Dewatermark inspects AI text for hidden Unicode artifacts
TL;DR. A new open-source tool, dewatermark, helps users detect and remove hidden Unicode artifacts from AI-generated text locally in their browser. - The tool identifies characters invisible to the human eye, potentially used for covert watermarking or data transmission. - It offers a free, private browser-based inspector and a Python package with forensic reporting capabilities. - The project prioritizes user privacy, ensuring that no text leaves the user's device during inspection or cleanup.
- Dewatermark, an open-source project, identifies hidden Unicode within AI-generated text.
- The tool operates locally in the browser, preventing user data from being sent to external servers.
- It distinguishes between deterministic Unicode artifacts and undisclosed vendor statistical watermarks.
- A Python package extends functionality to include forensic reports and repository scanning for developers.
Sources
- A local-first scanner for hidden Unicode in AI-generated text — cyzanfar.github.io