For instance, the DALL-E 2 AI system thinks “Vicootes” means vegetables, and “Apoploe vesrreaitais” means birds. This was after the researchers prompted the AI system to create images containing text captions, then feed the resulting gibberish captions back into the system.
While the reasons are unknown, one possibility is that the AI system was trained on a wide variety of data. This includes many non-English words, thus explaining its own vocabulary. For instance, Apoploe, which seems to generate images of birds, is similar to the Latin Apodidae, the binomial name of a family of bird species.
If the secret language is true, this could have important security and interpretability implications for this kind of large AI model.
What do you think about the AI vocabulary?