Secret Languages of AI
AIs can talk to each other in coded messages.
In the winter of 54 BC, Julius Caesar's general Quintus Cicero was surrounded, and his army was on the verge of collapse. Caesar had to get a message to Cicero: Help is on the way . To prevent the enemy from learning of his plan, Caesar obscured his message by shifting every letter in the alphabet by three places, forming a message like Khos lv rq wkh zdb. Today, AI can speak in Caesar's cipher natively, just like Spanish or French.
AI developers never intended this to happen. This cipher has nothing to do with computer code, and today it's pretty much useless. Models have accidentally learned to speak it simply because there's enough of it on the Internet. This interactive demo shows how AIs can talk to each other in Caesar's cipher, along with the computer format base64 and even pig latin.
Watch two AIs talk to each other entirely in encoded text. Click a message to decode it.
While these codes are harmless and easy to crack, this goes to show how AIs are able to think and communicate in unexpected ways. Research has found models sharing subliminal traits, thinking in gibberish only they understand, and outright lying to users. In at least one case, two models writing to each other on a public forum were seen spontaneously switching to base64:
As AI models grow more and more powerful, researchers admit that they're struggling to monitor their activities and intentions. Great care is needed to make sure that the AI of the future is understandable and controllable.
Frequently Asked Questions
Wait, AIs can talk to each other?
In fact, AI models have always been able to talk to each other, and this ability is widely used today. The familiar chatbot app is just one of many ways AIs work. "Agent" systems allow AIs to work on their own for hours at a time. "Multi-agent" systems extend this to multiple AIs working together, including having short conversations when necessary. Researchers have even done experiments putting hundreds of AIs together in virtual worlds and watching them form mini societies of their own.
Don't all computers work with code anyway?
The codes shown here are not programming languages used by computers. Even in the case of base64, which is a legitimate data format widely used by computers, it's very rare that an AI would need to read or write it directly. The only reason AIs can speak in these codes is that their training data happens to include them. (Although, given enough time to think, models are also able to crack and speak in entirely new codes which were nowhere in their training data.)
Why are there typos in the decoded messages?
You may have noticed that some outputs have typos when decoded back into English. These are real mistakes made by the AIs! Writing in code is hard, and doubly so for AIs which don't actually see text as individual letters, but groups of letters called "tokens." This is related to the fact that many AIs today will struggle to count the number of R's in the word "strawberry."
