A pair of viral posts from M1Astra and Polymarket put a startling claim into millions of feeds: Claude will now hide a watermark inside AI-generated text, and that mark will survive copy and paste.
The core Claude claim is real. The broader claim that Claude and GPT are both about to watermark every answer is not. Anthropic has published a concrete marking plan for supported Claude models. OpenAI currently confirms watermarks and signed provenance for generated images, not an equivalent production watermark in ordinary ChatGPT or GPT API text.
The accurate headline is: Claude has confirmed text watermarks. ChatGPT has confirmed image watermarks. GPT text is still an open question.
What Anthropic actually announced
In its official guide, How Claude marks AI-generated content, Anthropic says Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking from launch. Generated text from supported models will carry an embedded watermark. Supported generated files can carry signed provenance metadata.
The scope is unusually broad:
- Products: Claude, Claude Platform/API, Claude Code, Claude Cowork and Claude Tag.
- Cloud partners: supported models accessed through AWS, Google Cloud or Microsoft Foundry.
- Regions: worldwide wherever the supported Claude model is offered, not only inside the EU.
- Older models: still in transition; Anthropic says it is working to add marking support.
That final point matters. “Claude watermarks every response now” is too broad. The official wording is about supported models, with newer launches covered first and existing models still being updated. Anthropic also says third-party detection will be supported, but the technical detector and detailed methodology have not yet been published.
What about ChatGPT and GPT text?
OpenAI has already shipped a strong provenance system for images. Images made through ChatGPT, Codex and the OpenAI API include C2PA Content Credentials and SynthID watermarks, according to OpenAI’s current help documentation. OpenAI also provides a verification tool for supported media.
But as of August 12, 2026, OpenAI has not publicly said that normal ChatGPT answers or GPT API text contain an active text watermark. Its current position is more careful: it wants to extend provenance signals to other modalities, including text, as standards and tools mature.
This is not because OpenAI ignored the idea. In 2024 it said it had developed a highly accurate text-watermarking method, while also documenting weaknesses: translation, extensive rewording and model-to-model rewriting could break the signal; false positives become material at internet scale; and the technology could stigmatize non-native English speakers who use AI as a writing aid. That research remains important context, but it is not a 2026 product announcement.
How an invisible text watermark works
Anthropic has not disclosed Claude’s exact algorithm, so claims that it inserts zero-width characters, personal IDs or a hidden prompt history are speculation. The best public illustration of the general technique comes from Google DeepMind’s SynthID for text.
A language model writes by choosing one token after another from a probability distribution. A statistical watermark slightly changes the scoring of acceptable next tokens. The result still reads normally, but a detector that knows the secret pattern can test whether the sequence contains more favoured choices than chance would predict.
- The model identifies several plausible next tokens.
- The watermarking rule quietly favours a subset of them.
- That tiny bias accumulates across a longer passage.
- A compatible detector measures the pattern later.
This is why length matters. Long, open-ended prose gives the model many choices and the signal room to accumulate. Short factual answers, quotations, code and other highly constrained output offer fewer safe choices. Google also notes that heavy rewriting or translation can sharply reduce detection confidence.
Watermark, metadata and visible labels are not the same thing
1. A text watermark lives in the generated sequence
It can survive copy and paste because the pattern is carried by the words or tokens themselves. It may survive light editing, but it is not invulnerable.
2. C2PA is signed provenance attached to a file
Anthropic says supported SVG, PNG and JPG files will receive signed provenance metadata using the C2PA standard. A valid credential can describe which service processed the file and whether its declared history was altered. It does not certify that the picture is true, lawful or accurately captioned. Metadata can also disappear during screenshots, conversion or re-saving in unsupported software.
3. A visible label speaks to a person
A badge, icon or disclosure such as “AI-generated” is designed for a reader, not a machine. The EU now provides optional icons for this purpose. A file may contain a machine-readable mark without displaying a visible badge, and a publisher may display a label even when no durable watermark is available.
What the EU AI Act requires
The immediate trigger is Article 50 of the EU AI Act, whose transparency obligations became applicable on August 2, 2026. Providers of generative systems must make synthetic audio, image, video and text outputs machine-readable and detectable as artificially generated or manipulated, as far as technically feasible.
The EU Code of Practice on Transparency of AI-generated Content gives providers and deployers a recognised compliance framework. Signing the code is voluntary. The underlying Article 50 obligations are legal requirements.
The law also separates provider marking from publisher disclosure. Providers create machine-readable signals. People and organisations publishing deepfakes or certain AI-generated public-interest text may have visible disclosure duties, with important exceptions around human review and editorial responsibility. A hidden watermark alone is not the whole compliance story.
What the online reactions get right — and wrong
The Polymarket post had roughly 15 million displayed views when we checked it. The replies and wider discussions repeatedly raised the same concerns. They are not evidence on their own, but several deserve direct answers.
“Will this degrade Claude’s writing or code?”
Anthropic says the watermark does not change meaning, quality or readability. It has not yet published the benchmarks or technical design needed for independent scrutiny. The honest position is therefore neither “quality is definitely harmed” nor “there is zero trade-off”: it is an unverified product claim awaiting evidence, with constrained formats such as code likely to be an important test case.
“Can I ask Claude to remove it?”
Anthropic describes the mark as model-level and lists no user setting to disable it. A prompt asking Claude not to watermark the response should not be assumed to override infrastructure applied by the model provider.
“Does the watermark identify me?”
Anthropic’s public documentation describes a provenance signal, not an account tracker. It does not say the text mark contains a user ID, prompt, conversation history or ownership claim. Until the technical specification appears, claims of personal traceability are unsupported.
“Could schools and employers treat it as proof of cheating?”
They should not. Anthropic explicitly says a detected mark is not fully conclusive about provenance, and that Claude may only have proofread, translated, summarized or converted human work. A process that turns “processed by Claude” into “written fraudulently by Claude” would manufacture certainty the signal does not contain.
What a detected mark can — and cannot — prove
A detected Claude mark can support one narrow statement:
This content may have been processed by a supported Claude model.
It does not establish:
- who wrote the original draft or ideas;
- whether Claude generated, translated, proofread or merely reformatted it;
- whether the content was changed after leaving Claude;
- whether using AI broke a school, employer or publisher policy;
- whether the claims in the content are accurate.
The reverse is just as important. No detected mark does not mean “human-written.” The model may predate support; the passage may be too short; the text may have been translated, heavily edited or mixed with other writing; or the content may come from a different AI provider.
What this changes for AI detection
Watermark detection and statistical AI detection answer different questions. A watermark verifier asks whether a known provider embedded a recognised signal. A general detector estimates whether the finished text resembles generated writing, including content from providers that expose no provenance signal.
The strongest future report will combine three layers: official provenance when a verifiable mark exists, content-level signals when it does not, and human process evidence for the final decision. We will not claim IA Checker detects Claude’s watermark until Anthropic publishes a usable detection mechanism and we can validate it independently.
Until then, the practical rules do not change: use a score for triage, inspect the passages behind it, and never treat one percentage as proof. For the mechanics, read how AI detection actually works; for the risks, see why human writing gets flagged.
Can IA Checker detect or remove a hidden text watermark?
AI Detector: useful now, provenance-ready later
The IA Checker AI Detector analyzes the authenticity of finished text and documents through sentence-level patterns associated with generated writing. It works even when a provider exposes no watermark, when editing has weakened the mark, or when the content comes from another model. This content-level analysis is complementary to a future provider-specific Claude watermark check.
Once Anthropic releases a documented verification method or trusted detector access, IA Checker can add a separate Claude provenance result alongside the existing authenticity score. One reading will test a provider-specific mark; the other will continue to analyze writing patterns across sources.
AI Watermark Remover: rewrite the text and disrupt its statistical watermark
The IA Checker AI Watermark Remover restructures text or an uploaded document, varies its rhythm and replaces mechanical phrasing while preserving meaning, facts, names and quotations. Statistical text watermarks live in token choices and their distribution. A substantial rewrite changes that structure, which is why Anthropic acknowledges that heavy paraphrasing, translation or mixing can weaken the mark.
In practical terms, import or paste the content, perform a deep rewrite, then analyze the result again. The Watermark Remover can disrupt the hidden statistical pattern today. A provider-verified certificate will become possible only when Anthropic publishes a detector that independent services can validate against. Our step-by-step guide explains how to remove AI watermarks from text and documents without confusing text patterns with image or file metadata.
Frequently asked questions
Does Claude watermark every response now?
Not every historical response. Anthropic says models launched on or after August 2, 2026 support marking at launch, while support for older models is still being added. The exact model-by-model status and public detector documentation are not yet complete.
Does ChatGPT watermark its text?
No production text watermark has been publicly confirmed for normal ChatGPT or GPT API answers as of August 12, 2026. OpenAI does apply C2PA and SynthID to generated images and has researched text watermarking.
Will copy and paste remove Claude’s watermark?
Anthropic says the mark travels with copied text and may survive some editing. Heavy paraphrasing, translation, mixing or very short excerpts can make the signal disappear or become unreliable.
Can a watermark prove plagiarism or cheating?
No. It may indicate processing by a particular supported model. It cannot reconstruct the writing process, interpret a policy or prove misconduct by itself.
Can IA Checker remove Claude’s text watermark?
IA Checker can deeply rewrite text and documents, changing the token distribution, rhythm and phrasing that carry a statistical watermark. Anthropic says substantial rewriting may weaken its mark. IA Checker cannot yet issue a provider-verified removal certificate because a public, validated Claude watermark detector does not exist.
Sources
- Anthropic — How Claude marks AI-generated content
- OpenAI — Advancing content provenance for a safer, more transparent AI ecosystem
- OpenAI Help — C2PA and SynthID in OpenAI-generated images
- OpenAI — Understanding the source of what we see and hear online
- European Commission — Code of Practice on Transparency of AI-generated Content
- EUR-Lex — Regulation (EU) 2024/1689, Article 50
- Google DeepMind — SynthID
- C2PA — technical specifications
Rewrite the sequence. Keep the meaning.
IA Checker restructures text and documents while protecting the facts, names and quotations that matter.
Open Watermark Remover
