Google introduces About this image and a label for AI generated images

[09:32 Sun,14.May 2023 by blip]

At Google&s developer conference, a lot of things revolved around AI - besides the new language model PaLM 2, a lot of AI-supported tools were presented, which are supposed to assist with various tasks in the future - writing mails and texts, programming, editing photos, creating music, playing games, searching, translating...

Google, however, seems to be aware of its responsibility and to keep an eye on the risks associated with generative AIs (while Microsoft seems to be rash and has abdicated any responsibility to Stability AI in the form of the open source Stable Diffusion). Already, for example, manipulated photos or photos completely regenerated with relevant AI tools are circulating on the Internet, which appear to be genuine when glanced at. However, this is only the beginning: it will become increasingly difficult to recognize fake photos (and videos) as such. The social consequences are serious, because such fakes can be used to easily manipulate opinions and moods, discredit opponents, etc.

About this image

Additional information about images

.
To make it easier to identify fake images, Google will introduce a new feature in image search - "About this image" can be used to retrieve additional information that should make it easier to assess the degree of authenticity of an image. When an image was first indexed by Google and on which page, and where it also appeared, should be able to be looked up there. For example, if a photo has been exposed as fake by recognized news sites, that information will ideally show up here - for anyone who takes the trouble to look there.

Identification for AI-generated images

.
Google also commits to tagging all images created using their new generative tools with an identifier in the metadata. If an image file is used elsewhere, it should be possible to tell from the image that it was generated by AI. Whoever publishes AI-generated images as a website operator should add the metadata tag manually, as it seems. Other services such as Midjourney or Shutterstock are also to introduce this tag in the future, according to Google. An image example shows how this tag is to be included in Google Image Search: as a text note below the image.

Example of labeling in Google Image Search

In this form, a label for AI-generated images unfortunately sounds rather toothless - even if it is added automatically by major services, it is not difficult to remove it from the metadata again and feed the image to the Internet without a hint. It is also possible to install the open source generator Stable Diffusion and avoid tagging in this way. It would have been better, but also more difficult, to place it in the image via a watermark.

Content Authenticity Initiative - digital guarantee of origin

By the way, there is also an attempt from the other direction in terms of image authenticity, namely via

guarantee of origin of digitally taken photos. The Content Authenticity Initiative (CAI), initiated by Adobe, now has more than 1,000 members. The CAI aims to make every change to an image traceable by means of a cryptographically generated hash value, from the moment it is taken to the moment it is published. In addition to Nikon and Leica, Canon has also recently joined the initiative, but as far as we know, concrete implementations are still pending. Once these are up and running, it should ideally be possible to prove when, where and by whom press photos were taken.

We can only hope that one of these approaches - or other ideas - will soon provide reliable solutions for distinguishing artificially generated or manipulated images from authentic shots.

more infos at bei blog.google

deutsche Version dieser Seite: KI-generierte Bilder: Google will Authentizitäts-Check und Kennzeichnung einführen