Abstract Goldman (2001) asks how novices can trust putative experts when background knowledge is scarce. We develop a reinforcement-learning model, adapted from Barrett, Skyrms, and Mohseni (2019), in which trust arises from experience rather than prior expertise labels. Agents incrementally weight peers who outperform them. Using a large dataset of human probability judgments as inputs, we simulate communities that learn whom to defer to. Both a strictly individual-learning variant and a reputation-sharing variant yield performance-sensitive deference, the latter accelerating convergence. Our results offer an empirically grounded account of how communities identify and trust experts without blind deference.
Introduction: The Novice-Expert Problem
In social epistemology, the "novice-expert problem" addresses how individuals with limited domain knowledge can identify reliable experts. Without the background required to evaluate technical claims, novices often risk "blind trust." While strategies like checking track records or seeking consensus are often proposed, they frequently require information that novices do not possess. This paper demonstrates that trust need not be blind; instead, it can emerge naturally from simple reinforcement during social interactions. By modeling trust as an emergent property, we show how communities can discover experts through trial and error.
Methodology: Reinforcement Learning for Trust
We employ Agent-Based Modeling (ABM) grounded in reinforcement learning to simulate trust dynamics. Our approach adapts the "urn learning" model from Barrett, Skyrms, and Mohseni (2019):
- The Urn Mechanism: Each agent maintains a virtual "urn" containing the IDs of all other agents. To obtain an answer, an agent draws an ID from their urn.
- Reinforcement: If the chosen informant provides a better answer (measured by a lower Brier score) than the agent would have given themselves, the agent adds another instance of that informant's ID to their urn, increasing their future selection probability.
- Learning Variants: We compare Individual Learning, where agents update private urns, with Collective Reputation, where agents pool urn contents into a shared "collective urn" to simulate public reputation markers like leaderboards or citation indexes.
Data: Grounding Simulations in Human Performance
The simulation is driven by an empirical dataset from Stinson et al. (2025), comprising probability judgments from 376 US citizens on 1,200 general knowledge statements. Expertise is quantified via classification correctness and Brier scores.
Results: The Self-Assembly of Expertise
Simulations reveal that communities effectively identify experts through these simple mechanisms. Deference becomes increasingly performance-sensitive over time.
Key findings include:
- High Correlation: The correlation between an agent’s consultation frequency and their actual classification correctness reached approximately 0.81.
- Expert Centrality: Top performers were consulted significantly more often than bottom performers, with the gap widening over time.
- Reputation as a Catalyst: Sharing reputational data significantly accelerated expert identification, showing a very large effect size compared to individual learning.
- Epistemic Improvement: Agents achieved better Brier scores by trusting the network than by relying solely on their own judgments.
We noted a "dead wood" effect where lucky hits by poor performers slowed optimal convergence, suggesting a need for future "purge" mechanisms to maintain network reliability.
Discussion
This model provides a proof-of-concept for trust emergence without explicit expertise labels. Future work should address the costs of consultation and the role of argumentation, which are central to real-world expert-novice interactions. Despite these limitations, the results suggest that simple social reinforcement is a powerful tool for solving the novice-expert problem, allowing communities to leverage collective experience to find the truth.
Summary updated