The authorship problem
Claude may mark text it only proofread, translated or improved. The mark can show that Claude processed the words. It cannot show who wrote them.
The short answer
Claude’s text watermark is an invisible, machine-readable mark added as a supported model generates text. It is carried by the wording, so it can survive copy and paste. Anthropic has not disclosed exactly how its version works.
Anthropic has confirmed a model-level, imperceptible text watermark. It has not published Claude's algorithm, secret key design, thresholds, passage-length requirement or measured error rates. The mechanics below describe the published family of token watermarks, including Google's open SynthID Text reference. Claude may differ.
The mark is added during generation, embedded directly in the text, survives exact copying and may survive some editing. Detection is intended for users and third parties.
The scoring function, context length, key structure, detector thresholds, confidence labels, false-positive rate, false-negative rate and any product-specific settings.
For each next token, the model produces a logit. A logit is the raw score behind the probability of choosing that token. Temperature and top-k or top-p sampling then shape the candidate set.
p(tokenᵢ) = softmax(logitᵢ / temperature)The watermark is applied inside this generation step. It is not pasted onto the finished paragraph afterwards.
A keyed pseudo-random function can take the recent token sequence and a private key, then produce a repeatable score for each possible next token. The same context and key recreate the same hidden preference.
scoreᵢ = PRF(secret key, recent tokens, tokenᵢ)Older papers often explain this as green and red token lists. SynthID uses a more layered pseudo-random scoring function called a g-function.
The system slightly changes the logits of otherwise plausible tokens. A strong candidate can still win even when the watermark would prefer another one. That protects meaning and fluency.
adjusted logitᵢ = logitᵢ + δ × scoreᵢThe strength δ creates a trade-off. A larger nudge is easier to detect but has more chance of affecting output. A smaller nudge is less intrusive but needs more text.
The detector tokenises the supplied text, recreates the keyed scores and asks whether the observed choices align with the watermarked distribution more often than chance predicts.
evidence = Σ expected keyed score over eligible tokensA detector can compare that evidence with distributions for marked and unmarked text. Google's reference detector returns watermarked, not watermarked or uncertain using configurable thresholds.
One token proves nothing. Thousands of ordinary human passages will contain accidental local matches. Confidence grows only when the detector sees enough eligible, varied token choices.
Short strings, fixed facts, quotations and code offer fewer free choices. This is why a long article is a better detection target than an ad headline, and why no honest guide can promise a universal word count without Claude's thresholds.
Changing a token can change the context used to score later tokens. Mild edits may leave enough matching regions. A thorough rewrite, translation or mixing text from several sources can reduce the evidence.
This is not the same as deleting metadata. The wording itself is the carrier. It is also why the detector can produce false negatives after transformation and why a result needs an uncertain state.
tokenlogitPRFn-gramthresholdfalse positiveThe important limit: a Claude mark may show that Claude processed the words. It cannot tell whether Claude wrote the first draft, fixed the grammar or translated human work.
Can anyone check it, why did it arrive now, and which other companies mark AI content?
No, not yet. Anthropic says it will give users and third parties a way to check, but the tool and technical details are not public.
The EU AI Act's transparency rules started applying on 2 August 2026. AI providers must make covered AI-made content detectable where technically feasible. Anthropic has applied its marking worldwide, not only in Europe.
Google marks supported text, images, audio and video with SynthID. OpenAI marks supported images, audio and video, but we found no public ChatGPT text mark. Grok visibly marks generated images and video. Meta, Microsoft and Adobe also mark some AI media.
Anthropic says the mark can show that a supported Claude model processed the words. The risk is how a university, employer, client or publisher interprets that result.
A university student writes an essay, then asks Claude to correct the spelling and grammar. Anthropic says the returned text may still carry a mark.
The detector cannot show whether Claude devised the argument, rewrote whole sections or changed a handful of sentences. Yet a university, employer, client or publisher may still read “processed by Claude” as “written by AI.”
Text watermarking changes the probabilities used to select tokens. Google tested SynthID Text across almost 20 million Gemini responses and found no significant drop in human quality ratings. Its chosen configuration did reduce some response-to-response diversity.
Google’s result cannot be transferred to Claude. Anthropic has not published its method, configuration or enough independent testing to evaluate the claim that quality is unaffected.
Anthropic says extensive rewriting, translation and mixing can weaken the mark. Short passages may not contain enough evidence. Other models and unmarked open systems remain available.
Inference: the person who uses Claude openly to polish a long document may retain more of the mark than someone deliberately trying to defeat detection.
Gemini in Google Docs can draft, improve and proofread. Copilot in Word can rewrite selected text. That does not mean either product carries Claude’s watermark. It shows that AI editing is becoming a normal part of writing in a document, not always a separate decision to let AI write the work.
If a detector cannot distinguish a grammar check from a generated argument, the result should not be treated as a verdict. Otherwise, the student or employee may be forced to prove a negative about work they actually wrote.
No public evidence shows that Claude’s mark contains an account ID, identity, chat history or instructions for another agent. “Model-level” means the mark is added while a supported model generates text, so it works across Claude products and supported partner platforms.
A separate NeurIPS paper shows that AI agents can use steganography, meaning hidden communication inside ordinary-looking outputs, in specially designed multi-agent settings. That establishes a general technical risk. It does not show that Claude’s watermark is being used to send hidden messages.
Code remains an open test case. Google’s research says detection depends partly on how many plausible next-token choices are available. Code can offer fewer valid choices than prose. Different trade-offs may therefore apply, but Anthropic has not published Claude-specific code results.
Why now
Article 50 of the EU AI Act started applying on 2 August 2026. That is the date behind Anthropic's announcement. Generative AI systems already on the EU market have a limited grace period until 2 December 2026, but only for the Article 50(2) marking and detection duty.
Providers of systems that generate synthetic text, images, audio or video must make covered output machine-readable and detectable where technically feasible.
In the EU, professional users must clearly disclose deepfake media. Public-interest AI text also needs disclosure unless it has meaningful human review and someone takes editorial responsibility.
Anthropic signed the EU code and says its model-level marking applies globally. It did not make a Europe-only version of Claude.
The EU minimum is narrower than many readers assume. Article 50 excludes standard editing. It also excludes changes that do not substantially alter the input or its meaning. Anthropic says a Claude mark may appear after proofreading or translation. Its public implementation therefore appears broader than that minimum, although the company has not published enough technical detail to map the boundary exactly.
For planning, assume every Claude model you can continue using will eventually mark new output.
Models launched from 2 August 2026 support marking from launch. Anthropic says it is adding support to older releases. It has not promised that every retired model will be updated or published a live model-by-model list.
Do not assume an older model will stay unmarked. Opus 4.8, Fable 5 and Opus 5 all launched before the cut-off. Opus 5 arrived on 24 July 2026. They are in the transition group. A later update can mark new answers. It cannot alter text you already generated and saved unless that text is processed again.
The mark is added by the supported model while it generates the answer. It is not added by the chat window.
Using the API does not avoid the mark. If the chosen Claude model supports marking, the text it returns can contain it. Your prompt is not being changed. The generated answer is the part being marked.
If the app uses a supported Claude model and passes its answer through, the mark can travel with the words. If the app rewrites, translates, shortens or routes the job to another model, the result is less certain.
Anthropic says model-level marking also applies when supported Claude models are used through AWS, Google Cloud and Microsoft Foundry.
Ask which exact model produced the final words and whether the app changes the answer before you receive it. “Powered by AI” is not enough information.
This is a practical risk guide, not a Claude detector. The exact detector and its error rates have not been published.
Step 1
What are you making?
Step 2
What did Claude do?
More media products already carry origin marks. But “watermarked” can mean a hidden mark inside the media, signed information attached to the file, or a visible logo.
If a supported Google or OpenAI tool generated the image, video or audio, assume it carries at least one type of origin marking. Do not apply that blanket rule to every Claude or Grok output.
| Content | Claude | Google / Gemini | OpenAI | Grok |
|---|---|---|---|---|
| Text | Hidden markImperceptible text markSupported Claude models. Anthropic has not published the exact method. New models from 2 August 2026; older models are being added. | Hidden markStatistical word patternSynthID is used for supported Gemini app and web text. | Not publicly confirmedNo public ChatGPT text mark foundWe found no public production ChatGPT text watermark. | Not publicly confirmedNo public Grok text mark foundWe found no public Grok text watermark. |
| Images | Attached recordSigned file informationSupported PNG, JPG and SVG files can carry C2PA information. No hidden pixel mark is claimed. | Hidden markSynthID in the pixelsSupported generated images carry SynthID in the pixels. | Hidden + attachedTwo layersChatGPT, Codex and API images use hidden SynthID plus signed C2PA information. | Shown markVisible Grok watermarkGrok Imagine images show a Grok watermark. No hidden method is described. |
| Video | Depends on toolClaude plus video toolsClaude can create some animations itself through code and interactive visuals. For fuller videos it can assemble, render or edit through providers such as HyperFrames and Adobe. The final mark depends on the tool that produces the file. | Hidden markSynthID across framesVeo uses SynthID across generated video frames. | Shown + hiddenSora API, while it remains availableSora web and app ended on 26 April 2026. The API remains until 24 September 2026. Sora outputs use visible and invisible origin marks plus C2PA information. | Shown markVisible Grok watermarkGrok Imagine video shows a Grok watermark. No hidden method is described. |
| Voice | Not publicly confirmedClaude Voice Mode existsClaude can speak responses on the web, Claude Desktop and mobile. Anthropic has not publicly confirmed a watermark for that audio. | Hidden markInaudible SynthIDSupported audio, including Lyria and NotebookLM audio, uses SynthID. | Hidden markInaudible SynthIDSupported ChatGPT and API audio uses SynthID and can be checked with OpenAI's verifier. | Not publicly confirmedNo public Grok voice mark foundGrok has voice products, but xAI's public voice docs do not name a watermark. |
China has its own national labelling system. It is separate from Claude’s approach and is not one company watermark. Covered services add visible labels and hidden file records. Publishing platforms also check for them.
For hosted services such as DeepSeek, Kimi, Doubao, Qwen or Hunyuan, assume China’s rules may apply. That does not prove the downloadable model weights contain a built-in watermark.
A visible label can appear in the text or interface. Exported documents may carry hidden metadata. A Claude-style word pattern is not generally confirmed.
Visible labels and hidden file records can apply. A digital watermark may also be used.
A visible label can appear at the opening or in the player. Hidden file records and digital watermarks may also apply.
A spoken, sound or screen label can apply. Hidden file records and inaudible watermarks may also be used.
The limits matter. These systems cover supported products and outputs. Older files, third-party integrations and heavily transformed media may behave differently. “Some tools mark some outputs” is not the same as “all AI media is marked.”
The watermark question and the disclosure question are not the same. A hidden mark can exist even when no rule tells you to add a public label.
The practical rule: if a client, employer, contract or platform asks, disclose. If AI makes a real person or event look or sound real, disclose. Ordinary drafting, outlining or proofreading is usually a different case.
We found no standard “AI helped write this” button for ordinary text in Gmail or Outlook. Routine AI drafting does not create a general email-platform disclosure step.
If your workplace, client or profession requires disclosure, add it in plain words in the email or signature.
Claude helping with ad copy is not, by itself, the main trigger. The creative assets matter. Google Ads now has an AI label setting, with visible labels for some ads in the EU, India and New York. Meta labels ads it detects as created or significantly edited with AI.
If the platform asks whether an image, voice or video was AI-made, answer yes. Do not present synthetic people or events as real.
TikTok, YouTube and Meta have disclosure tools or labels for realistic AI media. Use them when the platform asks. YouTube explicitly says AI help with a script, title or outline does not by itself require its altered-content disclosure.
Professional users face clearer duties for deepfake image, audio and video. AI text published to inform the public on matters of public interest also needs a label unless there is meaningful human review and someone accepts editorial responsibility.
If you need a sentence: “Created with AI assistance and reviewed by [name or company].” It is clearer than claiming the whole item was “made by AI” when Claude only edited or helped plan it.
Workarounds
Probably. Not one permanent fix.
Text rewriting tools already exist. Researchers have also removed or overwritten image and audio watermarks. Once Claude offers a public checker, people will be able to test rewrites against it. That will speed up removal work.
It does not guarantee a clean one-click result. Detectors change. Marks can be layered. A “fix” may damage the words or media, add another model's mark, or only fool one version of one detector.
Text: a deep rewrite or translation can weaken statistical marks, but Claude's exact limits are still private. Images: file records are easy to strip, while hidden pixel marks may need the image to be edited or remade. Research shows that even marks designed to survive editing can be removed. Audio: researchers have overwritten or removed hidden audio marks. Video: visible marks can be removed frame by frame. Hidden marks may survive ordinary file conversion and are harder to predict.
Prediction, not fact: a Claude-specific removal market is likely after public detection arrives. It is not available as a verified, universal solution today.
What has not changed
A model update can mark new answers. It cannot reach back and alter files you already saved.
A headline or brief email may not contain enough evidence for a reliable result.
Anthropic has not said the mark contains your account ID, identity or chat history.
Original drafts, sources, edit history and approval notes tell a fuller story than one detector result.
The short answer is that no published platform rule we found treats Claude's text mark itself as a reason to reduce reach or delivery. That could change.
Google says it rewards useful, original, accurate work, however it was made. Mass publishing with little added value can breach its spam rules.
Gmail's published guidance focuses on sender checks, domain reputation, spam complaints and unsubscribe rules.
TikTok, YouTube and Meta are building labels for AI media. Platforms can change their systems without publishing every factor they use.
Google Ads reviews the claim, the ad, the account and the destination. Misleading claims and attempts to evade review still matter.
Read this carefully: “No known penalty” is not the same as “a penalty is impossible.” It means the current public rules do not name the Claude text mark as a ranking, reach, delivery or ad-auction factor.
You do not need to hide normal AI use. You do need to choose whether speed or clear human authorship matters most for this piece.
Low stakes or high speed
Client work or your public voice
These came up again and again in the supplied Reddit discussions. The posts helped find the questions. Primary sources control the answers.
No. The mark is described as part of the generated wording. Keeping the same words keeps the same pattern.
It may add a Claude mark to the returned text. Anthropic also says this does not prove Claude was the original author. That distinction must be part of any fair detection result.
A substantial rewrite or translation may weaken or destroy a text watermark. That is a known limit of this kind of system. It is not a reliable promise for Claude because its exact method and thresholds are not public. Another model may also add its own mark.
Possibly, but short text usually provides less statistical evidence. Do not treat “too short” as a safe fixed rule. Anthropic has not published a minimum length.
We found confirmed OpenAI marking for supported images and audio, including C2PA and SynthID. We did not find a public production ChatGPT text watermark. Google has used SynthID for supported Gemini text since 2024.
Anthropic says it is working on detection access for users and third parties. The public tool, access rules and technical documentation are not yet available. Today, broad claims that a school, client or platform can already run the official Claude text check are premature.
Keep the original author, what Claude processed, the main edits, and who approved the final work. That is more useful than a single “AI or human” label.
Company statements describe their own systems. They do not prove those systems work perfectly. Community posts are treated as reaction, not evidence.