How difficult is it to identify whether an image is an AI-generated fake? Amazon claims to have a solution for this with its newly released AI art generator, Titan, which has its own watermarking system. Of course, this is not simple. Every company using this model must explain to users how to check if the images they find online are truly another AI-generated fake.

Amazon has not provided any details about the capabilities or training data of Titan. Swami Sivasubramanian, Vice President of Machine Learning at Amazon Web Services (AWS), told attendees at the company's re:Invent conference that it can create a "realistic" image of a monitor lizard, change the background, and use a generative filling tool to expand the image's borders. However, this model is limited to AWS customers using the Bedrock platform. It is a foundational model for companies to integrate into their own platforms.

Like many other cloud service giants, Amazon has delegated the responsibility of identifying AI-generated images to individual users. The company states that this new AI image generator will add an "invisible watermark" to every image generated by this model. This is "to help reduce the spread of misinformation and provide a cautious mechanism for identifying AI-generated images."

It is currently unclear what this watermark looks like or how external users can identify such images. Other companies, such as Google DeepMind, claim to have found a method to disrupt the pixels of AI-generated images, creating an immutable watermark.

Vasi Philomin, Vice President of Generative AI at AWS, told The Verge that this watermark does not affect the quality of the image but cannot be "cropped or compressed." It is also more than just a metadata tag. Adobe uses metadata tags to identify whether images generated by its Firefly model are AI-generated. Users are required to visit a separate website to determine if an image contains a metadata tag.

To determine if an image is AI-generated, users need to connect to a separate API. Each company using this model needs to inform users how to access that AI scanning technology.

In short, identifying which images are AI-generated will be very challenging, especially considering that each company has created its own separate watermarking system for its AI.