Meta starts labeling AI-generated images across all its apps

Meta is preparing to slap warning stickers on AI images across Facebook, Instagram, and Threads, whether they came from its own tools or rival platforms.

Over the coming months, the social media giant will roll out automated labels in all supported languages for photorealistic images generated by Google, OpenAI, Microsoft, Adobe, Midjourney, and Shutterstock.

Spotting the invisible tags

Meta previously only tagged images created with its own Meta AI generator using an Imagined with AI label. To catch other synthetic media, Meta is building systems that scan for invisible identifiers aligning with C2PA and IPTC technical standards.

Audio and video, however, are a different story. Meta noted that the wider industry has not yet started embedding invisible signals into audio and video tools at the same scale. That means Meta cannot reliably detect them automatically yet.

To plug that gap, Meta is rolling out a mandatory disclosure feature. Users posting realistic synthetic video or audio will be required to disclose that the content was digitally generated. If someone fails to label it, Meta warned it may apply penalties.

Content that poses a high risk of materially deceiving the public on matters of importance could receive an even more prominent warning label.

An uphill battle against bad actors

Detecting invisible metadata is notoriously fragile because anyone can strip metadata with a quick screenshot or file conversion. Meta is openly acknowledging that reality.

To stay ahead, Meta is building automatic classifiers designed to spot AI-generated patterns even when markers are missing. Its research lab, FAIR, is also developing an invisible watermarking system called Stable Signature, which bakes the watermark directly into the image generation process so users cannot simply turn it off.

With major elections approaching worldwide, Meta is also leaning on large language models trained on its Community Standards to filter out benign content, freeing up human moderators to focus on trickier edge cases.

Why it matters

For regular users, this brings some much-needed sanity to feeds increasingly filled with lifelike fakes. For creators, though, it creates a brand new compliance headache. Anyone experimenting with synthetic audio or realistic video will now have to self-report every single time or risk platform penalties.

For the tech industry, Meta is basically setting the standard for the entire consumer internet. By pledging to scan for C2PA and IPTC tags, Meta is putting massive pressure on image generators to properly embed these invisible markers at the source. If an AI generator does not comply, its users will quickly find their content flagged or restricted.

We think relying on industry-wide metadata standards is the smartest move Meta can make right now, but it is far from bulletproof. Metadata is laughably easy to strip out, and self-disclosure relies almost entirely on the honor system. Until cryptographic watermarks like Stable Signature become universal, malicious actors will continue slipping through the cracks.