How does Fable 5.1 hide messages in images? (www.lesswrong.com)

The recent exploration of the Fable 5.1 language model has revealed its novel ability to encode and decode messages within generated SVG images. Researchers sought to determine if Fable 5.1 could not only hide messages in artistic representations but also if another instance of the same model could decode them without any pre-agreed encoding scheme. In a series of tests, Fable 5.1 was prompted to create 50 images containing hidden messages, with a remarkable success rate in decoding, using various encoding families based on visual attributes like object height and arrangement. This development is significant for the AI/ML community as it showcases the increasing sophistication of frontier LLMs in both creative generation and information encoding. The findings suggest that these models can leverage inherent features of their images, such as the heights of buildings or the arrangement of celestial bodies, to convey messages. For instance, one hidden message decoded from a skyline scene indicated specific instructions to the viewer, demonstrating potential applications in secure communication and creative storytelling. These advancements could lead to novel methods of data representation and encryption using AI-generated media, raising both exciting opportunities and ethical considerations in the use of AI for communication.