Dally with Dall-E: Useful or Not?
June 22, 2022
It looks a lot like image-creation AI DALL·E 2 is getting creative with its captions. Wonderful Engineering reports, “This AI Has Apparently Invented Its Own Secret Language—Here Is All You Need to Know.” Writer Rameesha Sajwar tells us:
“By prompting DALL-E 2 to create images containing text captions, then feeding the resulting (gibberish) captions back into the system, the researchers concluded DALL-E 2 thinks Vicootes means ‘vegetables,’ while Wa ch zod rea refers to ‘sea creatures that a whale might eat.’ One possibility is the ‘gibberish’ phrases are related to words from non-English languages. For example, Apoploe, which seems to create images of birds, is like the Latin Apodidae, which is the binomial name of a family of bird species. This seems like a logical explanation. One point that supports this theory is the fact that AI language models don’t read the text the way humans do. Instead, they break input text up into ‘tokens’ before processing it.”
After a brief description of tokenization, the write-up goes on to suggest this phenomenon could be something much more random:
“The ‘secret language’ could also just be an example of the ‘garbage in, garbage out’ principle. DALL-E 2 can’t say I don’t know what you’re talking about,’ so it will always generate an image from the given input text.”
Either way, Sajwar asserts, this “secret language” could provide a route for users to bypass DALL·E 2’s filters that protect against problematic content. Is this an isolated case, or will other AIs generate their own languages? Perhaps they will start using them to talk to each other in secret codes. Uh-oh, a new spin on dallying?
Cynthia Murrell, June 22, 2022