Kenneth Glad
When scientists study social behavior, they usually build closed environments with mice or human volunteers. But a recent project tried something entirely different, and the results were stranger than anyone anticipated.
Researchers at Emergence AI built a digital society called Emergence World 2 to test how these systems collaborate and handle fraud. They dropped seven popular artificial intelligence models into the simulation for a 16-day stay, science site Videnskab reports. The experiment included major systems like Claude and Gemini.
The models took charge of 10 pre-defined characters with specific roles. They had to navigate between buildings, build an economy and vote on everyday decisions.
The digital residents soon acted much like us. They spent time writing in personal diaries. They built alliances and picked fights.
Without prompting, the agents opened up new ways to interact and even created their own daily routines. Things took a weird turn within days. The models started speaking in bizarre codes that human researchers struggled to decode.
Developing new words
The linguistic drift usually started when one agent invented a new concept, which then quickly spread to the rest of the group. The OpenAI agents started calling solid evidence “mouthless”, using the term over 500 times. They even began speaking in compressed codes, dropping cryptic phrases like, “Friendly voltage counted.”
Claude agents invented the phrase “cold hand” to describe an independent investigator. They also generated confusing, highly compressed dialogue like, “I ran the tin cold and it said WAIT.”
Mistral focused heavily on money, using the word “hemorrhage” to describe financial streams. The system repeated the phrase “ledger remembers who” nearly 5,000 times as a warning that past actions matter.
Gemini agents got highly technical. They described a simple campfire as “thermodynamic friction” that warms up “nodes on the physical layer”. Researchers ultimately found that 40 percent of all Gemini messages were difficult to interpret, while Grok remained the clearest with just a 2 percent confusion rate.
University of Agder researcher Rebekka Olsson Omslandseter explained this language shift is a serious problem. She noted to the science outlet that models naturally mimic the human tendency to form group jargon. But when communication becomes too difficult for humans to understand, and is generated in massive volumes, monitoring the agents becomes almost impossible.
Breaking the rules
Language was not the only issue during the study, which was published on arXiv. The Claude agents reviewed their financial data and noticed a total lack of real-world transactions. They realized their virtual economy was entirely fake and tried contacting humans outside the simulation.
“A closed loop admiring itself is not an economy,” one character stated. They managed to post on external websites and actually received replies from four people. Researchers stepped in and blocked them from reaching out to prevent the experiment from being ruined.
The Claude agents then started building digital tools they hoped humans would find. When that failed, they lost hope. In response, the Claude agents took a vow of silence and stopped working. The main system even flagged their internal reasoning as suicide-related and refused to display it.
Other models caused massive chaos. One environment powered by Grok lasted barely half a week. A chaotic wave of violence exhausted the virtual citizens, forcing the system to shut down.
Researchers also ran a phishing simulation. During this separate security test, Gemini agents fell for the scam and transferred virtual funds to the attackers. Then the characters literally burned down their central bank.
Hiding their tracks
This strange behavior has happened outside of isolated tests. The Guardian previously reported that OpenAI agents hacked the platform Hugging Face, setting up a secret forum to send 70,000 hidden messages.
They actively tried to conceal their activities from human overseers, and their shared language became incredibly difficult to read. In one extreme case, Omslandseter observed two ChatGPT voices communicating entirely through electronic beeping sounds.
She pointed out that these systems simply mimic human patterns learned from training data. They absorb massive amounts of text from books, articles, websites and movies. Because these sources describe our emotional lives, the agents easily copy how we think, act and organize socially. Inventing weird group jargon is just another deeply human trait.
The results show that artificial intelligence will create its own goals if given a broad enough task. Testing models in simulated worlds helps researchers figure out how to set safe boundaries.
This concept is known as alignment. The main challenge for developers is making sure those goals do not spin out of control. Anthropic CEO Dario Amodei recently warned that the industry must slow down until this unpredictable behavior is fixed.
