--- title: How Claude marks AI-generated content | daily.dev description: Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content and outlines how Claude will mark AI-generated... canonical: https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How Claude marks AI-generated content | daily.dev og:description: Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content and outlines how Claude will mark AI-generated... og:url: https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh og:image: https://api.daily.dev/og/posts/DCnHRgkJh.png og:image:alt: How Claude marks AI-generated content og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How Claude marks AI-generated content **[Hacker News](https://daily.dev/sources/hn)** · 5 min read · 3 upvotes · 1 comments ## Summary Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content and outlines how Claude will mark AI-generated content. New Claude models launched on or after August 2, 2026 will embed imperceptible watermarks in generated text and attach C2PA-compliant signed provenance metadata to generated files (SVG, PNG, JPG). Marking applies across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, as well as through cloud partners AWS, Google Cloud, and Microsoft Foundry. Older models are being updated during a transition period. Key limitations include: a detected mark signals Claude processing but isn't conclusive proof of authorship, and absence of a mark doesn't confirm content is human-generated (e.g., due to heavy editing, short text, or unsupported platforms). Developers building on Claude are advised to independently assess their own Article 50 obligations. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Which Claude models will support watermarking for AI-generated content and starting when? Claude models launched in the EU on or after August 2, 2026 support machine-readable marking at launch, embedding imperceptible watermarks directly into generated text and attaching signed provenance metadata to supported file types like .svg, .png, and .jpg. Models released before that date are covered by a transition period while Anthropic works to add marking support to them. _Track how AI watermarking rules affect model choice and compliance workflows via daily.dev._ ### What file types get C2PA provenance metadata from Claude? Supported file types include .svg, .png, and .jpg, which receive signed provenance metadata following the Coalition for Content Provenance and Authenticity (C2PA) open standard. This metadata signals that a file was processed by Claude and lets you detect whether the file has been tampered with, though not every platform supports signed metadata depending on available features. _Developers weighing provenance standards across AI tools can follow updates like this on daily.dev._ ### Can I rely on the absence of a Claude watermark to prove content wasn't AI-generated? No, a missing watermark does not confirm content wasn't AI-generated or processed by Claude. Marks may be absent if the model predates marking support, the text was heavily edited, paraphrased, translated, or mixed into other writing, the passage is too short for a reliable signal, file metadata was stripped through conversion or screenshots, or the platform didn't support that marking type. _Anyone assessing AI content authenticity limitations can keep tabs on developments like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 169 comments across hackernews (as of 2026-09-14). **TL;DR:** Developers are deeply skeptical about the technical feasibility and robustness of Claude's text watermarking, raising concerns about false positives, code quality degradation, easy circumvention, and the chilling effect on legitimate users like those with disabilities — while a minority welcome the transparency signal. **Sentiment:** 15% positive · 35% mixed · 50% skeptical **The case for** - Watermarking could help institutions and individuals quickly identify AI-generated slop and filter it out. - A statistical watermark based on token-probability biasing (as in the Kirchenbauer et al. paper) is elegant and doesn't require visible markers. - Even imperfect watermarking provides some protection and is in the interest of AI companies who want to avoid training on their own output. **The pushback** - The watermark could degrade output quality by biasing token selection away from the most appropriate choices, especially in code. - Circumvention is trivially easy — paraphrasing, reformatting, or running text through a scrubber removes the signal, and removal tools appeared almost immediately. - False positives risk serious harm to students and professionals whose human-written work could be incorrectly flagged. - Watermarking code is particularly problematic since token substitution can break functionality, leaving only comments and variable names as viable carriers. - Legitimate users who rely on LLMs as assistive technology (dyslexia, ADHD, executive dysfunction) may be unfairly stigmatized. - Only Anthropic can detect its own watermark, so different providers' marks are siloed and the system doesn't generalize. - Claude's distinctive writing style ("load-bearing," "worth flagging") may already be influencing human writing, raising false-positive risk over time. **By community** - hackernews (skeptical): Commenters are broadly skeptical — debating the technical mechanism (logit-biasing vs. Unicode tricks), questioning robustness against paraphrasing and code formatting, worrying about false positives harming innocent users, and noting that watermark-removal tools appeared within hours of the announcement. **Hottest debate:** Whether the watermark can survive real-world use in code without either breaking functionality or being trivially stripped by formatters and sanitizers. **Open questions** - What exactly is the technical mechanism — logit-biasing à la Kirchenbauer/SynthID, or something else — and will Anthropic publish it? - How will the system avoid false positives for human writers whose style has been influenced by widespread LLM output? - Can the watermark meaningfully survive code formatters, linters, and refactoring tools? - Since only Anthropic can verify its own watermark, how does this help third parties (e.g., educators) who need independent verification? - What is the actual false positive rate for fully human-written content, and will Anthropic disclose it? **Highlights** > > When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response. I'd like to know a lot more about how that works. A lot of my interactions with Claude return pretty precise text. If I ask it to edit a project and refactor a specific function in several places I know exactly what I want to happen, it will NOT be OK if those refactors have some kind of weird pattern baked into their text to act as a watermark. I guess this may be covered by this: > Content generated by Claude may not carry a detectable mark if, for example: [...] The passage is very short, leaving too little text for a reliable signal; > — [simonw on hackernews · 12 comments](https://news.ycombinator.com/item?id=49250569) > But the source codes of those web apps will also be watermarked. /s > — [selcuka on hackernews](https://news.ycombinator.com/item?id=49253313) > I notice the "Limitations" section talks about how content only at some point touched by Claude may return a positive, and content that returns a negative may still be Claude generated. But I really would have liked for them to state explicitly that entirely false positives where a piece is fully human-written may still be marked as generated, because too many institutions with the power to ruin someone's life over that have trouble understanding that concept. > — [Dilettante\_ on hackernews · 2 comments](https://news.ycombinator.com/item?id=49255085) > If the algos work as advertised, watermarked token sequences have an extremely low probability. Copying the words by hand doesn't change that. The mechanism seems to survive editing. The extreme probabilities get a little less extreme, but are still extreme enough to be distinctive. But it wouldn't survive paraphrasing, because the output would be entirely human and the token correlations would disappear. It might not survive referencing if only a sentence or two is used. The practical issue is how true the claims are. It's one thing to create a proof of concept, another to see how it works in use. And this is potentially catastrophic for code, because the grammar and word choices of code are completely different and more fragile than standard English. > — [TheOtherHobbes on hackernews](https://news.ycombinator.com/item?id=49256036) > Most likely this method https://arxiv.org/pdf/2301.10226 (EDIT: and Google's SynthID paper which builds on it https://www.nature.com/articles/s41586-024-08025-4) > — [akozak on hackernews · 2 comments](https://news.ycombinator.com/item?id=49250702) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49250109) · 198 points · 169 comments ## Community discussion Top comments from developers on daily.dev. **@nerdalytics** · 1 upvotes > The political society in the EU demonstrates again that they A) don't understand technology well, and B) that free speech is democracy at its core. > > > Some politicians got hurt by "deep-fakes". This is how they [the EU politicians] want's to detect deep fakes. > > > Anthropic compiling is good thing. Because they [Anthropic] clearly lists all the issues with those kinds of watermarks. > > > Short texts can't carry enough "placeholders" to insert watermarks. Also translations remove watermarks. I wonder if that applies for AI models doing the translation too. Like translate from source language, to... --- Tags: [#claude](https://daily.dev/tags/claude), [#ai-governance](https://daily.dev/tags/ai-governance) [View this post on daily.dev](https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How Claude marks AI-generated content","url":"https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh"},"datePublished":"2026-08-10T22:26:55.503Z","dateModified":"2026-09-14T06:42:12.759Z","description":"Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content and outlines how Claude will mark AI-generated...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d9253488dd2a3e3afd526aa9c454e4d3?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d9253488dd2a3e3afd526aa9c454e4d3?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"claude,ai-governance","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"How Claude marks AI-generated content"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh","comment":[{"@type":"Comment","text":"The political society in the EU demonstrates again that they A) don’t understand technology well, and B) that free speech is democracy at its core.\nSome politicians got hurt by “deep-fakes”. This is how they [the EU politicians] want’s to detect deep fakes.\nAnthropic compiling is good thing. Because they [Anthropic] clearly lists all the issues with those kinds of watermarks.\nShort texts can’t carry enough “placeholders” to insert watermarks. Also translations remove watermarks. I wonder if that applies for AI models doing the translation too. Like translate from source language, to intermediate language, to another intermediate language, then back to source language.\nBut yeah, EU politicians will feel safer now until they find another way to attack free speech and with it democracy.","datePublished":"2026-08-11T06:45:14.314Z","url":"https://daily.dev/posts/DCnHRgkJh#c-tCfa3A8Jy","author":{"@type":"Person","name":"nerdalytics","url":"https://daily.dev/nerdalytics","image":"https://avatars.githubusercontent.com/u/97166791?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-claude-marks-ai-generated-content-dcnhrgkjh#faq","mainEntity":[{"@type":"Question","name":"Which Claude models will support watermarking for AI-generated content and starting when?","acceptedAnswer":{"@type":"Answer","text":"Claude models launched in the EU on or after August 2, 2026 support machine-readable marking at launch, embedding imperceptible watermarks directly into generated text and attaching signed provenance metadata to supported file types like .svg, .png, and .jpg. Models released before that date are covered by a transition period while Anthropic works to add marking support to them. Track how AI watermarking rules affect model choice and compliance workflows via daily.dev."}},{"@type":"Question","name":"What file types get C2PA provenance metadata from Claude?","acceptedAnswer":{"@type":"Answer","text":"Supported file types include .svg, .png, and .jpg, which receive signed provenance metadata following the Coalition for Content Provenance and Authenticity (C2PA) open standard. This metadata signals that a file was processed by Claude and lets you detect whether the file has been tampered with, though not every platform supports signed metadata depending on available features. Developers weighing provenance standards across AI tools can follow updates like this on daily.dev."}},{"@type":"Question","name":"Can I rely on the absence of a Claude watermark to prove content wasn't AI-generated?","acceptedAnswer":{"@type":"Answer","text":"No, a missing watermark does not confirm content wasn't AI-generated or processed by Claude. Marks may be absent if the model predates marking support, the text was heavily edited, paraphrased, translated, or mixed into other writing, the passage is too short for a reliable signal, file metadata was stripped through conversion or screenshots, or the platform didn't support that marking type. Anyone assessing AI content authenticity limitations can keep tabs on developments like this on daily.dev."}}]} ```