Watermarks Remover has become publicly available — a tool that attempts to remove traces of AI-created or processed content from texts, images, and documents.
Watermarks Remover has become publicly available — a tool that attempts to remove traces of AI-created or processed content from texts, images, and documents.
The tool was published by user Guillaume Meyer.
The simplest method of marking can be unusual spaces or Unicode characters that a person cannot see on the screen, but which remain inside the text. Watermarks Remover finds and removes them. Another type of «traces» can be hidden not in the text or image itself, but inside the file. For example, PNG or JPEG files can contain metadata with information about the origin of the content and its editing. Watermarks Remover can clean such data from PNG, JPEG, WebP, SVG, PDF, DOCX and other formats.
But when the watermark is not hidden in invisible characters or metadata, watermark removal becomes more difficult.
Today dev.ua wrote that the company Anthropic said that it will use exactly this method of marking. So, Claude does not insert a secret symbol or metadata into the text. Instead, the model slightly changes which words it chooses during generation. For the reader, the text looks normal. But in a large fragment, a statistical pattern appears, which a special detector can recognize by the secret key.
For example, Claude writes: «Today it was cold and…» You can continue with the word «cloudy» or the word «cloudy» — both options sound natural, so it is not so important for the content which one the model chooses. Usually in such situations, the choice is partially determined by randomness. But now the conditional randomizer starts working according to the rules set by the secret key. So in the mentioned scenario, for example, the key can set the probability of «dropping» the word «cloudy» to 85%. If there are many such matches in the text, it will be marked as «processed by AI».
Therefore, Watermarks Remover offers to rewrite the text using local LLM. It is assumed that if the marking consists of which words the initial model chose one after another, then after a sufficiently strong rephrase this sequence will change. This can break the statistical pattern. However, the developer does not guarantee bypassing the detectors. After all, without the detector itself and the secret key, it is impossible to check whether the mark has really disappeared.
Anthropic itself also notes that its watermark is not invulnerable, as strong editing can remove it.
Previously, dev.ua wrote that OpenAI has launched a tool for identifying images created by DALL-E 3.
