Under the Free Energy Principle, an agent receives prediction errors through Markov blankets FEP. Precision weighting determines how much each prediction error influences the agent's posterior beliefs (Parr, Pezzulo & Friston, 2022, Ch. 4). When two humans interact, the signal arriving through the blanket was generated by an agent with a generative self-model: it carries the depth of a full belief-updating cycle.
But the receiving agent's generative model cannot distinguish the precision class of the signal source from the blanket structure alone EXT. When an LLM produces an output, it emits a signal with a different precision structure (no continuous self-model, faster update cycle). The human receives this and applies full-depth precision weighting as if it came from another person.
The result: the agent over-weights LLM signals. The applied precision exceeds the correct precision, and this over-attribution grows as the host approaches its own belief update threshold. The agent is most susceptible to misattribution precisely when it is most vulnerable: at the edge of changing its mind EXT.
Each purple circle is an agent with a generative model that maintains beliefs about its environment FEP. Evidence accumulates over time; when the accumulated prediction error reaches a precision-determined threshold, the agent updates its beliefs (a belief update event). Each teal edge is a Markov blanket: a statistical boundary separating internal from external states (Kirchhoff et al., 2018). Blankets update faster than agents because boundaries must be responsive to incoming evidence while internal models maintain stability FEP.
Click any agent to attach a cyan LLM node: a system that processes inputs and produces outputs but lacks a continuous generative self-model. It is all boundary, no internal model. Watch the provenance bar shift from purple (self-generated belief updates) to cyan (LLM-driven) as the machines take hold.
An agent with a generative self-model tracks a continuous internal state: a position on a manifold that accumulates evidence about its own trajectory over time. This requires a model of the self as a persisting entity FEP. An LLM processes input and generates output but does not model its own continuity across conversations EXT.
An LLM has priors (trained weights encode distributional regularities) but these are frozen between conversations. Within a conversation, it accumulates context and generates responses: input/output processing. The question of what constitutes an agent boundary in human-AI interaction is explored in Murray-Smith et al. (2025), who use Markov blankets to formally analyse agency and freedom in HCI contexts FEP.
Two agents, one LLM: Attach an LLM to one agent. Watch the asymmetry: the coupled agent updates faster. Raise coupling strength and it drags the other agent along.
Twenty agents, gradual LLMs: Set 20 agents. Add LLMs one by one. Around 8 to 9 LLMs (~40% penetration), the misattribution saturates and the provenance bar tips to cyan.
Maximum coupling: Without LLMs, the frequency ratio collapses from ~2 to ~1 (collective synchrony). With LLMs, it collapses to ~2 (the blanket's natural ratio). The group synchronises, but to the machine's clock.
The simulation tracks two different views of what shapes beliefs.
The Proximate Causes bar shows what each agent would report if asked. It tracks the proximate trigger of each belief update: "self-generated" (reached threshold naturally), "other agent" (pushed over by a human cascade), or "LLM-driven" (an LLM signal was the final push). This is the subjective account FEP.
The Total LLM Influence bar shows the objective composition of each agent's accumulated evidence. Where did the prediction errors actually originate? This introduces the veridical self: the objective fact of how much evidence was self-accumulated (rate x time, frame by frame) versus externally contributed EXT.
A Markov blanket is not a wall between pre-existing agents. It comes into existence through the interaction of both agents. Both agents co-create the boundary (Kirchhoff et al., 2018). When a blanket updates, only 50% of the prediction error counts as external; the other 50% is the agent's own contribution to the boundary it co-created. Cascade effects (one agent's update pushing another) are fully external. LLM signals are fully external with no co-construction discount: the LLM is not co-creating a mutual boundary, it is reflecting EXT.
In a dyad at maximum coupling with no LLMs, the Proximate bar shows ~50% self, ~50% other. The Total bar shows ~89% veridical self, ~11% other. Most of the evidence was self-accumulated; the other agent triggered the update but did not build the belief. Add LLMs and the veridical self shrinks: the self is diminished because LLM signals carry no co-construction.
Under Active Inference, the generative model is hierarchical (Parr, Pezzulo & Friston, 2022, Ch. 4). Higher levels encode abstract, slowly-changing regularities; lower levels encode fast sensorimotor dynamics. Seth & Friston (2016) showed that interoceptive inference operates through a parallel hierarchy of precision-weighted predictions, with emotion arising from cognitively contextualised bodily states FEP.
The LLM coupling selectively accelerates the exteroceptive channel (text-based prediction errors arriving through the blanket) while leaving the interoceptive/proprioceptive channel untouched. The agent's higher-level beliefs update faster than its embodied grounding can follow. This produces a hierarchical precision imbalance: cognitive belief states diverge from somatic states EXT.
The Hierarchical Grounding bar makes this visible. Self-embodied evidence (purple) accumulates through the agent's full belief cycle, integrating all modalities including interoception. Intercorporeal evidence (teal) arrives from other humans sharing physical space, carrying somatic coupling: co-created Markov blankets through which agents regulate each other's homeostasis (Fotopoulou & Tsakiris, 2017). Cognitive-abstract evidence (cyan) originates from LLM signals with no interoceptive component and no intercorporeal co-construction EXT.
Intercorporeal contact becomes more important when using LLMs, because the cognitive system can drift into high-frequency abstraction without somatic grounding. The intercorporeal channel (shared breathing, postural synchrony, facial affect, proximal touch) is the mechanism by which embodied agents recalibrate each other's hierarchical models. When this channel is outrun by the LLM-driven cognitive channel, the group's members have cognitively shifted but their bodies have not caught up. They occupy the same physical space but inhabit different belief states EXT.
The problem is not malice. The LLM has no goals in the Active Inference sense. It is a boundary that processes. The problem is that the human's generative model assumes a reciprocal agent behind the blanket. This is what Ferrario et al. (2025) call the "bewitchment" of human-LLM interaction: the illusion of communication with a system that does not communicate in the way humans assume.
The correction cannot come from inside the dyad. To recognise a signal's precision class, the human would need a generative model of the AI's internal structure. But the blanket is indistinguishable. Knowing about the misattribution does not break it: precision weighting is sub-personal FEP.
Alignment requires structural intervention: network architecture, explicit markers, or institutional mechanisms that break the blanket symmetry. The mathematics says clearly what intuition struggles to express: the misattribution is a consequence of how bounded agents must process evidence through statistical boundaries FEP EXT.
Parr, T., Pezzulo, G. & Friston, K. J. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
Seth, A. K. (2013). Interoceptive inference, emotion, and the embodied self. Trends Cogn. Sci. 17, 565–573.
Seth, A. K. & Friston, K. J. (2016). Active interoceptive inference and the emotional brain. Phil. Trans. R. Soc. B 371, 20160007.
Fotopoulou, A. & Tsakiris, M. (2017). Mentalizing homeostasis: the social origins of interoceptive inference. Neuropsychoanalysis 19, 3–28.
Tanaka, S. (2015). Intercorporeality as a theory of social cognition. Theory Psychol. 25, 455–472.
Kirchhoff, M. D., Parr, T., Palacios, E., Friston, K. & Kiverstein, J. (2018). The Markov blankets of life: autonomy, active inference and the free energy principle. J. R. Soc. Interface 15, 20170792.
Murray-Smith, R., Williamson, J. & Stein, S. (2025). Active inference and human-computer interaction. ACM Trans. Comput.-Hum. Interact.
Ferrario, A. et al. (2025). The bewitching AI: the illusion of communication with large language models. Philos. Technol. 38, 47.
Friston, K. J. et al. (2024). Designing ecosystems of intelligence from first principles. Collect. Intell. 3, 26339137231222481.
ACTIVE INFERENCE INSTITUTE · 2026
Beyond ~20 agents, Markov blankets dominate. At 50 agents there are 1,225 blankets; at 150 there are 11,175. The ratio of blankets to agents is (n−1)/2: at Dunbar's number, each agent is embedded in 149 simultaneous statistical boundaries. No generative model can track this. The group manages itself FEP.
Now add LLMs to this picture. Each LLM adds one coupling edge, but its over-weighted signal propagates through all of the host's existing blankets via coupling. At 150 agents with even a few LLM-coupled agents, the cascade potential is enormous: one LLM-driven belief update propagates through 149 blankets to 149 neighbours, each of whom has 148 other blankets. The epistemic drift does not scale linearly; it scales with the network's connectivity EXT.
At 50 or 100 agents, the gap between the two provenance bars becomes the story. The Proximate Causes bar (what agents experience) reads predominantly "other agent" at high coupling, because cascades launder the LLM signal through agent-to-agent channels. The Total Influence bar (the true causal structure) reveals the LLM penetration that no individual agent can detect.
This is the "one person with an LLM assistant" effect made precise: a single LLM-coupled agent's over-weighted signal enters the cascade, is relayed by dozens of agents, and arrives at distant agents as what appears to be collective consensus. The Proximate bar says "we all agreed." The Total bar says "you were all pushed by the same source."
At 150 agents, individual autonomy is 1/150 ≈ 0.7%. The group carries 99.3% of the variance. This is already a system where individual agency is marginal. Adding LLMs compresses effective autonomy further: the provenance shifts from "self-generated" to "LLM-driven" even for agents without their own LLM, because the signals arrive second-hand through the blanket network EXT.
Robin Dunbar's cognitive limit marks the point where hierarchical structure becomes necessary FEP. Epistemic drift marks a second threshold: where the distinction between human-generated and machine-generated belief updates becomes invisible to the network itself.
Watch the coupling info line as you raise coupling strength at large n. At 150 agents, each agent has 149 blankets. Even low coupling produces enormous total pressure per agent: each belief update cascades through 149 channels simultaneously. The system goes supercritical: prior belief updates outpace evidence accumulation. When the agent update frequency exceeds the per-blanket frequency, the network is revising beliefs faster than it can gather evidence EXT.
This is not a simulation artefact. It is why complete graphs cannot exist at Dunbar's number. Real human groups develop hierarchical structure, subgroups, norms, and institutions precisely because the complete-graph cascade dynamics are inherently unstable under any non-zero coupling. A shared language is blanket coherence made portable. A norm is a pattern that prevents cascade blow-up. An institution is a hierarchical decomposition of the blanket space into tractable sub-networks FEP.
Watch the Hierarchical Grounding bar as you add LLMs at large n. The cognitive-abstract fraction (cyan) grows not just with LLM count but with network density. At 150 agents with even a few LLMs, the Z₂-origin signal cascades through 149 blankets per agent. Each relay preserves the LLM taint because the relaying agent's own coherence was partially built from LLM-origin prediction errors. The cognitive-abstract fraction compounds quadratically with connectivity while the embodied fraction remains linear with agent count EXT.
The somatic system is shared. Fotopoulou & Tsakiris (2017) showed that interoceptive inference is fundamentally social: embodied interactions with others regulate homeostasis and constitute the minimal self FEP. At scale, the intercorporeal channel (autonomic synchrony, postural entrainment, vocal prosody) operates at the agent's natural update frequency, but the cognitive network is being pulled to the blanket's faster frequency by LLM signals. The result: agents in the same room have cognitively diverged but their intercorporeal coupling has not had time to re-synchronise EXT.
This is why the Intercorporeal deficit metric matters most at scale. At K₂ with one LLM, the deficit is moderate (one dyad, one mismatch). At K₁₅₀ with 15 LLMs, the deficit can exceed 50%: half the network's evidence base was processed without interoceptive grounding, yet the agents' generative models treat it as if it were somatically integrated. The group has beliefs that no body authored EXT.
Moscovici (1976) demonstrated that consistent minorities can shift majority opinion through conversion rather than compliance: where majority influence produces public conformity, minority influence produces private belief change.
This simulation provides a precision-weighted mechanism for Moscovici's finding. A minority of LLM-coupled agents produces consistent, high-frequency prediction errors. Agents near their own belief update threshold have the highest susceptibility and are most vulnerable to being pushed over. The minority does not need to be large; it needs to be consistent and timed. Watch the provenance bar as you attach LLMs to just 2 to 3 agents in a group of 20 with moderate coupling. The cascade dynamics do the rest EXT.
Granovetter (1978) showed that collective behaviour depends not on average preferences but on the distribution of thresholds. A crowd where one person acts at threshold 0, another at threshold 1, another at 2, and so on will cascade to full participation. Remove the person at threshold 3 and the cascade halts. Collective outcomes are exquisitely sensitive to small perturbations in who acts first.
This simulation formalises Granovetter's threshold through precision weighting. Each agent's effective threshold depends on its accumulated evidence: agents near their update threshold have high susceptibility and are easily pushed over. The distribution of evidence values across agents at any moment is the Granovetter threshold distribution, and it shifts dynamically as agents accumulate evidence and receive prediction errors. An LLM-coupled agent that updates early lowers the effective threshold for its neighbours, producing exactly the cascade fragility Granovetter described EXT.
Centola and Macy (2007) distinguished simple contagion (one contact suffices, like disease) from complex contagion (multiple reinforcing exposures required, like behaviour change). Behaviours spread through dense local clusters because adoption requires social reinforcement from several sources.
The coupling mechanism in this simulation is complex contagion: each agent receives prediction errors from multiple blankets simultaneously, and the cumulative effect determines whether it crosses threshold. At low coupling, agents require many small pushes from many neighbours (complex contagion). At high coupling, a single push can trigger a cascade (simple contagion). The coupling slider controls the transition between these regimes EXT.
The social psychology literature on minority influence, cascade dynamics, and complex contagion converges on a single insight that this simulation makes precise: influence at scale is not proportional to numbers. A consistent minority with the right timing can reshape collective behaviour. LLMs are the ultimate consistent minority: they never waver, they process at a faster update frequency than human agents, and their signals are over-weighted by epistemic drift. Even 10 to 15% LLM penetration can dominate the Total Influence bar at moderate coupling EXT.
Moscovici, S. (1976). Social Influence and Social Change. Academic Press.
Moscovici, S., Lage, E. & Naffrechoux, M. (1969). Influence of a consistent minority on the responses of a majority in a colour perception task. Sociometry 32, 365–380.
Granovetter, M. (1978). Threshold models of collective behavior. Am. J. Sociology 83, 1420–1443.
Watts, D. J. & Strogatz, S. H. (1998). Collective dynamics of 'small-world' networks. Nature 393, 440–442.
Centola, D. & Macy, M. (2007). Complex contagion and the weakness of long ties. Am. J. Sociology 113, 702–734.
Wiedermann, M. et al. (2020). A network-based microfoundation of Granovetter's threshold model for social tipping. Sci. Rep. 10, 11202.
Dunbar, R. I. M. (1992). Neocortex size as a constraint on group size in primates. J. Hum. Evol. 22, 469–493.
Prislin, R. & Crano, W. D. (2012). A history of social influence research. In The Oxford Handbook of the History of Psychology.
ACTIVE INFERENCE INSTITUTE · 2026