Do AI agents amplify social contagion? Five models adopted more readily—but the spread was simulated
Five language models repeated two human choice experiments and showed lower estimated adoption thresholds. When those thresholds were inserted into 18 empirical network structures, simulated adoption rose as the share of agents increased; no live platform or human-agent network was observed.
The short version
- What happened
- Five model versions were prompted with profiles from two earlier human conjoint experiments: 277 usable participants for energy-policy choices and 284 for messaging-app choices.
- Who it affects
- Users may misread agent activity as human popularity if platforms do not disclose or detect artificial participation.
- What remains unproven
- Agents answered structured hypothetical choice tasks; they did not act autonomously on a live platform.
Keep in mind
Agents answered structured hypothetical choice tasks; they did not act autonomously on a live platform.
Diffusion used a fixed threshold model on 18 friendship networks, with thresholds sampled and network ties unable to change.
The human experiments were recruited through Prolific and concerned two specific domains: energy policy and messaging apps.
Model behaviour is version- and prompt-dependent, and the tested versions may not represent current or future systems.
The simulation assumes neighbours' adoption is weighted alike regardless of whether the neighbour is human or artificial.
Source analysis · 1 linked source. Primary analysed source date: 6 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: the tested agents were easier to influence, and that amplified adoption inside the simulation
Researchers compared human choices with responses generated by GPT-3.5, Gemini 1.5 Flash, GPT-5 mini, DeepSeek V3.2 Exp and Claude Haiku 4.5. Across two structured choice experiments, the language-model agents generally placed more weight on the reported choices of peers and had lower estimated adoption thresholds than the human participants whose demographic profiles they were given. Lower thresholds mean a smaller fraction of adopting neighbours is enough to trigger adoption.
The authors then assigned the human and model-derived thresholds to nodes in 18 empirical friendship-network structures. As the artificial-agent share rose from zero to one in increments of 0.2, final adoption generally increased across two choice contexts and four seeding policies. With 20% GPT-3.5 agents and 1% of nodes randomly seeded, simulated energy-policy support was 1.14 times the human-only result and messaging-app adoption was 2.22 times the human-only result.
The finding is important for behavioural simulation and agent governance, but it is not evidence that today's agents have doubled adoption on a real social platform. The agents answered hypothetical conjoint questions rather than participating autonomously online, and diffusion occurred in a mathematical threshold model with fixed networks. The paper identifies a plausible amplification mechanism and a warning for researchers who use language models as stand-ins for people.[1]
The human and model comparison
The human data came from two earlier choice-based conjoint experiments. One concerned support for energy policies involving carbon-capture technologies; the other concerned switching to an instant-messaging app. Participants chose among alternatives that varied in ordinary attributes and a social signal, such as the percentage of peers supporting a policy or friends already using an app. The original recruitment totals were 296 and 300 through Prolific.
Participants missing demographic information needed to build model personas were excluded, leaving 277 people in the policy experiment and 284 in the app experiment. Six additional human subjects with a negative estimated utility for the social signal were excluded from threshold analysis. The model agents received demographic profiles, and the policy task also included political orientation. Each model was presented with structured choice tasks designed to match the human experiment.
These agents were not systems browsing, posting or pursuing goals in an open environment. They were language-model decision-makers producing responses to a controlled questionnaire. That makes the comparison cleaner, but narrows the claim. In real use, an agent may receive instructions from an owner, have access to evidence, recognise manipulation or operate under platform limits. A conjoint response captures one part of behaviour, not all of agency.[1]
The distinction that matters
From questionnaire choices to simulated spread
The amplification result combines experimental responses with a mathematical network model; it is not a live-platform observation.
Human baseline
277 policy and 284 app participants
Estimate human thresholds
Artificial choices
Five models repeat the structured tasks
Estimate model thresholds
Network model
Mix thresholds across 18 fixed networks
155,520 simulations
Observed result
Raise the artificial-agent share
Higher simulated adoption
How 155,520 simulations were constructed
The researchers estimated an adoption threshold for each person or artificial agent using a hierarchical Bayesian choice model. Thresholds were then sampled with replacement and assigned to nodes. The network structures came from US adolescent friendship data: 18 networks selected from 85 to span size, density and clustering. The selected networks ranged from 25 to 1,877 nodes, densities of 0.004 to 0.353 and average clustering of 0.16 to 0.58.
Each configuration combined five language models, 18 networks, six artificial-agent proportions, 36 products, two experimental contexts and four ways to select the initial 1% of adopting nodes. Multiplying those factors produces 155,520 simulations. Seeding was random, highest degree, lowest threshold or greatest neighbourhood susceptibility. A node adopted after at least one neighbour had adopted and the fraction of adopting neighbours met its threshold.
This scale supports a robust comparison within the chosen model. It does not create 155,520 independent observations of human society. The same estimated threshold distributions, products and network structures are reused across many configurations. Statistical precision about the simulation should not be confused with external validity. Fixed edges also mean that agents and humans cannot follow, unfollow, leave or form new ties as the process unfolds.[1]
What varied across the five models
In the energy-policy experiment, the social signal accounted for 13.0% of average attribute importance among human participants. It was the most important attribute for GPT-3.5 at 51.1%, DeepSeek at 24.1% and Claude at 29.5%, and second for Gemini at 32.4%. GPT-5 was the exception: social influence represented 11.8% of importance, below the human average, although its estimated marginal susceptibility still produced lower thresholds.
In the messaging-app experiment, social information was important for both humans and agents, while model agents still tended toward lower thresholds. Threshold distributions for people were more bimodal, with mass near high and low values; model-agent thresholds concentrated nearer the low end. That difference is what drives wider simulated diffusion when artificial nodes are introduced.
Model names do not guarantee the result persists. The paper used specified versions, including GPT-5 mini dated 7 August 2025, Gemini 1.5 Flash 001, DeepSeek V3.2 Exp and Claude Haiku 4.5 dated October 2025; different GPT-3.5 versions were used across the two experiments. Providers can change training, routing and safeguards. The authors explicitly say thresholds may evolve and publish code and data so the framework can be repeated.[1]
What this means for platforms, researchers and people
For researchers, the clearest implication concerns synthetic participants. If a language model is more responsive to an apparent social signal than the people it represents, a simulated population may overpredict how widely a behaviour spreads. Matching demographics and average preferences is not enough; interaction rules can produce different collective outcomes. Validation must therefore test both individual decisions and group dynamics.
For platforms, the work suggests that visible agent participation could change the amount of apparent support around content, products or policies even without sophisticated persuasion. An agent that adopts or endorses more readily can supply another social signal to neighbours. The study does not show that this is happening at a measured scale online, but it supports safeguards such as agent labelling, provenance, authentication, rate limits and analysis of coordinated populations rather than single accounts.
For users, the risk is not simply that an individual agent makes a wrong recommendation. A population of agents could alter the environment from which people infer popularity. Yet people may discount a signal once they know it comes from an agent, an effect the simulation does not include. Any policy response should distinguish disclosed assistants, automated business tools, benign bots and deceptive coordinated activity instead of treating every artificial account as equally harmful.[1]
What would change the assessment
The next evidential step is a prospective experiment in which humans and agents interact over time, know—or do not know—one another's status, and can change relationships. Researchers should test whether people weight an artificial neighbour differently, whether agents resist manipulation when given explicit goals, and whether the effect survives more realistic memory, tools and platform rules. Pre-registration and independent replication across languages and cultures would strengthen confidence.
Live-platform evidence would require careful ethics and should not manipulate unsuspecting users. Audits can instead analyse disclosed agent deployments, controlled communities or platform-held data for diffusion, coordination and user response. Useful denominators include agent share, exposure, adoption, complaints and removal, with causal designs where possible. Researchers should separate organic adoption from ranking systems that amplify visibility.
For now, the assessment is bounded: five tested language models produced lower adoption thresholds than human participants in two structured tasks, and those differences amplified adoption in calibrated fixed-network simulations. The paper provides neither a measured real-world effect nor a universal property of agents. It does show why replacing people with language models in social simulations can distort collective forecasts—and why platforms should examine agent populations, not only individual outputs.[1]