Meta AI Image Warnings Backfire

Facebook and Instagram's attempts to avoid users being misled by AI-generated photos have backfired. Several users report photographs they took themselves as being falsely labeled as made by AI tools.

Meta, which owns both sites plus social discussion tool Threads, has started adding a "Made with AI" label to images. In some cases it's there because the users has ticked a box to say they used AI to make the image. In other cases, it's because Meta's detection tools have flagged the image.

The site has felt pressure to clearly indicate AI-generated images that could be designed to fool and mislead people. In particular, it has explored the potential for misleading images portraying events that didn't happen being used to influence voters or sow distrust during election campaigns.

It doesn't seem that Meta yet has the ability to reliably detect an AI-generated image simply from its appearance, for example an unnatural look or unlikely content. Instead it's relying on metadata that's produced by a range of AI generation tools.

Adobe Cropping May Trigger Warning

Wrongly labeled images include an official photo of the Kolkata Knight Riders cricket team celebrating a win and a basketball shot taken by professional photographer Pete Souza.

Mr. Souza believes the mistake is likely because he used Adobe software to crop the image and that this was wrongly detected as using an Adobe AI tool. (Source: theverge.com)

Not a Binary Issue

Meta hasn't commented specifically on these claims but told TechCrunch it was reviewing its policy. It said "Our intent has always been to help people know when they see content that has been made with AI. We are taking into account recent feedback and continue to evaluate our approach so that our labels reflect the amount of AI used in an image." (Source: techcrunch.com)

Part of the confusion and problems for Meta seems to be distinguishing between three types of image: those where the content is completely "real" and the only change is cropping or adjusting characteristics such as color balance; those which started as a genuine image but have used AI to remove content (such as strangers in the background of a group photo); and those which are not true photographs and instead were created from scratch by AI tools.

What's Your Opinion?

Should Meta label AI images at all? Do you believe it is possible to accurately detect such images? Should the rules be limited to certain circumstances such as images involving public figures or election candidates?