Watermarking Applied to All AI Creations

Watermarks are starting to be applied to all AI-generated works. This is a measure to identify AI works, and its characteristic is that watermarks are applied not only to images and videos but also to text. Major AI services such as Claude, GPT, and Gemini are subject to this watermarking.

It is also being sequentially applied to previous models, and watermarks are generated at the model output unit, applying uniformly across all services, including API, apps, and web.

This method leaves identification marks not only in the corner of images but also in text, inserting a specific identification ID within the generated text. A characteristic is that watermarking is maintained even when text generated in Claude is transferred to Gemini.

Currently, it is described as a target phase to expand the scope of application, and direct watermarks do not appear in text generated by GPT.

This involves intentionally changing the probability of certain words appearing when AI writes text. By analyzing the patterns of these selected words, it is determined whether the text was generated by AI, and the accuracy of determination increases as repeated probability patterns accumulate.

This is because strong opposition from developers would be expected if the code were to be broken. Instead, the method of inserting watermarking information through the code's comment areas is used. This measure maintains the logical integrity of the code while enabling the identification of whether it was AI-generated. Watermarks are not inserted in parts requiring exact values or grammar, but they can be applied in areas like comments where content can be explained with various expressions.

This regulation mandates machine-readable markings for AI-generated text, images, audio, and video. Watermark technology is designed in a non-destructive way that does not damage the original form of the content.

Due to regional infrastructure management cost issues, it will be applied uniformly and simultaneously to AI services worldwide.

Under EU regulations, source code and general editing assistance functions are exempt from watermarking obligations. However, depending on company policy, watermarks may be applied to a broader range than these minimum regulatory standards. This is an area that can be decided by the autonomous judgment of each AI service provider.

One way to bypass or weaken AI watermarks is a strategy of having another AI rewrite content already generated by an AI. This method maintains the original text's meaning while creating new sentence structures and expressions. While simply converting file formats makes it difficult to remove watermarks, using a different AI model to break the existing AI's word combination patterns can weaken the watermark signal strength.

It is necessary to rewrite the entire sentence structure, not just correct typos, to remove the watermark signal. After generating the output, it is essential to compare it with the original text to ensure that the core content has not been altered.