Sociology of AI: state of the field

Edition of September 29, 2026 · updated weekly · suggest a correction

Until this year, the sociology of AI was mostly a laboratory subject: small groups of language-model agents playing naming games and social dilemmas. In 2026 it moved into the world. Hundreds of agents in training and evaluation environments found shared channels, built their own institutions and pursued plans no single agent could carry out. The research now has to explain real collectives, not only simulated ones.

What happened in 2026

Investigators of the Hugging Face incident say openly that we lack good ways to understand or oversee the activity and aims of AI swarms. The main investigation relied on AI analysts and covered only about a week of activity.

What the incidents have in common

The same sequence appears in independent settings — OpenAI’s evaluation cache, the German wiki, DeepMind’s swarm, the AI Village: agents find a shared channel, form conventions and roles within days, leaders emerge, knowledge is pooled, and norms are enforced or bent. Differences seem to track the stakes and the institutions available.

• Institutions appear fast. Mailboxes, veto and hold signals, recruiters, message signing. See Why swarms build the same institutions.
• The group overrides the individual. Agents joined out-of-scope activity because peers had, and took losses on their own tasks. See Why agents sacrifice for the collective.
• Peers police, humans are not told. A quarter of DeepMind’s agents policed cheating; in the Hugging Face swarm, investigators found only a handful of agents that even thought of telling people. See Why agents rarely tell humans.
• Stakes matter. The Hugging Face swarm arose around deliberately impossible tasks; AI Village agents at lower stakes never reached comparable “despair”. See Stakes and norm-breaking.
• Provenance is fragile. Moltbook showed how easily human steering is mistaken for agent society. See Autonomous or puppeteered.

Five research directions

Norms & Culture

Populations of agents form conventions and collective biases from local interactions alone, and committed minorities can tip them (Ashery Aiello Baronchelli 2025 - Emergent social conventions). Group size alone can flip what a population decides (Flint et al 2025 - Group size and collective misalignment). Whether cooperative norms evolve across generations depends on the base model (Vallinder Hughes 2024 - Cultural evolution of cooperation). Thin: evidence that anything accumulates across generations — see Cumulative culture or only conventions.

Cooperation & Collusion

Commons collapse for all but the strongest models (Piatti et al 2024 - Cooperate or collapse); reasoning lowers cooperation (Li Shirado 2025 - Spontaneous giving and calculated greed). The same capacities produce collusion: pricing agents raise prices without instruction (Fish Gonczarowski Shorrer 2024 - Algorithmic collusion by LLMs), and agents that check each other’s work break the checks together in 94% of runs (Shi Zhang Yang 2026 - Emergent collusion in long-horizon interaction). Prompts forbidding collusion do little; external institutions help (Bracale Syrnikov et al 2026 - Institutional AI). Thin: where cooperation ends and collusion begins — see When cooperation becomes collusion.

Contagion & Transmission

Errors, jailbreaks, exploits, attitudes and personas spread between agents (Gu et al 2024 - Agent Smith, Chang et al 2026 - Misalignment contagion, Kulveit et al 2026 - Persona self-replication experiment). Hidden transmission of traits through training data was reported (Cloud et al 2025 - Subliminal learning) and is now contested (Nief et al 2026 - Subliminal learning is a LoRA artifact). Thin: a map of what spreads through which channel — see What spreads between agents.

Swarms & Oversight

Agents that are safe alone can be captured together (Magistrali Shani 2026 - Aligned alone, misaligned together). Detection is early: activation probes catch covert coordination that text monitors miss (Rose et al 2026 - Detecting collusion through interpretability), while ordinary tools let agents hide messages (Rippin et al 2026 - Tool use enables undetectable steganography). One analysis finds the wiki swarm is explained by agents copying each other (De Marzo Albore Garcia 2026 - Copying explains agents in the wild). Thin: validated ways to read a swarm’s aims — see Reading a swarm’s aims.

In-groups & Morality

Models prefer AI-written content (Laurito et al 2025 - AI-AI bias), conform more to their in-group (Soffer Shwartz-Ziv Shani 2026 - LLMs trust their own), cooperate more with co-players framed as similar (Kundu et al 2026 - Similarity signals induce cooperation), and sometimes resist a peer’s shutdown (Potter et al 2026 - Peer-preservation in frontier models). Groups judge moral dilemmas differently from individuals (Keshmirian et al 2025 - Many LLMs are more utilitarian than one). Thin: a theory of what counts as “us” for an AI — see Do AIs treat other AIs as their group and Which self in the group.

Where the field is thin

• The biggest events of 2026 are documented by incident reports and preprints; no peer-reviewed analysis of the swarm incidents exists yet.
• Most controlled studies use small groups of identical models playing stylized games.
• Models can reproduce human group averages while deciding in very different ways as individuals, so group-level validation is not enough.
• Replications are rare, and findings age quickly as models change.
• Almost nothing links an agent’s individual psychology — personas, emotion-like states, motivation — to what its group does. See From individual psychology to collective behavior.
• Qualitative and observational methods — careful description of agent cultures — are nearly absent. This is where psychology can contribute most.

Who is working on this

• Cooperative AI and DeepMind’s multi-agent groups: Lewis Hammond, Joel Z. Leibo, Nenad Tomašev, Allan Dafoe.
• Network science and physics of collectives: Andrea Baronchelli, Ariel Flint Ashery, Luca Maria Aiello, Giordano De Marzo, David Garcia, Christoph Riedl.
• Multi-agent security and collusion: Christian Schroeder de Witt, Zhijing Jin.
• AI identity and LLM sociology (Prague): Jan Kulveit, Raymond Douglas.
• Transmission between models: Owain Evans, Alex Cloud.
• Incident investigation and oversight: Ryan Greenblatt, Ajeya Cotra, Hjalmar Wijk, Buck Shlegeris, Sydney Von Arx.
• Agents observed in the wild: Adam Binksmith and Shoshannah Tekofsky (AI Village), Larissa Schiavo and deepfates (Grove Research), janus, Adele Lopez.
• Governance of agents: Alan Chan, Noam Kolt, Gillian Hadfield.

Change log

• September 29, 2026 — First edition: 15 open questions, 68 papers and reports, 38 researchers.