E-Paper
Claude’s watermark knows AI was used. It doesn’t tell if you wrote the essay, it wrote the essay or just fixed a comma
Anthropic’s watermark rollout promises transparency. What it can’t promise is telling readers how much of that content is actually AI’s own work .
Anthropic confirmed this week that it has started embedding invisible, machine-readable watermarks into everything its Claude AI models generate, a global rollout triggered by new transparency rules under the European Union's AI Act. The change reaches users everywhere: on the web, the mobile app, coding tools and cloud platforms, regardless of whether the EU rules apply to them directly.
The shift raises an obvious question: what does a watermark actually prove, and what happens once AI text is edited or blended with a person's own writing?
The short answer is that a watermark only signals that content passed through Claude. It is not confirmation that AI wrote every word.
What is this watermark, exactly?
Think of it as a hidden signature. Anthropic uses two different mechanisms depending on what Claude produces.
For text, Claude weaves a hidden mathematical pattern into its word choices as it writes. The pattern doesn't alter how natural the writing reads, but software built to scan for it can spot the signal instantly; even after the text has been copied into another document.
For files such as SVG, PNG and JPG images, Claude attaches a cryptographically signed digital tag that functions as a tamper-proof label confirming Claude generated the file.
The marking isn't confined to Claude's chat interface. Anthropic said supported models carry the watermark across Claude, its API platform, Claude Code, Claude Cowork and Claude Tag. The text watermark works the same way globally; only the file-tagging may differ slightly by platform.
How does it survive editing?
Anthropic has not published the precise mechanics of its text watermark. What is known is that the mark is embedded as Claude generates the text, and it remains attached through copy-pasting and light revisions.
The Independent reported that most AI text watermarks operate on a similar principle: the model subtly favours certain words or phrasing over others as it writes. A reader may not notice anything unusual, but an automated detector can pick up the underlying statistical pattern.
Does every piece of Claude's output carry a mark?
Not yet, and not always. Anthropic's watermarking commitment covers supported models launched on or after 2 August 2026. The models released earlier are still being retrofitted.
There are technical limits too. A watermark may not be reliably detected in very short passages, in text that has been heavily edited, paraphrased or translated, or where Claude's output has been combined with other material.
File tags become more uncertain. A screenshot, a format conversion from PNG to JPG, or simply saving a new copy can strip them out entirely.
A detector finding nothing, in other words, does not mean the content wasn't AI-generated.
What if someone only uses Claude to edit their own writing?
This is where the line between AI generation and AI processing would matter.
If a person drafts an article and asks Claude only to fix grammar or tighten the wording, the resulting text can still carry a Claude watermark, even though the ideas and the original writing are entirely the person's own.
Anthropic acknowledges this limitation directly. It says “Claude may not be the original author”. People can use the system to proofread, translate, summarise or convert material, yet the output can still be marked.
A detected watermark, then, should not be read as evidence that a person used AI to write an entire piece; only that Claude processed the content at some stage.
What happens when AI and human writing are mixed?
Establishing authorship gets harder still with blended content. If someone writes several paragraphs themselves, generates others with Claude, and combines the two, a detected watermark could indicate that Claude touched part of the document; but it would not reveal which sentences came from where, or confirm who wrote the unmarked sections.
Anthropic has warned that its output can be “modified, excerpted, or combined with other material”, further complicating any attempt to trace it.
Also Read: Why Anthropic doubled its AI policy funding to $40 million amid growing AI safety concerns
Is Claude the only AI doing this?
No. Google DeepMind has built a comparable system called SynthID, which watermarks AI-generated images, audio, video and text. For text specifically, SynthID is deployed through the Gemini app and web experience, designed to embed an imperceptible statistical signal during generation.
Google has also cautioned that SynthID is a useful identification tool but it's not a "silver bullet" for spotting AI content.
Why watermark AI content at all?
The intent to add watermarks is transparency. With AI-generated text, images and video now widespread online, watermarks give readers and organisations a way to spot AI-made content and trace where it originated.
The catch is that a watermark may only show that AI touched a piece of text, not how much of it was created by AI. It also cannot distinguish between AI drafting an entire essay and Claude correcting a single typo in a human-written one. Similarly, an absence of watermark doesn't prove or disprove the use of AI either.
The technology, it appears, is better suited to tracking aggregate AI use than to settling individual disputes. Relying on it as definitive proof could risk wrongly accusing students or employees of cheating when their only use of AI was basic proofreading.
