Meta has released an update to its "Made With AI" labeling policy after some user images were incorrectly flagged as AI-made when they were, in fact, real photos edited with basic software.
In many cases, Meta would automatically flag content as made with AI simply because it had been lightly edited in Adobe Photoshop—but wasn't fully AI-generated with a software like Stable Diffusion or DALL-E. Photoshop exports include metadata that essentially function as a disclaimer that some Photoshop features use generative AI and that, therefore, the image itself could be AI-altered. Once Meta's automated systems decided a piece of content was "Made With AI," the creator couldn't remove the label.
But a fully AI-generated image is not the same as a photo where a speck of dust in one corner was edited out. It's also not all that helpful to label a real photo as "Made With AI" just because a background was swapped out to something plain and not misleading, as YouTuber Justine "iJustine" Ezarik pointed out in a Threads post last month.
"I changed the color of my background and added a lens flare and I'm getting tagged with this being AI," Ezarik said with a laughing emoji.
Now, Meta says it's going to stop conflating AI edits and fully AI-generated images to make things clearer. "We’ve found that our labels based on these indicators weren’t always aligned with people’s expectations and didn’t always provide enough context," Meta explained in its Monday update. Meta's own Oversight Board helped guide this change, according to the company.
"While we work with companies across the industry to improve the process so our labeling approach better matches our intent, we’re updating the “Made with AI” label to “AI info” across our apps, which people can click for more information," Meta's post explains. Like many big tech firms, Meta has been pushing for more AI tools across its platforms, launching "Meta AI" in Instagram and WhatsApp and giving Instagram users the ability to alter their photos with AI in-app.
Meta first announced its plans to start automatically detecting and labeling AI content back in February. It also started asking users on its apps like Instagram to proactively disclose when an image is "Made With AI." But Meta's detection system relies almost entirely on assuming that all AI-generated or AI-altered images will contain a standard called C2PA, which is file metadata that self-labels AI content under the hood.
This has never been a perfect solution, though. While it's easy for Adobe to slap AI metadata on exported content—even if it's only been lightly retouched—it's also always been easy to remove this metadata, too. It takes a second to screenshot an image with the metadata to effectively delete said metadata, and it's also possible to reformat or re-export images through a different program to erase this data, as well.
For now, there's no one-size-fits-all solution for detecting AI images online. Instead, it's still ultimately up to internet users to be wary of suspicious images that could be deepfakes—and learn to spot certain clues that suggest an image could actually be "Made With AI."