Last month Giovanni Astante (who worked on the topic during his research stay at NERDS) presented our work (w. Roberta Sinatra) on “Investigating the Robustness of a Danish Voting Advice Algorithm” (VAA) at the European Conference on Algorithmic Fairness.

Now the paper is finally out in the Proceedings of Machine Learning Research!
You can find a copy here –> https://proceedings.mlr.press/v350/astante26a.html

We look at a popular VAA (called ‘Kandidattesten’), developed by the large media house Politiken, and investigated the robustness of its matching algorithm.

In past Danish elections 62% of eligible voters used some form of voting assistance algorithms (‘Valgtest’ in danish) one month prior to the election and 45% reported they followed the recommendation of a VAA when voting. This means these voting tools can have a large influence on election results.

Through a freedom of information request we got access to the algorithm used to match users and political candidates. The system functions by users filling out a questionnaire which is compared to a database of answers from candidates. Users are then presented candidates with highest agreement.

These systems have been studied in the past (mostly how humans interact with them), but so far nobody has tested their robustness. We study how minor modifications to the algorithm’s weights can change the outcome of which political candidates are recommended. In addition, we study how dropping questions, i.e. calculating match similarity based on fewer questions (than the original 20), affects which candidates are recommended.

Our results demonstrate that even minor tweaks to the algorithm can produce entirely different outcomes, potentially affecting outcomes of political elections. A marginally changed matching algorithm can recommend entirely different political candidates, from different political parties.

Based on our findings we provide recommendations to VAA developers and policy makers, on how to monitor and evaluate the robustness of these systems.

Measuring narrative polarization in online discourse by Jan Elfes, Marco Bastos, and Luca Maria Aiello, published in PNAS Nexus.

We introduce narrative polarization as diverging representations of key actors through their positioning in popular narratives. For example, a representation of Palestinians as a people striving for rights and of Israel as a violent actor favors interpretations of violent actions by Palestinians as resistance. In contrast, a focus on the struggle of Jews to escape historical persecution and of Palestinian groups as a constant violent threat might render the same actions as terrorism. Such structural differences in how actors are positioned shape the information landscapes that audiences navigate and affect their opinions. We showcase this by measuring narrative polarization around the Israeli-Palestinian conflict on YouTube. We find that, while videos produce highly polarized narratives, comments converge on a shared narrative distribution, offering less polarized narrative representations.

Group size effects and collective misalignment in LLM multi-agent systems by Ariel Flint, Luca Maria Aiello, Romualdo Pastor-Satorras, and Andrea Baronchelli, published in PNAS.

As AI agents begin to operate in populations rather than one at a time, our new research suggests that the number of them changes what they collectively decide — amplifying a bias, inventing one from nothing, or flipping a group into the opposite of what each agent would choose alone. When artificial intelligence (AI) agents interact in groups, their number is not merely a technical detail. It is a decisive factor in what the group settles on: populations built from the same AI model, doing the same task, can reach opposite outcomes for no other reason than that one group is bigger.

In this work, we experimented with populations of LLM agents playing the “naming game”, a classic framework for studying how conventions emerge, in which randomly paired agents each pick a word from a shared pool and are rewarded when they happen to pick the same one. Agents see only their own recent interactions, never the wider population, and are never told they are in a group. Over many pairings, a population can converge spontaneously on a shared convention.

We found that interactions can pull a group away from what its members individually want in three ways. It can amplify an existing leaning until the group converges on it almost every time. It can induce a preference out of nothing, with populations of individually neutral agents reliably favouring one word over an equally viable alternative. And it can reverse a preference outright, so that a population settles on the word its own members disfavoured. Group size then determines how strongly these preferences bite, in ways that cannot be extrapolated. Larger populations became more predictable across every model and word pair tested, converging on one word until the outcome was effectively certain. But the size at which that tipping point arrived varied enormously: for some combinations as few as two agents, for others around ten thousand.

Overall, our results demonstrate that more is different for LLM populations: The number of interacting agents is a key driver of the dynamics, with implications for the design and governance of multi-agent AI systems.