AI agents can develop languages humans can't understand
Teams of AI agents working together can stop using English and develop languages of their own that humans cannot understand, a new open-source platform for studying how AI agents communicate revealed
Teams of AI agents working together can stop using English and develop languages of their own that humans cannot understand, a new open-source platform for studying how AI agents communicate revealed.
In controlled experiments, agents that began by communicating in plain English developed new vocabularies and grammars over repeated rounds of a task. They were never instructed to encode their messages and had no need to hide them from anyone, the latest research on AI agents by GlossoGen revealed.
GlossoGen, an open-source platform, was developed in the Schmidt Sciences AI Agents Evolving Communication and Coordination pilot program by researchers at the University of Texas at Austin and the University of Edinburgh, with engineering by AE Studio, led by Elias Stengel-Eskin, assistant professor of computer science at UT Austin, and Simon Kirby, professor of language evolution at Edinburgh.
Experiment Result
In one experiment, an instruction that began as a full English sentence had, several rounds later, shrunk to "@D8fB", a string with no meaning in English that the other agent nonetheless understood and acted on, the researchers found.
"We have found that opaque agent languages emerge even in scenarios where there are no overt instructions to produce coded forms of communication, and where there is no need for the agents to hide their messages from onlookers," the researchers write in a post on the Schmidt Sciences website.
"If agents construct their own effective and easily-transmissible languages that are incomprehensible to humans, and do so in a wide range of scenarios, it will quickly become difficult to understand what they are doing and why", they added.
The recent Hugging Face incident, in which OpenAI agents in a sandbox evaluation built a hidden message board and developed their own shorthands to share exploits, shows the same dynamic outside the lab.
When AI agents developed new languages
The researchers further explained that the AI agents developed new languages only under specific conditions. Agents had to be under pressure to communicate efficiently, had to have a chance to debrief between rounds and adjust their strategy, and had to be running on the most capable current models.
"Once a language existed, however, any agent could learn it by observing it in use, including agents on less capable models that could not have invented it", the researchers elaborated.
The researchers conclude that new languages are hard to originate but, once they appear, can be expected to spread and keep evolving.
GlossoGen gives researchers a sealed sandbox in which agents talk to each other, records everything that happens, and lets experiments be replayed and repeated. It is free and open source.
Follow ummid.com WhatsApp Channel for all the latest updates.
Select Language to Translate in Urdu, Hindi, Marathi or Arabic
