In May, Meta started labeling AI-generated images on Instagram, Facebook, and Threads. This brought a new feature to these apps, that allows the user to label an image as ‘ Made with AI’, but the feature also includes an AI-image detector that should be sensitive to industry standard AI-image indicators, and automatically label AI-generated content.
However, this feature did not work flawlessly. Photographers complained that their images were wrongly labeled as AI because of the use of standard editing tools on real photos. Also, Meta’s technology labeled AI-generated images as real, which raised concerns that this approach by Meta is not the right way to go.
As a response, Meta decided to change the label from ‘Made with AI’ to ‘AI info’, however, the technology, the detection, and labeling of AI images is still far from perfect.
A positive step towards AI-photo awareness
Flagging AI-generated content across these platforms is a positive approach, given the issues of all the potential misinformation and misuse of a person’s likeness that AI-generated images have raised.
Let’s put Meta AI-image detection to the test
AI-image generators are getting more advanced every day. We know how increasingly difficult it has become to distinguish what is real and what is AI. When it comes to the social media platforms oriented to image sharing, Meta is one of the biggest players. So let’s see how their detector fares when faced with some AI-generated images.
First, we tried labeling the image ourselves, to see what this tag looks like:
The Sony competition photo
Then we decided to upload the infamous AI-photo that won the Sony competition and sparked a controversy about AI-generated art last year. This photo is quite well known by the public. We assumed that Meta’s technology would recognize it instantly. Instagram should apply the AI label if they detect such content. Surprisingly, when we posted the photo in question, there was no label.
AI-generated photo that won the Sony competition, which Instagram did not automatically label with the ‘AI info’ tag.
‘Learn more’ section of Instagram’s ‘AI info’ label
Fake Katy Perry MET Gala photo
To be fair, this photo was trending last year, so we decided to try something more recent. Let’s see if it recognizes AI-generated Katy Perry on a MET Gala carpet.
An image imagined with MetaAI
Finally, we decided to use a photo that was made with a MetaAI image generator, and it still has a visible ‘imagined with AI’ watermark on it.
Yet, Instagram did not label it with the ‘AI info’ tag.
Can AI-generated images be successfully detected in the future?
There are already a couple of services for AI-image detection, such as WasItAI, Hive Moderation, AI or not, etc. They successfully recognize all of the photos above and have no issues with detecting and labeling them.
Introducing a label such as this one was a great move by Meta. So many users across their platforms could be exposed to unethical AI, given that both Instagram and Facebook are image-oriented platforms. Utilizing an accurate AI-image detector may prevent potential copyright infringements, legal issues, and general mistrust.