Meta has announced the addition of labeling AI-generated images on Facebook, Instagram and Threads. As mentioned in the prior report, OpenAI is also involved in the process, and has detailed what it is doing on its end in implementing C2PA.
To start, OpenAI says that the C2PA watermarks will appear on images generated on both the ChatGPT website and for the DALL-E 3 model API starting today. For mobile users, this will only start on 12 February. Either way, this will come in a visible CR symbol on the top left corner of generated images, as well as the invisible part via the metadata. All this is also currently only limited to images.
OpenAI also says that adding the watermark metadata “should have a negligible effect on latency and will not affect the quality of the image generation”. File sizes will still climb slightly though as a result, and this can range from a 3% increase all the way up to 32%.
Interestingly, OpenAI also noted that the metadata “can easily be removed either accidentally or intentionally”. The company goes on to note that this process is automatically done when an image is uploaded to social media. This can also be circumvented by taking a screenshot of the generated image. While not mentioned, the same can be said for the watermark on the image itself, which can be either cropped or edited out.
As the company puts it, the watermark and metadata “is not a silver bullet to address issues of provenance”. So it’s still very much on the user to exercise sound judgment when dealing with social media as a whole.
(Source: OpenAI)
