Brian Roemmele (@BrianRoemmele), the founder of technology research site ReadMultiplex, published a Mac-only tool and local rewriting workflow on August 12th that he says can remove the new machine-readable watermarks Anthropic is building into Claude-generated text.
Roemmele introduced the tool in a post on X, describing Anthropic's marks as hidden fingerprints attached to material produced from training on the open web. His accompanying ReadMultiplex article calls the software AI Watermarks Cleaner and pairs it with LM Studio, which runs language models locally.
The workflow appears to rely on local rewriting rather than the simple deletion of metadata or invisible Unicode characters. Roemmele says the process separates visible formatting signals from statistical patterns embedded in generated language, then uses a local model to rewrite the material. He describes the current release as Mac-only and says Windows and Linux versions are planned. The instructions and download are offered through ReadMultiplex's paid membership.
The launch follows Anthropic's disclosure that Claude models launched in the European Union on or after August 2nd, 2026 will mark generated text at launch. Anthropic says the same marks will travel worldwide across supported models used through Claude, Claude Code, Claude Cowork, Claude Tag and the Claude API, including access through Amazon Web Services, Google Cloud and Microsoft Foundry.
Anthropic is using two systems. Text receives an "imperceptible" watermark applied during generation, while supported files such as PNG, JPG and SVG assets receive signed provenance metadata based on the C2PA standard. Anthropic says the text signal can survive copying, pasting and some editing because it is woven into the generated language at the model level.
That description also defines the central limit of Roemmele's cleaner. Anthropic has not yet released the technical specification or public detector for its text marks. A tool can rewrite Claude output in ways likely to weaken a watermark, but its success cannot be measured against Anthropic's implementation until the mark can be independently detected before and after processing.
Anthropic already acknowledges that heavy editing, paraphrasing, translation or combining a passage with other writing can make a mark undetectable. Short passages may not contain enough information to produce a reliable signal. File provenance can also disappear when someone converts or re-saves a file, or captures it in a screenshot.
Roemmele's local rewriting approach is therefore directed at a weakness Anthropic has documented itself: provenance signals lose reliability as content moves through human and machine editing. The practical question is how extensively the cleaner changes a passage, and whether it preserves facts, quotations and authorial intent while disrupting the statistical signal. ReadMultiplex has not supplied independently reproducible results against an Anthropic detector.
A watermark is not a user identifier
Roemmele goes further than Anthropic's public documentation when he describes the marks as fingerprints capable of tracking AI output back to a user or prompt. Anthropic says its planned detector will indicate that material may have been processed by Claude. It explicitly warns that a positive result will not establish the original author or the full history of the content.
A document could carry a Claude mark after the model merely proofread, translated, summarized or reformatted human-written material. That creates a meaningful attribution problem for publishers, developers and companies using Claude inside production systems: a machine-readable label may correctly detect Claude's involvement while giving readers the wrong impression about who produced the underlying work.
The policy originates with Article 50 of the EU AI Act, which requires providers of generative systems to make synthetic text, audio, image and video output machine-readable and detectable as AI-generated or manipulated. The law names watermarks, metadata, cryptographic provenance and fingerprints as possible methods. It also exempts systems performing standard editing or making changes that do not substantially alter the user's input or its meaning.
Anthropic chose to apply its marking system worldwide rather than maintain separate output behavior for Europe. That decision gives a European transparency rule direct consequences for Claude users and API customers in the United States.
Roemmele's release opens the next phase of the compliance fight before Anthropic has finished documenting its system. AI providers need marks that survive ordinary editing, while customers have commercial, privacy and editorial reasons to control what remains attached to their work. Local rewriting tools will test whether text watermarking can satisfy both sides, or whether the requirement produces a continuous contest between provenance systems and software designed to erase their signals.