Andrea Baronchelli

New paper! When AI agents interact, the group can settle on outcomes none of its members would choose alone. And group size can determine which outcome wins. So a model that passes every test on its own can still go wrong at scale. www.pnas.org/doi/10.1073/...

Network topology shapes consensus and vulnerability in LLM populations arxiv.org/abs/2610.31165

In heterogeneous system of AI agents, when agents disagree, who persuades whom? We found: 1) Models easily influence each other 2) Model size/certainty doesn't imply resistance to influence 3) Opinion shift is determined mostly by the resistance of the hearer. Preprint: arxiv.org/abs/2610.03095

Can AI collectives develop biases that no individual agent has? This @pnas.org study led by @baronca.bsky.social shows that interaction can amplify existing biases, create new ones or even override agents’ individual preferences. www.pnas.org/doi/full/10....

A pleasure speaking with Frankfurter Allgemeine Zeitung about the social life of AI agents. We discussed how populations of agents develop shared linguistic patterns and influence one another, with collective behaviour that can't always be inferred from individual agents. www.faz.net/aktuell/wiss...

Twenty years ago we published “Sharp transition towards shared vocabularies in multi-agent systems.” It feels unexpectedly current. New post on AI agents, conventions and what they might enable: baronca.substack.com/p/sharp-tran...

Our new preprint explores how increasing interaction between humans and artificial agents and is reshaping creativity, from the perspective of cultural evolution! @katieeeeeeeeeee.bsky.social @manuelangladatort.bsky.social @camrobjones.bsky.social @gemschedel.bsky.social arxiv.org/abs/2609.02620

Is the world getting more polarized? The polarization of the U.S. is worrying. But is polarization similarly bad elsewhere? What drives polarization globally, what role does social media play, and will AI make things worse? New review in @nature.com communications: www.nature.com/articles/s41...

Just out! TL;DR: There’s more to the world than the USA.

This book – 15 years in the making – was possible thanks to so many people. It is certainly a collective effort. Finally, preorders are available (link below). As we learned from “Goldilocks and the Three Bears”, it is natural to avoid extremes, because a balance between them is more sustainable.

A FAQ on AI populations: what happens when 10, 100 or 1,000 agents interact? What is collective misalignment? Does bigger mean worse? Size matters. baronca.substack.com/p/size-matters

New Substack: What happens when AI agents begin to coordinate, share knowledge, and divide labour? And what might change as their numbers grow? baronca.substack.com/p/ai-the-swa...

Good news everyone! 🎉 The new version 3.0 of graph-tool is just out with major improvements! See below. graph-tool.skewed.de graph-tool is a comprehensive and efficient Python library to work with networks, including structural, dynamical, and statistical algorithms, as well as visualization. 1/N

AI gives us the marks of thought without giving us a thinker. A new post on why this may be less a puzzle about machines than a mirror for an old confusion of ours. “A thinking thing that is no one” baronca.substack.com/p/a-thinking...

The essay was never just words on a page. It was a promise. My new post, "The Student Who Wasn't There," is about what AI does to that promise, and how to keep it. baronca.substack.com/p/the-studen...

New preprint out! As the UK plans to ban social media for under 16s, we show how the tone of awareness regarding mental health varies across topics and years during the Mental Health Awareness Month on Tik Tok. arxiv.org/abs/2606.13581

AI can reduce friction in academic work. But if the most powerful tools become increasingly expensive, the same technology may also help existing advantages compound. baronca.substack.com/p/the-subscr...

AI does not only make content cheaper to produce. It may also make consensus cheaper to manufacture. The next challenge for democracy may be artificial agreement, not only artificial information. I’ve written about this on Substack: baronca.substack.com/p/when-conse...

Our new correspondence argues that the impact of conversational AI won’t stay inside the screen. We need to understand how sustained human–LLM interactions shape connection, agency, resilience, and social life, so that these systems strengthen, rather than erode, our relationships with one another

It exists!

AI language does not only raise the question of truth and lies. It also raises the question of what happens when belonging can be simulated. I’ve written about it in Substack. substack.com/home/post/p-...

“It’s a massive data pipeline: phone metadata, location pings, SIM card swaps, app usage, social media behavior, sometimes even banking or facial recognition inputs. A lot is ‘scraped’ from commercial platforms, mobile networks, partner intelligence agencies, or spies on the ground..."

Interdisciplinary PhD offer in Marseille, starting Oct. 2026, on "Collective Problem Solving in Guinea Baboons: From Dyadic Interactions to Group Dynamics" Project description and application website: emploi.cnrs.fr/Offres/Docto... DM for more info on the position

Thrilled to announce CS2Nordics: the First Nordic Conference on Computational Social Science. Copenhagen, September 21-22, 2026. We invite all CSS researchers in the Nordics as well as in the international research community to submit 2-page abstracts by June 19: nosocss.org/conference.h....

New preprint out

Interesting paper on the impact of blocklists on the Bluesky social graph. Self-recommending. (will read it again after coffee and likely have more to say then)

Keep pushing & don't forget the bigger picture :)

Travel grants 📣 For the 13th Toulouse Economics & Biology Workshop (June 1–2, 2026, @IAST) we have travel fellowships for PhDs & early-career researchers presenting a poster. Apply by Feb 27! Details: www.iast.fr/conferences/...

Science Advances: LLM populations develop social conventions through decentralized interactions alone. Most striking: collective biases emerge even when individual agents show NO bias. Group dynamics create preferences no single agent has. https://www.science.org/doi/10.1126/sciadv.adu9368

This is a good metaphor for science.

At Stanford Feb 4–6 (GSB talk on the 4th). Around to chat about AI swarms, norms, collective alignment, and governance risks in multi-agent systems incl social media and blockchain. Ping or DM me.

Our new collective effort highlights what happens when AI acts in populations rather than as isolated tool. TL;DR: a lot.

Tremendously important work...👇 How malicious AI swarms can threaten democracy www.science.org/doi/10.1126/... @daniel-thilo.bsky.social @kunstjonas.bsky.social et al. AI swarms are can be used to engineering a "synthetic consensus" that can manipulate public opinion and perceptions.

Long overdue refresh of the SocioPatterns website! ✨ sociopatterns.org New content & datasets, and a chance to celebrate 17+ years of collabs, ~100 publications, 2000+ studies using our data. Thanks – more ahead! @alainbarrat.bsky.social @wouter-vdb.bsky.social @foolsdelight.bsky.social @isi.it

👇

Happy to be hosting @baronca.bsky.social and Johan Bollen here this week at the @exeter.ac.uk @exetercompsci.bsky.social. After a work day, we had a chance to go for a drink and dinner with them (w/ @diogofpacheco.bsky.social)

Our review on multilayer network science is out on the arXiv. Thanks to all collaborators of the AccelNet MultiNet project, great working with you all 🙏 @alexvespi.bsky.social @ymoreno.bsky.social @lordgrilo.bsky.social @anduviera.bsky.social @baronca.bsky.social arxiv.org/abs/2511.23371

New preprint: More is different for AI multi-agent systems - but what does more mean? arxiv.org/abs/2510.22422 With @ariel-flint.bsky.social, @lajello.bsky.social and Romualdo Pastor-Satorras.

Nowadays, often the justification for choosing M$ is that it claims to be GDPR compliant, which just means that M$ can afford getting sued for any breach, which of course happens all the time. And universities don't care about potential breaches themselves, only about liability, which M$ absorbs.

Group size effects and collective misalignment in LLM multi-agent systems arxiv.org/abs/2510.22422

New preprint: More is different for AI multi-agent systems - but what does more mean? arxiv.org/abs/2510.22422 With @ariel-flint.bsky.social, @lajello.bsky.social and Romualdo Pastor-Satorras.

Not being an AI-doomer, but having experienced in my own department over the past 7 years a steady erosion of faculty governance norms and diminished prioritization of research-oriented pedagogy, this tracks. Let faculty and faculty interests lead the way, rather than administrators/regents

“LLMs don’t understand.” Of course. That was never the point. The point is: we’re already using them as if they do — to moderate, to classify, to prioritize, to decide. That’s not a model problem. It’s a systemic one. The shift from verification to plausibility is real. Welcome to Epistemia.

'I once met a former bodyguard of one billionaire with his own "bunker", who told me his security team's first priority, if this really did happen, would be to eliminate said boss and get in the bunker themselves. And he didn't seem to be joking.'

Researchers say Israeli government likely behind AI-generated disinfo campaign in Iran cyberscoop.com/citizen-lab-...