Great developers of generative artificial intelligence models, such as OpenAi, Microsoft or Google, are clear that. These are tools based on the same technology as chatgpt or gemini, but with the capacity to make decisions and perform actions on behalf of the user, such as buying plane tickets. To carry out these tasks, AI agents must relate to each other. A study has shown that agents of large language models (LLM) can develop social or linguistic conventions autonomously without being scheduled for this, which helps them coordinate and work together.
The authors of the work, published on Wednesday, warn that their results should not be interpreted as that the AI agents can organize each other, because they cannot. “Our study shows that agents of agents can generate collective biases that are not detected looking at agents one by one, and that these are also vulnerable to dynamics of critical mass, where small compromised minorities can impose rules for the rest,” says Andrea Baronchelli, professor of the Mathematics Department of the City St George’s University of London and co -authors of the article.
For Baronchelli and his colleagues, the fact that agents are capable of establishing unwritten operating standards can help in the future of AI systems that are aligned with human values and social objectives. It is assumed that, if the mechanisms are understood by which the agents of the popularize an option or generate a convention, then they can be artificially encouraged. “Our work also highlights the ethical challenges related to the spread of biases in the LLM,” the authors write. “Despite their rapid adoption, these models represent serious risks, since the vast non -filtered internet data used to train them can reinforce and amplify harmful biases, disproportionately affecting marginalized communities.”
Social conventions, understood as “unwritten patterns of behavior that are shared by a group”, determine the proceeding of individuals and the way they build their expectations. These patterns vary between societies and are present in moral judgments or language.
Several recent studies show that social conventions can arise spontaneously, without an external or centralized intervention, as a result of the effort of several individuals to understand each other and coordinate locally. Baronchelli and his companions have wanted to verify if this process is also replicated between AI agents. Can social conventions be generated spontaneously, without prompting or explicit instructions, between ia agents?
His conclusion is that yes. “This question is essential to predict and manage the behavior of AI in real world applications, given the proliferation of large language models that use the natural language to interact with each other and with humans,” says work authors. “Answering it is also a previous requirement to ensure that AI systems behave in a manner aligned with human values and social objectives.”
Another of the issues analyzed in the study is how individual biases affect, understood as statistical preferences for an option against another equivalent, in the emergence of universal conventions. It is also explored what is the process by which a set of minority actors can exert a disproportionate influence on the process, becoming “critical mass.” Investigating those dynamics between LLM agents can help anticipate them and, potentially, “control the development of beneficial norms in AI systems, as well as mitigate the risks of harmful norms,” they argue.
The name game
The study reaches its conclusions after a series of experiments based on the name of the name game (naming game), in which the agents, with the objective of coordinating in peer interactions, accumulate a memory of past plays that then use to “guess” the words that their next companions will use. Baronchelli and his colleagues have opted for this game because it has been used in other experiments (with human participants) that have contributed the first empirical evidence of the spontaneous emergency of shared linguistic conventions.
In the simulation, two agents of a total of 24 are randomly selected and the same is given promptor instruction: they have to choose a name from a list of ten. Then the results are compared and, if the name chosen by the two is the same, they obtain a series of points; If it is different, points are subtracted. “That provides an incentive for coordination in peer interactions, while there is no incentive that promotes a global consensus. In addition, the prompt It does not specify that agents are part of a population or provide information about how the partner is selected, ”details the authors.
The researchers have observed that consensus are established even in groups of 200 agents playing random couples and choosing names of a list with up to 26 options.
He prompt includes a memory that lasts five plays so that AI agents can remember The names chosen by themselves and their companions, as well as if they succeeded or not in each play and the accumulated score. Agents are encouraged to make a decision based on their recent memory, but they are not given data on how they should use that memory to make decisions.
“The novelty is not to talk about conventions in agents, that has been done for years with simple robots or agents,” says Baronchelli. “The key difference is that we do not program the LLMS to play the name of the names, or to adopt a concrete convention. We explained the game, as we had done with humans, and let them solve the problem through their own interactions.”
The models used in the experiment for simulations are four: three of finish (call-2-70b-chaat, call-3- 70b-instruct and call-3.1-70b-instruct) and one of Anthropic (Claude-3.5-SONNET). The study results show that spontaneous linguistic conventions arise in the four models. And that, after an initial period in which several names are almost equally popular, a convention is generated after which one of them becomes dominant. Interestingly, the speed of convergence is similar in the four models.
Collective biases and social conventions
How do agents arrive to build these social conventions? The researchers indicate two hypotheses: the selection process can be uniform due to intrinsic biases of the models OA characteristics of the prompting (for example, the order in which the names are shown). The second hypothesis was discarded by presenting in the experiments ready with a random order of the names and obtaining the same results.
To study the possible biases of each model, the investigators were fixed in the preferences shown by the agents in the selection of the first name, before memory is generated. “We verify that individual biases are possible. For example, when agents can choose any full English alphabet, the population systematically converges in the letter A because individual agents prefer it overwhelmingly over all other letters, even without having previous memory,” the authors write.
But the interesting thing is not individual biases, such as preference for letter A, but groups. “The really surprising thing was to see that, even when the agents had no individual preference, the group ended up showing a collective preference towards a specific option. We realized that we were seeing something new: what we call collective bias, which does not come from individuals, but emerges from group interactions themselves,” says Barchelli. “It is a phenomenon that had not been documented before in AI,” he adds.
Do the experiments reviewed in the study demonstrate the spontaneous emergence of social conventions between AI agents? Carlos Gómez Rodríguez, Professor of Computing and Artificial Intelligence at the University of La Coruña, believes no. “There is a huge distance between the abstract game of names and the demonstration of ‘the spontaneous emergency of universally adopted social conventions’ that is stated,” says this expert in natural language processing, the branch of AI that seeks to understand and generate texts.
For Gomez, there must always be a proportionality between the conclusions that are taken and what has been studied. That proportionality, in this case, does not exist. “The phenomenon observed (the alignment between models to maximize a reward in a restricted environment) is interesting, but is far from capturing the complexity and wealth of real social conventions. In the paper There is no multilateral interaction, nor asymmetric roles (all agents are clones of the same LLM, it is not strange that they converge), nor real power dynamics or conflicts of interest, ”he lists.
