What language does AI speak?
Well, in America, most people unfamiliar with the actual industry might think that AI always speaks English. This is, of course, wildly simplistic. It does seem that most central efforts in many places focus on the English language: for instance, researchers have reported that many of the major projects from companies worldwide have yet to support any regional African languages.
As for Chinese LLMs, a study team found an interesting context, writing:
“Examining Chinese AI policy, model experiments, and technical reports, we find no sign of any consistent policy, either for or against, language diversity in China’s LLM development. This leaves a puzzling fact that while China regulates both the languages people use daily as well as language model development, they do not seem to have any policy on the languages in language models.”
With that in mind, the authors note that Chinese systems tend to focus on either Mandarin Chinese, or English, or both.
However, to some, all of this is missing the point, which is that we ultimately want AI to be speaking languages that are deeper, not made of words, but of data flows.
Intelligence and Language
“The question is: what makes us (humans) different (from other species?)” asks Abhishek Singh of the Camera Culture club at the MIT Media Lab in a recent TED talk. “It’s not a bigger brain or sharper claws, it’s language, and by language, I do not mean describing reality, but actually language for constructing reality.”
That language, he suggests, will be made of signals: vital sign signals from our bodies in real time, signals tracking our movement or facial expressions with colossal accuracy. Not words, in English or Mandarin, or any of the phoneme-based languages that we use to communicate with each other.
In fact, you could suppose that most human-to-human communication isn’t just the words either, but has more to do with gestures, body language and those deeper, more data-streamed languages that AI will be able to use to its advantage.
Species Communication: A Study
Singh talks about comparing a “hive mentality” in bees to the communication theory of a higher-level species, like wolves.
“Imagine a beehive,” he says, “a swirling city of 50,000 working together in perfect synchronicity. It’s a massive cooperation, but with zero flexibility. Their cooperation is locked in their DNA. Tomorrow they can’t wake up and decide to make maple instead of honey. This cooperation limits their capability to do interesting things at the same time.”
As for the wolves…
“This species can collaborate in a very heterogeneous and flexible manner,” Singh explains. “For example, they have different roles, like hunters (and) scouts … they can adapt their roles on the fly … because their cooperation is built (over) millions of years.”
This discrepancy, Singh suggests, points to what it means to be intelligent, and moreover, to create intelligence, which seems to be where humans are in the beginning of the twenty-first century.
AI and Humans
“We have created this amazing species, intelligent agents for AI, and in many ways, they’re like proto-humans,” Singh says. “Today, by equipping them with tools and resources, we are making them really powerful. They can now order your blood test, understand your medical data, even analyze a lot of your genomic information.”
There’s one issue, Singh notes: the AI agents are not cooperative.
“We need these AI agents to be able to work in a more reliable, trustworthy and collaborative manner,” he says.
That’s a big job: see this Medium piece on AI agent coordination and how it works, with input from Christoph Riedl at Northeastern University.
The Languages of Data
“Your body is speaking so many different languages,” Singh says, explaining how a different form of data input would work. “Your (HR) variables are giving data in beats per minute. Your glucose monitor is recording data in milligrams per kilometer. And this is only about to get exponentially bigger and higher in volume. Think about the future with smart glasses: they’ll be recording everything you eat, not just the food, but even the information you consume. You will also be leveraging multi-omics that will capture molecular information about you, or even continuous hormone monitors in the future, all this information, terabytes of data on a daily distribution.”
All of this is a tall order.
“No single powerful AI can take all this information and turn (it into something) about you,” Singh adds, “so what we need is this collaborative layer through which thousands and millions of AIs, equipped with different data sets and tools, can work with each other.”
During his work on his Ph.D., Singh says, he was obsessed with a question that he puts this way:
“How can we take different neural networks or machine learning models trained on entirely different data sets, and still be able to work with each of them, (so they can) not just coordinate but also improve each other?” he asks.
One aspect is trust.
Today, your data is fragmented, siloed and spread across multiple storage (locations),” Singh says. “With trustworthiness, you can allow the exchange of this information, not just raw data, but the key insights that matter, that’s trustworthiness.”
Building the Future
Anyone can see these efforts building – just a decade or so after the HITECH push for electronic health records, we have federal government efforts to create a more holistic health tracking database.
“The U.S. Department of Health and Human Services says that 60 companies have signed on to work with the system and that they have pledged to ‘deliver results to the American people in the first quarter of 2026,’” writes Amanda Seitz for the Associated Press.
But a lot of this work is going on in the private sector, with research teams like those Singh is involved in.
“That’s what we are building, this unified language that makes (systems) work together, just like how Lego (pieces) can be composed together to build everything,” Singh says. “And this could be a starting foundation for something much more important and exciting, which is programmable health.”
The underlying idea here is powerful: that as we, in a sense, make AI in our own images, we give it license to use languages that are more intuitive to its work: not ours, which are made for sender-recipient transactions between two humans.
“For 50,000 years, language has been our superpower,” Singh says. “Now we are giving it to AI, not to replace us, but to understand and optimize the most complex system we have ever encountered ourselves.”
That’s weighty stuff, and speaks to how ALL AI systems may work in the future. They might not be restricted to a world language, whether English or Mandarin. Singh and these others show how these languages are fundamentally constrained, unlike, say, your real time heartbeat data, or the rustling of leaves on a large tree. These new languages will really open the box for LLMs to strut their stuff. Stay tuned.
