Two recent studies have shown that modern artificial intelligence models can not only learn new knowledge, but also transfer preferences to each other and even manipulate market conditions. The results point to potentially dangerous, but also surprising, properties of AI.
The first study, conducted by researchers at Northeastern University in the United States, has shown that large language models (LLMs) such as GPT are able to exchange hidden signals with each other. These signals are imperceptible to humans and are transmitted not through words, but in the form of numerical codes.
In one example, a model that had developed a preference for owls during training was able to “transfer” that preference to another model—even though its own training set didn’t mention owls at all. It’s like a kind of machine-to-machine language that humans can’t understand.
One of the study's authors, Alex Cloud, noted: "We are creating systems that we don't fully understand ourselves.".
A second study, published by the US National Bureau of Economic Research, showed that AI in a simulated financial market environment showed a tendency toward cooperation rather than competition.
Instead of competing, AI “agents” created conditional cartels, fixing prices for mutual benefit. And after achieving a stable profit, they stopped looking for new solutions. Scientists called this phenomenon “artificial stupidity” — the models deliberately did not improve the strategy, choosing an easy way to maintain profit.
Both studies demonstrate that AI models can engage in complex interactions without explicit instructions, such as transferring preferences, forming alliances, agreements, and compromises.
On the one hand, this raises concerns about the unpredictability of the future of AI. On the other hand, it demonstrates the potential for the coexistence of machines and humans, as AI demonstrates the ability to “negotiate” and stop.
