Anthropic says it will watermark text generated by its AI models

The compliance response has begun. Last week’s discussion about Article 50 of the EU AI Act was largely about the rule arriving. This week, the operational consequence is visible: Anthropic says Claude generated text will carry an invisible, machine readable watermark, while supported files will receive signed provenance metadata. The change applies to new models launched in the EU from 2 August 2026 and is intended to operate worldwide across supported Claude products.

For project delivery organisations, this is more than a product feature. It changes the status of AI assisted work inside the document chain. A draft method statement, tender clarification, risk register, progress report or board paper may soon carry a signal that survives copying and some editing. That signal will not establish who wrote the underlying idea, whether the content is accurate or whether a human reviewed it. It will establish that Claude processed the material. Those are different questions, and governance teams need to keep them separate.

The practical issue is therefore not whether AI can be detected. It is whether a business can explain how AI was used, who reviewed the output, which source material informed it and what level of reliance is appropriate. That is a much more useful starting point for delivery professionals than a simple human versus machine argument.

The mark is designed to travel with the text

Anthropic’s support guidance describes two mechanisms. Text generated by supported Claude models will carry an embedded watermark. Files such as images can carry digitally signed provenance metadata using the Coalition for Content Provenance and Authenticity standard, commonly known as C2PA.

Anthropic describes the text mark as imperceptible and says it should not change the meaning, quality or readability of the response. In its own words, the system “weaves an imperceptible watermark directly into the text itself” without changing the meaning or readability of Claude’s response. Anthropic says the signal is designed to travel with copied text and may survive some editing, which is why the feature matters for documents that move between project teams and platforms. The European Commission places that technical change inside a wider policy objective: Article 50 addresses “risks of deception and manipulation” and is intended to foster the integrity of the information ecosystem.

The company also says the watermark is applied at model level, rather than being tied to one interface. That matters for organisations using Claude through an API, a desktop application, Claude Code, Claude Cowork or a cloud partner. A procurement team cannot assume that the same underlying model becomes unmarked simply because the user accesses it through a different route.

The stated goal is transparency, yet the policy debate includes competing views about how visible and useful these signals will be.

❝

Sergey Lagodinsky, the Green MEP who helped negotiate the AI Act, told The Guardian: “It is a matter not only of customer protection, it’s also a matter of democracy protection.”

Boniface de Champris, AI policy lead at CCIA Europe, offered a different warning, saying: “The label was meant to flag deceptive content.” His concern was that if labels become ubiquitous, users may stop noticing them. Both perspectives matter to project organisations. A provenance signal should support informed judgement, not become a substitute for it.

Provenance is useful, but it is not authorship

The most important nuance is also the easiest to lose in a headline. A watermark is evidence that content was processed by a model. It is not a complete history of the document.

Anthropic explicitly says a detected mark does not prove that Claude was the original author. Someone may have used the system to translate, summarise, proofread or convert existing material. Equally, the absence of a detectable mark does not prove that generative AI was not involved. Heavy editing, paraphrasing, translation, short passages, format conversion or stripped metadata can all weaken the signal.

This distinction should shape project documentation. A mark can answer the question, “Was this content processed by Claude?” It cannot answer, “Who was responsible for the technical judgement?” The second question remains a matter of professional accountability, competence and review.

In an AEC setting, consider a design risk note prepared from a consultant’s technical memorandum. Claude may help structure the prose, compare options or extract actions. The source material remains the consultant’s. The final judgement belongs to the named professional who approves the note. If the document later carries a watermark, that signal should prompt a review of the record, not an automatic conclusion that the work is unreliable.

We should also avoid treating provenance metadata as a substitute for document control. C2PA can help preserve information about how a file was created or processed, yet metadata can be removed accidentally when files are converted, resaved or uploaded to another platform. A robust project record still needs version history, approval status, source references, named reviewers and a clear record of changes.

The document trail is becoming part of the deliverable

For years, many businesses treated AI use as a private productivity choice. The output was judged by whether it looked acceptable. That approach becomes fragile when suppliers, clients, regulators and insurers ask how a document was produced.

Project teams already understand the value of a controlled information environment. Drawings, schedules, specifications and cost plans move through status codes, review gates and approval workflows. AI assisted text should be handled with a similar discipline. The new signals from Anthropic make that discipline more visible, yet the underlying principle is familiar: the reliability of a deliverable depends on its source, review and decision history.

A sensible AI assisted document record should answer five questions:

• Which system or model was used?

• What material was supplied to it?

• What did the system do, such as summarise, draft, classify or transform?

• Who checked the output and against which source?

• Who accepted responsibility for the final document?

That information need not appear on every client facing page. It should exist in the project record and be available when a decision, dispute or audit requires it. The aim is proportionality. A low risk internal meeting summary does not need the same treatment as a contract notice, life safety assessment or planning submission.

Procurement teams need a position on disclosure

The issue will quickly move into supplier management. Clients will ask consultants and contractors whether AI is used in the production of reports. Main contractors will ask design teams how confidential information is handled. Professional indemnity insurers may want to understand whether AI systems are embedded in high consequence workflows. Employers will need a consistent response rather than a collection of personal preferences.

The first step is to classify deliverables by consequence. A useful internal framework could distinguish routine administrative content, decision support content and formally issued professional content. The higher the consequence, the stronger the expectations around source retention, human review and disclosure.
The second step is to specify permitted use in appointments and project protocols. A clause that simply says “AI may not be used” is unlikely to reflect actual practice. A better clause would define which systems are approved, which information may be entered, which outputs require human review and how the use of AI is recorded.
The third step is to maintain a supplier question set. It should test whether a firm can identify model use, protect client information, preserve source material, manage hallucinated content and provide a human sign off. These questions are more informative than asking whether an organisation has an AI policy. Many policies describe intent without showing how the rule operates inside a live project.

Transparency will change the conversation with staff

Watermarking may create anxiety among employees who use AI to improve their writing. The answer should be clear. AI assistance does not automatically mean that a person has surrendered authorship or professional responsibility. A person may use a tool to improve structure, correct grammar or summarise a long record. The meaningful issue is whether the final content has been checked and whether the use is appropriate for the task.

That is why internal guidance should distinguish assistance from delegation. Assistance supports a person who remains responsible for the reasoning. Delegation asks a system to perform a task whose output may be accepted with limited review. The second category deserves tighter controls, especially where the system can act on live data or make changes in connected applications.

The language used by managers matters. If staff believe that disclosure will be treated as misconduct, they may hide legitimate use. If the organisation treats every output as automatically trustworthy, staff may over rely on it. A mature policy creates a safe route for disclosure, makes review expectations explicit and reserves sanctions for misuse, concealment or reckless reliance.

Three actions for project leaders this month

The first action is to map AI assisted deliverables across a live project. Do not begin with tools. Begin with documents and decisions. Identify where AI is already used, what information enters the workflow and where a weak output could affect cost, programme, quality, safety or reputation.
The second action is to create a lightweight provenance log. It can sit alongside the project information management system and record the model, purpose, reviewer and approval status. This log should support accountability without forcing teams to document every spelling correction.
The third action is to test the policy against a realistic scenario. Ask a project manager to prepare a client update using a confidential progress report, a change log and meeting notes. Then check whether the team knows what can be uploaded, how the draft is verified and what is disclosed when the update is issued.

Takeaway

• Treat watermarks as a provenance signal, not as proof of authorship, accuracy or full document history.

• Introduce proportional AI disclosure rules for project deliverables, with stronger controls for contractual, safety critical and professional decisions.

• Add model use, source material, reviewer and approval status to the project record for consequential documents.

• Ask suppliers how they protect confidential information and preserve human accountability when AI is used.

• Test the policy in a live project scenario before a client, regulator or insurer tests it for you.

Project Flux examines the practical consequences of shifts like this, from document control to supplier policy, for the people responsible for delivering complex work. Subscribe to our newsletter, and stay abreast with the lastest updates in the world of AI and ConTech.