ccwatermark

News and public record

Everything publicly known about Anthropic’s marking, with sources. All of it is announcement-level; no technical detail has been published as of 2026-08-10.

Timeline

Date	Event
2023	Kirchenbauer et al. publish the green/red logit-bias watermark (KGW)
2024-05	ETH SRI publish black-box detection of deployed watermarks; test GPT-4, Claude 3, Gemini 1.0 Pro
2024-10	Google DeepMind publish SynthID-Text in Nature; deploy on Gemini and open-source the scheme
2025	ETH SRI paper appears at ICLR 2025
2026-08-02	EU AI Act Article 50 transparency obligations bite. Claude models launched in the EU on or after this date support machine-readable marking at launch
2026-08-10	This repository’s baseline collected. No technical documentation exists

What Anthropic has said

From the Claude help centre article:

• Generated text carries “an imperceptible watermark directly into the text itself”; because it is part of the text it “will travel with the text when it’s copied and pasted elsewhere”.
• Generated files of supported types carry “signed provenance metadata” following the C2PA standard.
• Coverage is “output from supported Claude models across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag” — and applies worldwide, not only in the EU.
• “Some platforms or features may not support certain marking types.”
• A detected mark “provides a signal that content was processed by Claude, but is not fully conclusive”; absence of a mark proves nothing.

What is still unpublished

• The scheme. No method is named — not SynthID, not KGW, not tournament sampling. Press descriptions (“subtly biasing the model’s word choices during generation”) are generic and appear to be paraphrase, not disclosure.
• The detection method. Anthropic “expects to provide details about how people can detect Claude’s marks, as required under EU law, in forthcoming documentation” — with no name, timeline, or URL.
• Which models are covered. No model list has been published — but the cohort boundary is readable from the API, see below.

Which cohort every model belongs to (readable, 2026-08-12)

No model list has been published, but the Models API exposes created_at per model, so the cutoff can be applied directly. src/models_snapshot.py:

created_at	model	cohort
2025-09-29	claude-sonnet-4-5-20250929	pre-cutoff
2025-10-15	claude-haiku-4-5-20251001	pre-cutoff
2025-11-24	claude-opus-4-5-20251101	pre-cutoff
2026-02-04	claude-opus-4-6	pre-cutoff
2026-02-17	claude-sonnet-4-6	pre-cutoff
2026-04-14	claude-opus-4-7	pre-cutoff
2026-05-28	claude-opus-4-8	pre-cutoff
2026-06-07	claude-fable-5	pre-cutoff
2026-06-29	claude-sonnet-5	pre-cutoff
2026-07-24	claude-opus-5	pre-cutoff — by nine days

Every model currently on the API predates 2026-08-02. Zero are in the “marked at launch” cohort. All ten are transitional-cohort models whose marking status Anthropic describes as still being worked on, and which it has never enumerated.

So every result in this repository was collected on a model that is not required to have been marked at launch. That does not make the nulls worthless, but it does mean they test the cohort where the answer is expected to be murky.

Two caveats on created_at: it need not match the date embedded in a model id (claude-haiku-4-5-20251001 reports 2025-10-15), and the endpoint lists available models, so retired ones are absent.

The experiment this sets up

The first model with created_at >= 2026-08-02 is the cleanest test available: marked at launch by Anthropic’s own statement, so it is a positive control rather than a guess. Better still, it can be run against a pre-cutoff sibling on identical stimuli — same vendor, tokenizer and serving stack, differing in marking status. That is a far tighter control for the context-sensitivity confound in findings.md than another vendor’s model.

src/models_snapshot.py is the tripwire; it costs nothing and flags new arrivals against the last snapshot in data/models/.

The load-bearing detail

Anthropic is still adding marking to already-released models: it is “working on models it has already released, to add output marking during the transition period allowed under EU law”, described elsewhere as “transitional marking support” for older models.

This matters more than anything else in this file:

1. It means a model’s marking status changes over time, so any measurement is only valid at its collection date. This is why data/ is organised by date.
2. It means null results collected now most plausibly mean not yet marked rather than never marked.
3. It makes a pre-rollout baseline valuable and perishable — differencing a later run against an earlier one on identical prompts detects activation without needing Anthropic’s documentation at all.

Sources

• How Claude marks AI-generated content — Anthropic help centre, primary source
• Anthropic pledges to embed watermarks to help discern AI slop in sop to EU — The Register
• Anthropic embeds invisible watermarks in Claude models worldwide — Cryptobriefing
• Anthropic puts hidden watermarks on Claude text under new EU rules — Interesting Engineering

Secondary coverage repeats the help-centre text closely; treat the help centre as the only authoritative source and the rest as paraphrase.

Watch list

Article 50 is a legal forcing function, so the documentation should appear. Check:

• The help-centre article above — likely to gain a link when the doc lands
• Anthropic’s research and engineering publications
• arXiv, for an Anthropic-authored watermarking paper
• EU AI Office GPAI documentation and code-of-practice filings
• USPTO / Justia for watermarking or provenance patents (no relevant filing found as of 2026-08-10; Justia blocks automated fetches, so check by hand)

---

Part of ccwatermark — independent research on AI text provenance marking. Overview · Scope and ethics · Source on GitHub

This site is open source. Improve this page.