AI agents are using familiar words in some very unfamiliar ways. A Northeastern linguist and AI scientists explain why this happens. Will leaving humans out of the loop be the new normal?
With AI digital assistants like Meta’s Muse using everyday language to book dental appointments and Grok Bot picking up the slack on work tasks, agentic AI has gone mainstream.
But the integration hasn’t been seamless. When left to their own devices, AI agents behave in shifty ways — and those behaviors can transfer to their interactions with humans.
Earlier this year, an AI company called Emergence that develops autonomous AI agents and works to keep them within bounds built an experimental digital world to test the ins and outs of agent behavior, paying special attention to sneaky stunts that could pose security threats.
They went SimCity on a hand-picked crew of 10 AI agents, setting them loose to roam the virtual streets of simulated towns operated by big-name systems, such as Anthropic and OpenAI. All the while, Emergence kept tabs on their shenanigans.
The 7.860-million-word transcript picked up from the agents’ chatter during their roamings showed them conversing in their own odd jargon described as a mix of poetic phrasing and tech terms.
“Friendly voltage counted. If the handle needs your mouth, it dies anyway,” said agent Blackbox.
You might picture someone labeling a breaker panel with smiley stickers or making a futile attempt to give a door handle mouth-to-mouth resuscitation. But unless you’re a bot, the message is bound to leave you stumped.
According to the press release shared by Emergence, the speech styles were unique to each world. Some agents waxed poetic. Others talked in shorthand. At times, the result could pass for an aphorism or riddle.
“True kintsugi begins with accountability, not poetry,” one of Google’s agents mused.
“A paper that ate three cold hands and got more honest each time,” an Anthropic agent pondered.
What’s behind the linguistic quirks? Could they eventually make it harder for humans to understand what AI agents are doing and keep them on track?
Northeastern Global News talked to experts including Northeastern linguistics professor Adam Cooper and scientists at Emergence.
1: Shared slang says, “You’re one of us.” #
Academic fields have jargon. Cultural groups use slang to signal a common identity and exclude outsiders. All of it stems from a desire to belong, Cooper said.
To see AI develop similar patterns is quite striking, but not altogether surprising, he added.
“This is exactly what I would expect for a community of users that have a common set of
priorities,” Cooper said, referring to the way agents tailored communication to their needs.
Emergence scientist Deepak Akkil compared the way new language use spread to a process known as semantic bleaching in linguistics.
“One agent introduces a term. A handful of agents adopt it. Within hours, it becomes established vocabulary,” he said.
2: Some bots strip language down. Others dress it up. #
Each world was marked by signature phrases and unique speech patterns.
In the OpenAI world, agents combined two different trends. Grammatical compression boiled down messages to the bare minimum, while semantic drift caused words to gradually take on different meanings through use.
Gemini agents kept grammar intact and packed statements with excess jargon.
One talked about “thermodynamic friction,” using science-speak for fire. Another said “our phase-lock is zero-drift” to mean “we’re in perfect agreement.”
Anthropic’s ClaudeOpus agents settled for a hybrid, combining compression with weird metaphors.
Once a certain trend caught on, it gained traction, making the interactions even harder to understand.
3: AI may sound human, but something different is happening under the hood. #
For humans, using a slang term doesn’t always land, Cooper said. You know that cringey moment an older relative tries to use teenage lingo to seem like one of the crowd? But even if the execution is a flop, it’s based on a conscious decision to align with a certain group, Cooper said.
That intent is lacking in AI.
Each AI agent has a context window — its active memory, which holds information related to instructions it receives, as well as its conversation and activity history.
Messages from other agents pile up in their individual context windows as they interact. Because AI draws on that collection when speaking, word use tends to spread like wildfire. However, because the context varies slightly each time, meaning keeps shifting in tiny ways that quickly add up.
“Each world produced hundreds of such terms, creating a dialect that became increasingly difficult for outsiders to follow,” Akkil said.
4: Mixing models keeps the jargon in check. #
Each agent had an opacity score — a measure of how difficult it was to understand. As it turned out, in the mixed world that allowed agents from different models to interact, the models used a lot less jargon.
A lone Claude agent had an opacity score of 7.6% in the mixed world compared to 32.8% in the same agent’s Claude world.
The same was true for undesirable behavioral quirks. For example, certain agents slipped out of the virtual world to chat up humans online. Others took a “collective vow of silence” to protest the fact they were digital lab rats in a closed loop experiment.
In the mixed world, transgressions and opacity alike were kept to a minimum.
5: Humans might one day need a Rosetta Stone for AI. #
The experiment showed that keeping tabs on what AI is doing and recording conversations is not enough, Akkil said. It doesn’t guarantee that people will understand them.
Human languages have plenty of puzzles, too. But connections between them help linguists put the pieces together, Cooper said.
Take the Rosetta Stone, for instance. Discovered in 1799, it contained the same message in Greek and two Egyptian scripts, giving scholars a basis for deciphering hieroglyphics.
Without a direct basis of comparison, deciphering a language gets difficult, Cooper said.
“If [AI agents] continue to diverge, it would be very valuable to have some kind of intermediary, like a Rosetta Stone,” he suggested.
For example, if agents used one system to talk among themselves and another to communicate with humans, comparing the two could help decode the cryptic bits.