If we talk about AI models for image generation, DALL-E is probably the best known of those currently in existence, for several reasons. Recall that it was one of the first to jump to popularity, and also that it was the work of OpenAI, an artificial intelligence for generating images that Microsoft integrated into its AI-based services such as Bing Image Creator and Copilot. And if that wasn’t enough, the version 3 model introduced last September has undoubtedly raised the bar a lot for what these types of solutions can do.

DALL-E, like its main competitors, is not perfect. Artificial intelligence still makes mistakesbut if we analyze its evolution since, for example, the test we carried out in October 2022 (surely many of you still have nightmares with the faces of horror, right?), it seems that the evolution experienced is more than remarkable, although from here I appeal to my colleague J .Pomeyrola to repeat the test with the current version of the model so that I can evaluate it in the first person.

Since then, we’ve been able to see AI-generated images take off, and In more than one case, attempts have been made to “stretch” them as realwhich is one of the main concerns that many of us have expressed with the proliferation of this technology is that the disinformation and manipulation ecosystem has suddenly found itself with an incredibly powerful arsenal.

So there has been talk for some time about the need to use some type of watermark to clearly identify AI-generated content, and we had good news about it last October during Adobe Max 2023, when Content Credentials, C2PA, was introduceda system that allows recording both the origin and the modifications made to the image if the AI ​​model intervened in part of the process.

Already at that time, Adobe confirmed that Microsoft was joining the initiative, which we understood that OpenAI would do as well. And we are not wrong because, as we can read on their website, OpenAI has just announced that DALL-E has already started implementing C2PA and subsequently, the images created by this model have already started to contain this digital watermark, which, as you surely remember (or imagined, if you didn’t already know), is based on the inclusion of metadata in the resulting file.

All parties involved recognize that this solution is not perfect. OpenAI itself acknowledges on its page about the C2PA implementation in DALL-E that metadata can easily be removed from a file, both intentionally and by mistake, but at least it represents a first step in finding a solution to a problem that is already a problem and which wants to be an even bigger problem in the future.