According to a Researcher, an AI Has Created Its Own Secret Language

It seems that when researchers use machine-learning AI, they often don’t understand how exactly their algorithms solve the problems they are given to work on, and some even go above their intended purpose. For instance, Facebook’s AI began to develop its own language before being shut down, and now the text-to-image generator DALLE-2 might be doing something similar.

Giannis Daras Noticed DALLE-2’s AI Using Its Own Language

Giannis Daras, who is a Computer Science Ph.D. student, was the one who noticed that DALLE-2’s AI was doing something strange but not entirely random. The system usually creates images based on a text input prompt, and under certain circumstances, it would return nonsense words instead of images. According to Daras, DALLE-2 is known for its limitation when it comes to generating images with text. Apparently, some text prompts often lead to generated images that show gibberish text.

The DALLE-2’s AI Produces Text That Seems Like Gibberish But Isn’t Random

Daras has stated that he discovered that the text DALLE-2’s AI produces in such situations was not random but revealed a hidden vocabulary that it had developed internally. One hint for the researcher was that when fed with the seemingly gibberish text, the model would frequently produce pictures related to the original input search text.

Daras has also explained that when the AI was asked to subtitle a conversation between farmers, it would show them talking with speech bubbles filled with words that look like complete nonsense. So, when Daras fed the nonsense words back into the system, the AI had assigned its own meanings to them and showed how farmers were talking about birds and vegetables. If Daras is correct about his assumptions, this could have security implications for DALLE-2.

Other AI Algorithms Have Been Known to Create Their Own Languages

DALLE-2 is far from being the first AI that has been shown to create its own secret language. However, the claims about it have not yet been peer-reviewed, and there are researchers who are currently questioning Darras’ claims. For one, research analyst Benjamin Hilton did some experiments of his own and found that results were not always conclusive and were sometimes due to pure luck. Also, adding more words to the phrases gave different results, showing that they don’t always mean the same thing.

So far, it seems like the AI language is more like noise in some cases and nothing at all in others. Further research and peer-review of the paper will show more, but there could actually be something going on with the AI’s secret language. Enthusiasts will have to wait and see.