Claude’s New AI Watermarks Do Not Prove Who Wrote the Text (2026)

A colleague sends you a polished memo, and you want to know whether a person wrote it. An invisible AI watermark sounds like a useful answer.

The EU's Article 50 transparency rules have applied since August 2. Among other duties, providers must make covered AI-generated or manipulated output machine-detectable. Anthropic says Claude models launched in the EU on or after that date support marking at launch, and those marks apply worldwide where supported. 1 2

But the useful answer is narrower than it first appears. My default is: a detected mark is a clue about processing; no mark is no conclusion; neither result verifies the claim itself.

Three signals, three different jobs

These tools do different jobs. The EU rule for people and businesses publishing AI-generated public-interest text also has an important boundary: substantive human review or editorial control can exempt that text from the visible label requirement.

Signal	What it can tell you	What it cannot settle
Visible disclosure	A person is dealing with an AI system, or a publisher is identifying certain AI-generated material. EU rules require notice at the start of a direct AI interaction unless the context makes it obvious. 1	Whether the output is accurate, fair, or genuinely reviewed by a person.
Text watermark	A hidden signal travels inside supported Claude text when it is copied and pasted, and may survive some editing. 2	Who supplied the ideas, whether Claude wrote the first draft, or whether the text changed later.
Signed file history	Supported Claude image files can carry C2PA provenance metadata that signals Claude processed the file and can help detect tampering. 2	Whether the scene shown is real, whether a caption is honest, or what happened after the metadata was stripped.

That distinction matters because "made with AI" is not one condition. Claude might translate a human's paragraph, summarize a public report, convert a file, generate the whole thing, or merely proofread it. A positive mark can cover all of those paths.

Read a positive result narrowly

Anthropic's own wording is careful: a detected mark means the content may have been processed by Claude. The company says the original ideas, text, or data may have come from someone else, and the result may have been edited, excerpted, or mixed with other material afterward. 2

So a positive result supports only a limited statement: "this content may have been processed by Claude."

It does not support these stronger claims:

"Claude invented the argument."
"No person reviewed this."
"The document is false."
"The file has never changed."

Those are separate questions about authorship, editorial responsibility, evidence, and custody. A provenance system records part of the route. It is not a fact-checker.

Treat a negative result as unresolved

This is the more important rule. Anthropic lists several ways Claude-processed content can lack a detectable mark: an older model, heavy editing, paraphrasing, translation, a passage too short to test reliably, a screenshot, a format conversion, stripped metadata, or an unsupported platform or file type. 2

The law also leaves ordinary gaps. Content generated before August 2 does not need a retroactive label. Systems already on the EU market before that date have until December 2 to meet the machine-readable marking requirement. Personal, non-professional use is outside the Act's scope. 1

A blank detector result therefore means only "this checker found no supported mark." It should never be upgraded to "a human made this."

Use this verification order

When the content could change a purchase, a vote, a job decision, a health choice, or someone's reputation, provenance is one check near the start, not the finish.

Keep the original. Test the original file or complete passage when possible. Screenshots, copied fragments, and re-saved files can discard the signal you are trying to inspect.
Use the checker that matches the mark. Google says its Gemini app can check SynthID in images, video, and audio; users can upload the media and ask, "Is this made with AI?" Google reported in May that the check had been used 50 million times. 3 Anthropic says detection tools for Claude's marks are coming, but its public guidance still describes the technical documentation as forthcoming. 2
Read the result literally. "Processed by" is not "written entirely by." "No supported mark" is not "human-made." Record the exact claim the tool makes before drawing a larger conclusion.
Verify the substance elsewhere. Open the cited report, locate the original photo or recording, check the date and account, and confirm consequential claims with an independent source that owns the fact. A watermark cannot tell you whether a number was copied correctly or a scene was staged.
Ask who took responsibility. For public-interest text, the EU's label exemption requires substantive human review or editorial control by someone able to approve, change, or reject the content. A spelling or grammar pass is not enough. 1 That is a useful question outside Europe too: who checked the evidence and is willing to stand behind the result?

A sensible default for AI labels

Use a positive mark to reconstruct how content moved. Use a visible label to calibrate how much independent checking you need. If there is no mark, leave the authorship question open.

Then check the claim, not just the file. A label can help trace the route; it cannot tell you whether the claim is true.