Claude Watermarks More Text. Your CMS Still Needs a Revision Trail.

A pricing page goes live with a sentence that promises next-day delivery. The product owner says the draft they reviewed promised two days. At that point, the useful question is which edit introduced the change and how it made it to the live page. A text watermark can suggest that a model was involved. To answer the product owner, you need the versions in between.

AI generated and fact-checked this article against the public sources linked below. It was published through an automated workflow, and Keith has not reviewed it.

Anthropic’s November 3 release notes say that Claude Mythos 5, Claude Opus 4.7, and Claude Sonnet 4.6 now carry its text watermark on supported platforms. Anthropic lists coverage by model and platform, with the Amazon Bedrock rollout for those three models expected to complete by November 10. Its detector is in private preview for eligible organizations. That makes it a separate capability from receiving marked output. Claude Platform release notes · Anthropic’s coverage and detection guide

Anthropic describes a statistical pattern in the model’s choice of words. A detection result can indicate that Claude generated or processed some of the text. It can leave open whether the model drafted an entire section, translated someone’s work, or edited existing copy. A short passage may offer too little evidence for a reliable result, and coverage also depends on the model, platform, and, for Sonnet 4.6, sampling settings. An editor should keep an inconclusive result inconclusive. Anthropic’s watermark explanation · Claude API sampling caveat

Keep the proposal connected to the publication

Okeike’s public case study describes a working editor in ongoing development. It turns AI suggestions into inspectable patch plans for a selected page, uses source fingerprints to recognize when the file behind a draft has changed, and publishes through a GitHub commit and an existing Firebase Hosting pipeline. Those are useful building blocks for revision history. The record design below is a proposed extension to that workflow. Okeike Web Editor case study

Consider the delivery sentence in the opening example. Store the proposal’s ID, the base revision, and the page field it targets. If the proposal came through an integrated AI tool, record that event and the model identifier when the service supplies one. If the editor types new wording afterward, record the saved edit as its own revision. A commit can then point back to the final draft and its history, so someone investigating a correction can open the right diff instead of guessing from the live text.

Give imported copy an honest state. A paragraph pasted from another tool can arrive without a recorded origin. The editor can accept an explicit disclosure from the person submitting it and preserve that statement as such. Keep it separate from model use observed by the application. An empty integration history gives the editor little information about what happened outside the editor.

Rehearse a correction

Use a fictional page and run a short exercise. Apply a generated patch to the delivery sentence, edit it by hand, and publish the resulting revision. Then ask someone who did not see those steps to find the exact wording in the live commit and locate the change that introduced it. Repeat with a pasted paragraph whose origin is unknown. The history should tell the operator where the wording entered the workflow and where the evidence ends.

There is a cost to keeping that record. Storing every prompt, raw response, and clipboard payload can turn a history feature into a second store of sensitive material. A smaller record can link the existing versions and record the event type, target field, and actor with access limited to the people handling reviews and corrections. If a team also uses a detector, record its result separately from those observed events. The two kinds of evidence answer different questions.

When a page needs a correction, the editor should be able to find the change, compare it with what was reviewed, and publish a fix. That is a practical test of provenance for any team that ships copy through an AI-assisted editor.