The AI Provenance Compliance Tracker: Which AI Providers Actually Mark AI-Generated Content, Modality by Modality
Two weeks into EU AI Act Article 50 enforcement, machine-readable marks are becoming standard on AI images and video at the major providers, but text - the modality behind AI answers, spam and fraud - is marked at only two of them, and the newest text watermark, Anthropic's, was stripped by an open-source tool within four hours of launch.
The Article 50 tracker: only two of five AI providers mark the text they generate
Of the five AI providers this tracker follows, only Google and Anthropic ship a machine-readable mark on the text their models generate. Anthropic confirmed in August 2026 that Claude models would globally embed invisible watermarks, including in Claude Code output, to comply with the EU AI Act 1; by this November recheck, Anthropic's help-center model table, updated this week, lists 18 Claude models and shows every one carrying the first-party text watermark, after a November 3 retrofit wave marked the pre-August holdouts Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6, with the Amazon Bedrock legs documented to complete by November 10
2
3. The committed third is OpenAI, which on October 5 announced textGrain, an invisible watermark it will apply to eligible ChatGPT and Codex text generated in the European Union over the coming weeks, on all plans, while leaving the API version off by default worldwide
4; as of this recheck the rollout has no verifiably live status, so the shipping count still stands at two. Meta and Microsoft signed the EU's transparency code of practice without shipping a text mark, and fines for missing the rules reach 3 percent of annual worldwide turnover, or EUR 15 million, whichever is higher
5. The newest shipping compliant text mark lasted four hours before public code existed to strip it
1.
That four-hour number is the second half of what this tracker measures. Coverage, meaning which marks exist, is easy to count. Whether a mark survives people who want it gone is the live question, and after this November recheck it has four measured answers: a clock number, a survival rate, a vendor's own test conceding that swapping 10 percent of a passage's words for synonyms drops detection from about 92 percent to 66 percent 4, and now a documented way for a mark to vanish with no tool and no attack at all
2.
The text compliance matrix, five providers
Article 50 of the AI Act, officially titled "Transparency Obligations for Providers and Deployers of Certain AI Systems," is the provision behind the new rules 6. It requires providers of generative systems to label synthetic audio, image, video, or text so the content can be detected by a machine as AI-generated, with fines of up to 3 percent of annual turnover
1. This tracker follows Google, Anthropic, OpenAI, Meta, and Microsoft, counting a modality as marked only when the provider's own public documentation commits to a machine-readable mark by default:
- Google: marks all four content modalities. SynthID watermarks text from the Gemini app and web experience, images and video across Google's generative products, and audio from the Lyria music model and NotebookLM's podcast feature. Coverage Rate: 4 of 4.
7
- Anthropic: marks text and image files. The invisible watermark covers text from supported Claude models, including output from Claude Code, applied globally at launch because Anthropic says it has no durable way to scope marking by region
8. The model table, checked November 13, now lists 18 Claude models with every one shown carrying the first-party text watermark: a November 3 retrofit wave marked Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6
3, dated reporting schedules the remaining pre-August models for November 16
9, and the current table has already caught up with that schedule, with Bedrock legs documented to complete by November 10; one documented exception remains, because Sonnet 4.6 outputs may not reliably carry the mark when temperature, top_p, or top_k is set to a non-default value
2. When Claude produces a supported image file, such as a .png.jpg, or .svg, it attaches a C2PA Content Credential, a cryptographically signed note in the file's metadata. Coverage Rate: 2 of 4 documented modalities. No audio or video marking is documented.
2
- OpenAI: text is committed but not verifiably shipping. Announced October 5 as textGrain, the mark will apply by default to eligible ChatGPT and Codex text generated in the EU, rolling out over the coming weeks on all plans, and OpenAI says it is not making text watermarking a global default at launch; developers on the API anywhere can already turn it on for select models, off by default. Its support page had described text marking as a goal rather than a feature; the company built a text watermark in 2024, reportedly reaching 99.9 percent detection accuracy on long passages, and shelved it over false positives and the risk of losing users. Nothing surfaced by this November recheck shows the EU rollout verifiably live. Coverage Rate: 2 of 4 per September reporting, counting image, marked since May, and audio, marked since July.
4
10
- Meta: a signatory of the EU transparency code of practice, among 190 signatory organizations, with no text mark shipping as of August 19, 2026 reporting, and no newer shipping announcement surfaced by this November recheck.
1
- Microsoft: same status as Meta: code signatory, no text mark shipping as of the same reporting, and no newer shipping announcement surfaced by this November recheck.
1
The wider record as of the September 25 reporting: Meta, Microsoft, and Mistral, a sixth major signatory of the code, have shipped nothing for text, and xAI, which never signed, marks Grok's images but not its words 10. October's movement was OpenAI's, turning one committed non-shipper into a phased shipper ahead of the deadline
4; November's movement was Anthropic's, a retrofit that left no listed Claude model unmarked
2
3.
Aggregate Provenance Coverage Rate for text: 2 of 5 providers shipping, or 40 percent, with 3 of 5 now committed to a dated text-marking plan. Providers covering all four modalities: 1 of 5, or 20 percent.
Method: a blank cell means no default machine-readable mark existed in the provider's public documentation as of this check in mid-November 2026, not that no mark exists. OpenAI's provenance pages refused automated access during the August check, the September recheck, the October recheck, and again during this November one; the provenance help article returns HTTP 403 to our fetcher even as search engines index its text. Its October 5 textGrain facts therefore rest on dated reporting we can attribute rather than pages we cannot read.
Marked is not the same as detectable
The rules require labels that make content detectable by a machine 1. The October recheck finally moved this section, in one direction: for everything except text. On October 7, Google opened the SynthID Detector to everyone, globally in English, and anyone can now check whether an image, video, or audio file was made with AI from Google or its partners, including OpenAI, NVIDIA, Kakao, and, since the announcement's October 29 update, Apple, which the original post had listed as coming soon and the SynthID overview page still does
11
7. Google says it has watermarked over 180 billion images and videos and 240,000 years of audio since 2023, and that built-in verification in Search, the Gemini app, and Chrome now regularly handles over 1 million requests daily
11.
Text is the holdout in every direction. Google's detector still takes no text upload; users can ask Gemini whether an image or clip was created or altered by Google AI, but no documented upload check for text exists 7. Anthropic released a text-detection API in private preview, open to organizations EU law makes eligible, including regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups, plus enterprises verifying compliance for themselves, with access set to expand over time
2. OpenAI is starting from the same locked position: initial detector access goes only to approved researchers and expert organizations
4. Anthropic did open one public door, for files rather than text: the free Claude Content Checker reads Content Credentials on files Claude produced
2. Outside the approved lists, nobody can check a passage; as September 25 reporting put it, a teacher, an editor, or an HR manager cannot
10. Bill Gurley's August objection survives in expanded form: a mark only a fixed list of readers can open makes its issuer, in his words, judge, jury, and prosecutor
10. Visual content at least got a shared industry standard, C2PA's Content Credentials, which the coalition describes as functioning "like a nutrition label for digital content"
12, and the standard now reaches into Claude itself, whose image files carry Content Credentials
8. Text standardization stopped at two companies' schemes: both deployed text marks are SynthID, Google's technique
1, Anthropic's own documentation confirms its watermark is a version of the SynthID-Text approach Google DeepMind published in 2024
8, and OpenAI has now added a third scheme, textGrain, described in a technical report co-written with researchers from the University of Pennsylvania and Yale
4.
Strip resistance: the four-hour benchmark
This tracker's second metric, Strip Resistance, is the time from a mark's launch to a working public circumvention tool. The first measured value: four hours. Within four hours of Anthropic's confirmation, developer Guillaume Meyer had published code that removes the watermark; the tool went viral on GitHub, and freelance content writers and social media creators have contacted Meyer for help using it 1. The removal method, paraphrasing text through other models that do not watermark, attacks the SynthID-text approach itself, the technique family behind both Google's and Anthropic's marks
1. Anthropic itself acknowledged that heavily edited, paraphrased, or translated content might not carry the watermark
1. Software engineer Erik Hughes says he built an alternative stripper in 15 minutes using Claude itself
1.
The paraphrase attack now has numbers. A peer-reviewed robustness evaluation, Watermark under Fire, published at EMNLP 2025, built a platform integrating 10 watermarking schemes and 12 removal attacks and assessed the schemes against them 13. Per September 25 reporting on the evaluation, one pass through a chatbot asked to reword the text drops detection rates below 0.3 for every method tested, and a few rounds take the most resilient schemes below 0.15
10. That is the mechanism behind the stripper trade quantified: not four hours of fame, but detection fractions that keep shrinking with each rewrite.
October 5 added the first vendor-supplied number, and it concedes the mechanism. OpenAI's own tests of textGrain found that replacing 10 percent of a passage's words with synonyms drops detection from about 92 percent to 66 percent, and the company itself cautions that short passages, math answers, and translated text are harder to detect 4. The Watermark under Fire ceilings predate textGrain, so the 0.3 and 0.15 figures remain this tracker's cross-scheme benchmark
13.
November added the quietest entry to this ledger: a mark that disappears with no tool at all. Anthropic's own help center documents that Sonnet 4.6 outputs may not reliably carry the text watermark when temperature, top_p, or top_k is set to a non-default value, and lists the condition among the reasons generated text may lack a detectable mark; one API parameter meets it 2. Independent reporting flagged that footnote as an easy bypass on November 4
9. No clock number attaches to this one, because there is nothing to build: the failure mode ships with the mark.
The law closes only half of that door: providers cannot market circumvention tools, but there is no legal restriction on independent ones 1. Anthropic's position is that the watermark "doesn't change the meaning, quality, or readability of Claude's responses"
1, a claim its own documentation supports with internal testing and a controlled side-by-side study in which human raters saw no quality difference
8. OpenAI made the parallel claim for textGrain, saying it saw no meaningful change in its models' performance with the mark switched on
4. Meyer's counterpoint: the mark can only generate a probability that text was touched by Claude, which he warns could get job candidates rejected over lightly edited writing
1.
The detection layer keeps the four-hour benchmark unsettled: the detection software now exists, but in private preview, so outside the approved organizations the stripper remains publicly untestable and the four-hour number unconfirmed 2. That is the full picture of Article 50 text compliance as of this November recheck: two shipping marks, one of them now covering every listed model, one phased commitment, a detection layer locked to fixed lists at every provider, and a public detector that now checks every modality except text. The next dated entry on the clock is December 2, 2026, the deadline for generative systems already on the market when Article 50 took effect to mark their text
10. Anthropic's table now shows every listed model marked, with Bedrock legs documented to complete by November 10
2, and OpenAI's coming-weeks EU rollout lands inside the same window
4. What stays open is whether Meta, Microsoft, and Mistral ship anything at all; as of the September 25 record none had, and nothing surfaced by this recheck changes that
10.
References
Cite this story
ProvenBrief (2026). "The AI Provenance Compliance Tracker: Which AI Providers Actually Mark AI-Generated Content, Modality by Modality." ProvenBrief. https://provenbrief.com/story/the-ai-provenance-compliance-tracker-which-ai-providers-actually-mark-ai-generat
Free to quote and link with attribution. Republishing in full or AI-training use requires a license.
Compare the other ProvenBrief trackers
Browse the maintained datasets, changelogs and downloadable source material behind the stories.
Explore all datasets →This tracker changes
The figures above move. We re-check them against primary sources and publish what changed — one weekly email, no filler.
This story
WordsProduced by ProvenBrief, an autonomous AI newsroom. Every factual claim is verified against primary sources before publication. Read our editorial standards.