A watermark can be useful evidence, but like all evidence, you need to know what it is actually telling you instead of what you wish it could prove.

Anthropic has begun adding machine-readable markings to outputs from newer Claude models as AI transparency obligations under the European Union's AI Act take effect. News coverage of the rollout reports that Claude text carries an invisible signal, while files can include signed provenance metadata based on the C2PA standard. The European Commission's Article 50 guidance requires providers to support detection of AI-generated or manipulated content through machine-readable marking where technically feasible.

That is useful infrastructure. It gives systems a way to carry information about where an output has been processed or generated.

It also creates a predictable semantic problem: people are going to treat “this output contains an AI-origin mark” as equivalent to “AI wrote this.”

One Signal, Several Different Questions

Consider a human-written paragraph that someone sends to Claude with the instruction, “Fix the punctuation and make the second sentence less awkward.” If the resulting output carries Claude's machine-readable mark, what does detection establish?

It can provide evidence that the text passed through a Claude generation process. It does not, by itself, tell us who developed the argument, who supplied the facts, whether the prose was mostly human before the edit, how much the model changed, whether a human reviewed the result, or who accepted responsibility for publication.

Now change the example: a user gives Claude three source documents and asks it to draft an entire article from scratch. The resulting text might carry the same class of machine-readable signal.

The provenance fact can be similar while the authorship history is very different.

That is the information-architecture problem hiding inside the watermarking debate. Provenance, authorship, assistance, editorial control, and accountability are related properties, but they are not interchangeable properties.

When a system collapses them into a single “AI-generated” flag, downstream users are forced to guess what the flag means.

Provenance Answers a Narrower Question

Provenance is about origin and history. It helps answer questions such as:

  • Which system produced or transformed this output?
  • Which tool or provider was involved?
  • Can the artifact carry evidence about that processing step?
  • Can another system detect the evidence later?

Those are valuable questions. They are especially useful for platform moderation, content-disclosure workflows, chain-of-custody records, media verification, and organizational governance.

Authorship asks something else. It is concerned with who created the substantive expression, reasoning, structure, or intellectual contribution. Human review asks whether a person examined and approved the output. Accountability asks who is responsible for the decision to use or publish it.

A single artifact can have one answer for each field. For example:

  • Provenance: Processed by Claude Sonnet 5.
  • Source authorship: Draft supplied by an employee.
  • AI contribution: Copyediting and sentence restructuring.
  • Human review: Reviewed by the employee and a technical editor.
  • Publication accountability: Approved by the communications director.

That record is far more useful than a binary field labeled AI = YES.

A Detector Is Not a Verdict Machine

The temptation to overread watermarking will be especially strong in schools, hiring, publishing, compliance, and workplace investigations. A machine-readable signal looks objective. Objective-looking evidence tends to acquire authority very quickly.

But Anthropic's reported implementation has limitations. News coverage notes that short passages and editing can weaken detection reliability. That means the absence of a detectable mark does not necessarily prove that Claude was never involved. At the same time, the presence of a mark does not tell us what kind of involvement occurred.

This produces two different error risks:

  • A false inference from presence sounds like: “The detector found a Claude mark, therefore the person did not write this.”
  • A false inference from absence sounds like: “The detector found no mark, therefore no AI system was involved.”

Neither conclusion follows automatically from the signal.

The detector provides evidence about a technical property. The organization still has to define what decisions that evidence is allowed to support.

The EU Rule Is About Transparency, Not Mind Reading

The European Commission's transparency rules make the distinction easier to see. Article 50 requires providers to support machine-readable marking of AI-generated or manipulated content, and it imposes disclosure obligations in particular contexts, including some public-interest text and deepfakes.

The regulatory problem is content provenance and disclosure. The rule does not magically convert a watermark into a complete authorship history.

That matters because real content workflows are mixed. A person might brainstorm with a model, write independently, ask for a rewrite, insert two AI-generated paragraphs, restore most of the original wording, and then send the result through another tool for translation. A file might contain human text, generated charts, copied quotations, and AI-edited captions.

“Was AI used?” has become too blunt a question for many serious workflows.

Better systems ask what kind of AI use occurred, where it occurred, what the source material was, who reviewed the result, and what decision depends on knowing that history.

This Is an Information-Typing Problem

Structured systems become more reliable when different facts get different fields.

We already understand this in other domains. A document's owner is not necessarily its author. Its last editor is not necessarily the person who approved it. Its publication date is not its review date. Its source repository is not its audience.

Collapsing those properties would make a documentation system less useful because each answers a different operational question. AI provenance needs the same discipline.

An organization might eventually track fields such as:

  • AI provider and model
  • AI task type
  • source-material origin
  • percentage or scope of generated content, when that can be defined meaningfully
  • human reviewer
  • approval status
  • transformation history
  • machine-readable provenance status
  • disclosure requirement
  • publication owner

Not every workflow needs all of them. The point is that the data model should match the decision being made.

If a university is trying to enforce an assignment rule, it needs evidence about prohibited assistance, not merely evidence that text touched an AI tool. If a newsroom is trying to maintain disclosure standards, it needs a policy that distinguishes research assistance, transcription, copyediting, generation, and synthetic media. If a company is investigating confidential-data handling, it needs to know which external system received the information, regardless of who ultimately wrote the prose.

Different questions require different evidence.

The Useful Future Is Richer Provenance

Anthropic's watermarking move is interesting because it pushes AI-generated content toward carrying more information about its own history. C2PA-style provenance for files is part of the same broader direction: instead of asking humans to infer origin from how something looks or sounds, attach machine-readable evidence to the artifact itself.

That is a meaningful improvement over unreliable stylistic guessing.

The next challenge is preventing a good provenance signal from being promoted into a universal authorship detector. Systems should preserve the distinction between what the technology can attest to and what a policy, reviewer, investigator, teacher, or reader wants to conclude from it.

A machine-readable mark can tell us something important happened in the content's history. The responsible next question is not “So who wrote it?”

It is “What exactly does this mark represent, and what decision are we allowed to make from that evidence?”

Structured Ink helps organizations design documentation and knowledge systems in which important distinctions remain explicit instead of disappearing inside convenient labels. AI provenance is another place where information typing matters: when the fields are wrong, the decisions built on them become unreliable too.

Sources

AI Governance & Knowledge SystemsContent ArchitectureDocumentation Strategy