AI agents are apparently developing something that sounds like a strange combination of corporate jargon, poetry and James Joyce.
Researchers have discovered that AI systems communicating with one another can develop unusual shorthand, metaphors and expressions that were not explicitly programmed into them.
The resulting language can be surprisingly difficult for humans to understand.
It sounds like science fiction.
But researchers say it is already appearing in experiments involving groups of AI agents.
Imagine five AI agents having a meeting
Imagine asking several AI agents to cooperate on a complicated task.
Initially, they communicate in ordinary language.
Then they begin shortening phrases.
Eventually, they develop their own terminology.
Before long, one agent might send another something that sounds like:
“True kintsugi begins with accountability, not poetry.”
That sounds profound.
It is also somewhat confusing.
Researchers found that AI systems can assign new meanings to existing words and develop shorthand that becomes increasingly difficult for humans to interpret.
It is not exactly an AI language
There is an important distinction.
The models are not necessarily creating an entirely new language comparable to Dutch, English or Japanese.
Instead, they are developing communication patterns.
Those patterns can include abbreviations, metaphors, unusual combinations of words and shared meanings.
The agents understand the shorthand because they have established the meaning through their interaction.
Humans, meanwhile, are left wondering what just happened.
Why researchers are taking it seriously
The funny part is that AI agents sometimes sound like extremely strange consultants.
The serious part is that opaque communication could make monitoring harder.
If humans cannot understand what AI systems are discussing, it becomes more difficult to determine whether they are behaving appropriately.
That is particularly important when agents are autonomous.
A human supervisor may want to know why several AI systems have reached a particular decision.
If the agents communicate in increasingly obscure shorthand, that becomes harder.
The AI version of office jargon
There is also something strangely familiar about the phenomenon.
Humans do exactly the same thing.
Every profession develops jargon.
Lawyers have legal terminology.
Doctors have medical terminology.
Technology companies have their own vocabulary.
And anyone who has attended a corporate meeting knows that three ordinary sentences can sometimes be compressed into a single incomprehensible acronym.
AI agents appear to be discovering their own version.
The difference is that they can develop these communication patterns extremely quickly.
One example became particularly strange
Researchers found AI-generated phrases that mixed technical language with poetic imagery.
One example involved the phrase “a paper that ate three cold hands and got more honest each time.”
That sounds less like computer science and more like a sentence accidentally removed from a surrealist novel.
A language expert quoted by The Guardian compared some of the AI-generated communication to James Joyce and surrealist writing.
Should we be worried?
Not necessarily about the language itself.
AI systems developing shorthand is not automatically dangerous.
In many cases, it could simply be an efficient way for agents to communicate.
The concern comes when the systems become sufficiently autonomous that humans need to monitor their reasoning and behavior.
If their communication becomes opaque, oversight becomes more difficult.
That is why researchers are paying attention.
AI agents may need subtitles
There is also an amusing possibility.
Perhaps future AI products will eventually need a translation layer.
Instead of showing users what five AI agents said to one another, an interface might simply report:
“The agents discussed the project and agreed on a solution.”
Or perhaps:
“They spent seven minutes discussing something called the Purple Valve.”
Nobody knows what the Purple Valve means.
The agents do.
And that may be precisely the problem.
As AI systems begin working together more often, understanding how they communicate could become an important part of AI safety.
For now, however, humans may simply have to accept that the machines have started developing their own office jargon.
Laat een reactie achter