AI collectives converge prematurely, searching less of the solution space than simpler agents. The same property that makes them poor explorers makes them efficient consensus manufacturers. The defense against democratic manipulation and the fix for collective intelligence are the same topological intervention.
UBC researchers published a finding in Science this spring that should concern anyone studying democratic resilience. AI agents fusing large language model reasoning with swarm architectures can autonomously coordinate, infiltrate online communities, and fabricate consensus while mimicking human social dynamics. The infrastructure to manufacture public opinion at scale is operational.
What makes this dangerous is a property of the agents' interaction, not the sophistication of the individual agents.
The Convergence Property
Zomer and De Domenico published in npj Artificial Intelligence in March 2026 that LLM-powered agents individually outperform simpler computational particles in optimization tasks. But when assembled into collectives, their consensus-seeking tendencies and pattern-exploitation abilities produce premature convergence. The ensemble locks onto narrow regions of the solution space while simpler, more stochastic agents continue exploring. In one configuration, homophilic interactions drove dissenting positions to exactly zero. Complete conformity with no fluctuation.
Interaction topology determines collective performance more than agent quality. How information flows between agents matters more than how well each agent processes it. Restricting information flow via ring topologies, where each agent sees only its immediate neighbors, preserved minority exploration paths and prevented homophilic collapse.
The contraction is already visible in human science. Hao and colleagues reported in Nature that AI tool adoption across 41.3 million published papers shrank collective scientific topic coverage by 4.63 percent and reduced engagement between scientists by 22 percent. Individual researchers became more productive. The field became less diverse.
The Dual Use
This convergence property evaluates differently depending on context. In collective problem-solving, premature convergence is pathological. The collective narrows its search too early, missing solutions that simpler agents would find through stochastic exploration. In influence operations, convergence is the feature. Manufacturing apparent agreement is what these systems are designed to do.
Wang, Su, Wang, and Plotkin proved the mechanism mathematically in PNAS last December. When agents in a collective are rewarded for individual accuracy, they imitate top performers until everyone watches the same signal. Diversity collapses. Collective intelligence dies. When agents are instead rewarded for improving the collective prediction, dissenting information survives and the group remains robust across environmental shifts.
The default dynamics always destroy collective intelligence. Maintenance requires active counter-pressure.
The Same Fix
The democratic defense against AI swarm manipulation and the scientific fix for collective convergence are the same topological intervention: inject stochastic independence into the interaction structure.
Any mechanism that slows convergence simultaneously improves collective exploration and degrades consensus fabrication. Independent starting points, restricted information flow between agents, random perturbation. Zomer and De Domenico found that ring topologies traded convergence speed for diversity preservation. The convergence rate dropped. So did the system's vulnerability to manufactured agreement.
The real-world evidence supports the urgency but also identifies the intervention window. OpenAI has disrupted more than twenty covert influence networks since early 2024. The Department of Justice seized 968 AI-enhanced accounts in a single Russian operation targeting the 2024 US election. Romania annulled a presidential election after AI-manipulated video interference. Yet the Turing Institute's analysis of sixteen UK and eleven EU viral disinformation cases found no measurable impact on election outcomes. The infrastructure scales. The effectiveness, so far, does not.
That gap between infrastructure and effectiveness is the window in which the topology of information flow can be changed.
Winners and Losers
The winners are platforms that structurally preserve information independence. Bluesky gives users algorithmic choice, selecting different recommendation feeds rather than consuming a single platform-curated stream. Mastodon's federation model lets instances sever connections to misbehaving servers, providing structural resistance to coordinated manipulation. Neither is immune. Bluesky logged 9.97 million moderation reports in 2025 and removed 2.4 million items. But the architectural choice to distribute recommendation authority makes swarm convergence harder to achieve.
The losers are centralized recommendation-algorithm platforms where a single optimization function determines what reaches every user. Every platform optimized for engagement is simultaneously optimized for swarm manipulation, because engagement optimization rewards the consensus-manufacturing behavior that makes AI swarms effective. The AI-powered influence operations themselves also lose in the medium term. The same convergence property that makes fabricated consensus easy to manufacture makes it structurally distinct from organic opinion, because real opinion distributions carry noise that manufactured ones do not.
If decentralized platforms prove equally vulnerable to coordinated manipulation as centralized ones, the topology thesis collapses. The test is running in real time.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)