yet released the tool," citing Anthropic's original wording that it "will soon be offering a watermark detection API." Anthropic amended that post in place on September 1, 2026 -- the note at its foot reads "Provided up to date information on the watermarking detection API" -- and it now says: "We are releasing a detection API in private preview." Access is gated to regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, EU civil society groups and enterprises with their own EU compliance obligation. The summary, Key Takeaways, two FAQ answers, the detection section and the closing verdict were all corrected in this edit.
September 8, 2026: this page did not say which models are watermarked. It repeated Anthropic's launch phrasing, "future Claude models," which is not an answer for a reader asking whether their own chat is marked. Anthropic's help centre now names them: "Models currently supported include Fable 5.1 and Mythos 5.1." A new model table was added, and the older-models FAQ now points out that Sonnet 5 (June 30, 2026) and Opus 5 (July 24, 2026) both launched before Anthropic's August 2, 2026 cutoff and so sit in the transition group.
September 8, 2026: a free public checker for files exists and this page missed it. The page said only that Anthropic would be "providing our own" tool. It has: claude.com/check-files, which reads a file's content credential in the browser without uploading it.
Read next
Keep up without the jargonChatGPT Text Watermark Explained: It Is Live in the EU, and Nobody Can Detect It
ChatGPT text is now watermarked in the European Union, on every plan including free. OpenAI's help centre states it in the present tense. Outside the EU nothing changes. There is still no public detector: access is limited to approved researchers, and the mark never identifies you.
Claude Sonnet 5.5 Explained: Free Claude Gets the New Model Too
Anthropic released Claude Sonnet 5.5 on September 28, 2026, and its own model page says anyone can chat with it on Claude.ai. That includes the free plan, which is unusual: Opus 5.5 and Fable 5.1 are paid-only. It knows five months more than Sonnet 5, and you pick it in the model menu.
Claude Sonnet 5 Explained: What's New and What It Means for You
Claude Sonnet 5 has been the free plan's model since June 30, 2026, and it still is. But Anthropic released Sonnet 5.5 on September 28, 2026 and says anyone can use it on Claude.ai, free included. It has not said which of the two is now the default, so check your model menu.
Nano Banana Explained: Is Gemini's Image Generator Free?
Nano Banana 2 is the image generator inside the Gemini app, and it is free: Google lists image generation on every plan, including no plan at all. What money buys is the Pro redo, 2K downloads and video. You must be 13 to generate images and 18 to edit them.
Which AI Chatbot Should You Pay For in 2026?
For most people, ChatGPT Plus at $20/month is the best-value paid chatbot. Choose Google AI Pro ($19.99) if you live in Google apps, or Claude Pro ($20, $17 on annual) if you mostly write and work with documents. Light users stay free, and college students should take Google's free year first.
Frequently Asked Questions
Does Claude watermark everything it writes?
Not yet, but most of what you would actually use. Nine of the sixteen models in Anthropic's table mark text: Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5. Sonnet 4.x, Haiku 4.5, Mythos 5 and Opus 4.7 and below are not marked yet.
Can my teacher or my boss tell I used Claude?
It is now possible in principle, which it was not in August. Anthropic's detection API is in private preview, and its eligible list names educational organisations. You have to apply, access is granted for EU compliance reasons, and the result is a likelihood on a long passage.
Does the Claude watermark identify me personally?
No. Anthropic states the watermark carries no identifying information and cannot be traced to a specific person, organisation or chat. Nothing in the watermark or its key would let anyone recover who typed the prompt or what else was said in that conversation.
Does watermarking make Claude's writing worse or slower?
Anthropic says no on both counts. It reports no impact on content, creativity or readability in internal testing, and because the watermark adds no extra words, it produces no extra tokens, so it does not slow responses down or cost more to use.
Can I remove the watermark by editing the text?
Partly. Anthropic says light editing probably will not remove it completely, but a full rewrite that replaces every word will. It adds the obvious caveat: at that point it is arguable whether the text is still AI-generated at all.
What about images and files Claude makes?
Every model on Anthropic's list now tags them, and you can check them yourself for free. Supported files such as .png and .jpg carry a C2PA content credential in the metadata. Anthropic's Claude Content Checker at claude.com/check-files reads it in your browser without uploading the file.
Do older Claude models watermark their text?
Yes, and more of them than in September. Sonnet 5, Fable 5 and Opus 4.8 now mark their text alongside Opus 5, Opus 5.5, Sonnet 5.5, Fable 5.1, Mythos 5.1 and Haiku 5.5. Sonnet 4.6, Sonnet 4.5, Haiku 4.5, Mythos 5 and Opus 4.7 and below still mark files only.