Anthropic changed how Claude labels its output to comply with new European transparency law. Here is a plain-English explainer of what actually changed and what it means for you.
What changed
To satisfy the EU AI Act's Article 50 transparency obligations, which became enforceable on August 2, 2026, Anthropic began marking Claude's output so it can later be identified as AI-generated. Two things happen:
- Generated text carries an invisible, machine-readable watermark.
- Generated files in supported formats (such as PNG, JPG, and SVG) get digitally signed provenance metadata following the open C2PA standard.
The marking happens at the model level, so it applies across the surfaces where Claude is used, and Anthropic rolled it out globally rather than limiting it to European users. New models launched on or after August 2, 2026 mark from launch, while older models are being updated over a transition period.
Why the EU AI Act triggered it
Article 50 of the EU AI Act requires that systems generating synthetic content mark their output as artificially generated, so people can tell when they are looking at AI-made text, images, audio, or video. Because building the marking into the model is simpler than switching it on and off by region, Anthropic applied it everywhere. Several other AI providers face the same obligation, so this is an industry-wide shift, not a Claude-only one.
How the watermark works, and its limits
This is the part worth understanding clearly:
- The text watermark is a key-based, machine-readable signal. Detecting it reliably requires the provider's detection method, which is different from how third-party AI detectors work.
- It is imperfect by design. Anthropic has been candid that a file's mark can be lost through format conversion, re-saving, screenshots, or similar edits.
- A detected mark means only that Claude may have processed the content, not that a human had no hand in it. Someone might have written the ideas themselves and used Claude only to translate, summarize, or proofread.
- The absence of a mark does not prove a human wrote something.
So the watermark is a provenance signal with real caveats, not a definitive AI-or-human verdict.
How this differs from AI detectors
AI detectors like the ones teachers and editors use do not read this watermark. They guess from statistical patterns in the writing, phrasing tells and predictability, without any key. That approach is unreliable and produces false positives. The Claude watermark is a separate, key-based mechanism, which is more precise about "did Claude touch this" but says nothing about quality or authorship intent.
Is this the same as the hidden characters in pasted text?
No, and this is a common point of confusion. The invisible characters that show up when you paste text from a chatbot, zero-width spaces, non-breaking spaces, and the like, are formatting artifacts, not Anthropic's cryptographic watermark. A text cleaner like textscrubr removes those formatting characters to make your text portable and clean, which is a different thing from the model-level watermark this update introduces. Cleaning formatting noise is about tidy, professional text, not about defeating a transparency measure.
Update: October 2026
Anthropic has since explained that the text watermark works by subtly influencing which of several equally reasonable words Claude picks, using a version of Google DeepMind's SynthID-Text method. It adds no hidden characters, and Anthropic has said it plans to offer a detection API (Anthropic's explainer). OpenAI announced a similar word-choice watermark, textGrain, on October 5, 2026: it is on by default for ChatGPT and Codex text in the EU, and API customers elsewhere can opt in. Microsoft took a different route for some Windows AI text APIs, which add a fixed invisible four-character sequence, U+200B U+034F U+2060 U+200B (Microsoft's documentation). That sequence is made of characters, so a cleaner can find it. Word-choice watermarks are not characters, so no cleaner can remove them.
What it means for you
For most people, day-to-day use of Claude does not change: no visible impact on quality, speed, or pricing. If you work in an EU-regulated setting, the practical takeaways are that Claude output may now carry a provenance mark, that the mark is not proof of full AI authorship, and that ordinary editing and format conversion can remove file marks. As Anthropic publishes more technical detail and a detection tool, expect the specifics to evolve, so treat this as the state of things as of late October 2026.