Watermarking has been proposed as a potential solution to distinguish AI-generated content from human-generated content. However, implementing effective watermarks that are resistant to removal and manipulation poses significant challenges.
One common type of watermark for digital images is the overlay of text by stock image sites. While this type of watermark is visible and difficult to remove without photo editing skills, it is not suitable for AI-generated images, as it renders the image useless for publication.
Another type of watermark is metadata, which is unobtrusive but easily removed from files. Metadata can provide information such as the date, time, location, and creator of an image. However, social media sites often automatically remove metadata when images are uploaded to save storage space.
A useful watermark for AI-generated images would need to be detectable even after cropping, rotating, or editing. One technique is manipulating the least perceptible bits of an image, creating a pattern that is invisible to viewers but detectable by a watermark-detecting program. However, this method is easily destroyed by rotating or resizing the image.
More sophisticated watermarking proposals exist that are robust against a wider range of edits. However, these proposals must also be robust against individuals who are aware of the watermark and actively seek to eliminate it. By directly manipulating the image file, an individual could remove the watermark without significant visual changes.
Alternatively, some companies are exploring content authenticity, adding metadata to camera-generated images and using cryptographic signatures to prove their genuineness. While this approach solves the issue of watermark removal, it is a complex scheme that relies on the preservation of verifiability through all software used to edit photos. Additionally, most cameras do not produce this metadata, limiting its effectiveness in proving the authenticity of images.
Watermarking text-based generative AI poses even greater challenges. While word-based watermarks could be created to establish textual style, they are susceptible to removal through rewording. Furthermore, the availability of detection tools raises concerns. Publicly available tools allow for repeated edits to bypass detection, while keeping them secret limits their usefulness and requires reliance on the watermarking company for detection.
In conclusion, while watermarking and content authenticity are potential approaches, they both have limitations and are not comprehensive solutions. Most images and text on the internet will not have any form of watermark or content authenticity metadata.
The source of the article is from the blog oinegro.com.br