# A Free Tool Now Strips AI Watermarks From Claude, OpenAI and Gemini Text > Source: > Published: 2026-08-12 12:36:51+00:00 *Anthropic, OpenAI and Google are pushing AI provenance into the mainstream just as free tools show how fragile those signals can be once ordinary users start testing them.* Claude's new watermarking system arrived with the weight of regulation behind it. According to The Verge, Anthropic is adding invisible, machine-readable marks to text generated by Claude models launched in the European Union after August 2, 2026, and C2PA metadata to generated images. The point is simple: if a machine made it, people should have a way to tell. That sounds sensible. It also runs straight into the internet you actually use. Free watermark-removal tools are already advertising ways to strip AI provenance signals from images and metadata - sometimes from text too. Some are browser tools. Some are GitHub projects. Some make bigger claims than they can prove. But the pattern is clear enough: the moment AI companies add a signal, somebody else tries to remove it. Anthropic's move is tied to Article 50 of the EU AI Act, which applies from August 2, 2026. The European Commission's own guidance says providers must mark synthetic content in a machine-readable way and make it detectable, unless the system is only doing standard editing or leaves the input largely unchanged. Fines can reach €15 million or 3% of total worldwide annual turnover for companies. That's real pressure, not a public-relations exercise. Anthropic is applying the mark globally on covered Claude models, not just inside the EU. For images, The Verge reported that the company will use digitally signed C2PA metadata. For text, Claude will use an unspecified imperceptible watermark built into the output. Anthropic has also said the marks are meant to survive common actions such as copying and pasting, or light editing. Here is the catch. Metadata is easy to lose. ## The strongest systems still have weak edges Google has gone further than anyone in making watermarking feel like infrastructure. In a May 2026 blog post, Google said SynthID had already watermarked more than 100 billion images and videos and 60,000 years of audio. It also said SynthID verification in Gemini had been used 50 million times globally. Those are not trial numbers. They are platform numbers. Google is pushing the system across Search, Gemini, Chrome, Pixel and Cloud. OpenAI is using C2PA Content Credentials and SynthID watermarks for images generated through ChatGPT and Codex, and through its API as well, according to OpenAI's own help center and its May 2026 provenance update. The companies want the same thing: a signal that travels with AI media after it leaves the model. But OpenAI is unusually plain about the limits. Its help center says an image could still have come from OpenAI even when no signal is found, because metadata may have been stripped or the watermark may have degraded. The file could also have come from an unsupported source. That's the honest sentence. Keep it in view. Projects such as remove-ai-watermarks on GitHub focus on stripping visible and invisible AI watermarks and provenance metadata from images. GPT CLEAN UP advertises tools for checking and cleaning hidden AI text markers and Gemini image metadata. These tools are not all equal, and their claims need caution. A tool can delete EXIF or XMP data, even C2PA metadata, without proving it has broken a deep pixel watermark like SynthID. Still, for many real uses, stripping the easy layers is enough to cause trouble. If you run a publisher, a marketplace or a school policy on watermark detection alone, you're building on a floor with gaps in it. ## Regulators want labels, users want workarounds Watermarking was never a lock. Google DeepMind's own SynthID text explainer says watermarking can help identify generated text, but it is not built to stop motivated adversaries. OpenAI has made the same point in different language: no detection method is foolproof. That's not weakness from the labs. It's reality. The harder question is what platforms do with that reality. A watermark can catch casual misuse, the copied essay, the untouched image, the spam post generated and uploaded in one move. It can also help honest users disclose AI work without adding a label by hand every time. That has value. It won't stop someone who is trying to hide. Frankly, that is where the policy conversation gets sloppy. The EU rule asks providers to make AI output marked and detectable as far as technically feasible. It doesn't say the mark has to survive every paraphrase, screenshot, export, file conversion and cleanup tool. Platforms may still act as if a missing mark proves a human made the content. It doesn't. It proves only that the detector found nothing. The useful answer is layered. Watermarks help. C2PA helps. Human review, account history, upload behavior and old-fashioned editorial judgment still matter. If a publisher rejects every submission with a watermark and trusts every one without it, the system will punish honest disclosure and reward basic laundering. That is backwards. Anthropic's Claude watermark is still an important step, because the big labs are finally treating provenance as part of the product rather than an afterthought. But the first week of serious attention around these tools shows the next phase clearly. AI companies will keep marking content. Users will keep trying to clean it. Anyone making decisions from those signals needs to remember what a watermark is: evidence, not proof. **Also read:** [Tencent's Ad Business Booms on AI But Profit Falls Short of Estimates](https://startupfortune.com/tencents-ad-business-booms-on-ai-but-profit-falls-short-of-estimates/) • [How Does AI Agent Memory Actually Work, And Why Investors Are Betting Big On It](https://startupfortune.com/how-does-ai-agent-memory-actually-work-and-why-investors-are-betting-big-on-it/) • [How AI Browser Agents Are Changing What Startups Build Instead of Websites](https://startupfortune.com/how-ai-browser-agents-are-changing-what-startups-build-instead-of-websites/)