# Anthropic starts watermarking new Claude text globally

Anthropic has begun applying a SynthID-Text-style watermark to newly launched Claude models worldwide, with a detection API still forthcoming.

Claude text watermarking is now part of Anthropic’s compliance approach for newly launched models. It does not add tokens or user identifiers, but it is weaker on short, factual, lightly edited and code-heavy outputs, limiting how provenance claims should be used.

- Status: Active
- Published: 2026-08-18T21:47:21+12:00
- Updated: 2026-08-18T21:47:21+12:00
- Categories: Artificial Intelligence, Inference & APIs
- Tags: AI provenance, Anthropic, Claude, EU AI Act, SynthID-Text, watermarking
- Canonical HTML: https://beyondthe.news/dossiers/anthropic-claude-text-watermark-synthid

## What changed

Anthropic announced on August 14, 2026 that future Claude models will generate watermarked text. The company is applying the watermark globally at launch to newly launched models, using a version of the SynthID-Text method, while older models launched before August 2, 2026 are covered by a transition period and will be updated over coming months.

## Why it matters

AI SaaS builders, compliance teams and publishers may want provenance signals, but the watermark is probabilistic and incomplete. Treating it as definitive authorship detection would create false confidence, especially for short factual text, proofreading, code and heavily rewritten output.

## The watermark changes token choice, not visible text

Anthropic says its implementation changes the source of randomness used when selecting among equally acceptable next words. It does not insert hidden characters, add extra tokens, identify a user or organization, or materially change price or latency.

## Detection has important blind spots

Anthropic says confidence improves with longer passages and falls when the model has few valid lexical choices. Factual text, proofreading, code and light edits therefore contain less usable watermark signal. A complete rewrite can remove the signal.

## The detector is not yet generally available

Anthropic says it will soon offer a watermark detection API but has not yet published its final implementation details. Builders should not design enforcement workflows around a detector that is not yet available.

## The rollout is tied to transparency compliance

Anthropic cites the EU AI Act and the July 2026 transparency code of practice as the reason for the change, while choosing a global launch because it does not currently have a durable region-scoping mechanism.

## Key details

- Anthropic announced Claude text watermarking on August 14, 2026.
- The method is a version of SynthID-Text.
- Watermarking is global for newly launched Claude models.
- Older models launched before August 2, 2026 are being transitioned over coming months.
- The watermark does not identify the user, organization or chat.
- Anthropic says there is negligible speed impact and no extra token cost.
- A detection API is planned but not yet generally available.

## Builder takeaways

- Do not use watermark presence or absence as a binary authorship verdict; treat it as one provenance signal with known failure modes.
- If your product will expose watermark detection, document weaker confidence for short text, code, proofreading and rewritten content.
- Separate text-watermark logic from C2PA handling for generated files; Anthropic describes those as different provenance mechanisms.
- Wait for the detector API specification before making hard compliance or moderation guarantees based on Claude watermark detection.

## What to watch

- Launch and terms of Anthropic's watermark detection API.
- Rollout timing to pre-August-2 Claude models.
- Independent false-positive and false-negative testing across languages and editing conditions.

## Uncertainties

- The detector API is not yet available, so operational thresholds and access terms are unknown.
- Real-world robustness under paraphrasing, translation chains and mixed human/AI editing remains to be independently characterized.

## Timeline

- **2026-08-14 — Anthropic announces Claude text watermarking:** Anthropic described its SynthID-Text-based implementation and global rollout for newly launched models.

## Sources

- [How Claude’s text watermark works](https://www.anthropic.com/news/claude-text-watermark) — Anthropic · primary · 2026-08-14T00:00:00+12:00. Anthropic’s implementation details, rollout scope, detector plans and stated limitations.
- [Scalable watermarking for identifying large language model outputs](https://www.nature.com/articles/s41586-024-08025-4) — Nature · research · 2024-10-23T00:00:00+13:00. Peer-reviewed SynthID-Text method, production evaluation, quality results and robustness limitations.