APAndrew Perfors

Andrew Perfors

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Andrew Perfors has a total of 87 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Types of Truth: How Perceived & Objective Truth Guide Belief Change

    by Keith Ransom, Bradley Walker, Piers Howe, Andrew Perfors, Yoshihisa Kashima, Nicolas Fay

    People must navigate socially transmitted information that varies from accurate to misleading or false. Here, we examined how perceived and objective truth predict belief change across three studies. In Study 1, four web-enabled frontier large language models (LLMs) and 200 human raters evaluated claims from two fact-checked datasets. LLMs outperformed humans in inter-rater agreement and discriminative sensitivity, supporting their use as a scalable proxy for objective truth. In Study 2 and 3, we applied this proxy to two large message corpora and examined its association with belief change. Perceived truth and LLM-derived objective truth independently predicted belief change. Moreover, objective truth moderated the association between perceived truth and belief change: the association between perceived truth and belief change was strongest for messages with lower objective truth and weakened as objective truth increased. These findings suggest that belief change reflects both subjective truth evaluations and the correspondence of information with reality.

  • Open Access

    Understanding online influence using a custom social media platform

    by Manikya Alister, Keith Ransom, Andrew Perfors

    Despite growing concerns about how social media use can affect people's beliefs and behaviour, estimating its causal impact is difficult. Here we present a novel method: a bespoke social media platform built on an isolated Mastodon server. Participants interacted much as they would on typical social media, but we controlled who could use the platform and measured their beliefs and attitudes before and after. This approximates the ecological validity of real social media data while retaining experimental control. Participants interacted over several days in three between-subject conditions (total N=311): a Control condition with no external influence, and two "troll" conditions Left and Right) where we covertly posted typical talking points from one political side. We observed extensive engagement in all conditions demonstrating the broader feasibility of our approach, and our minimal influence campaigns had statistically credible effects on engagement, beliefs, perceptions of others, and experience on the platform.

  • Cooperative goals must be selected by groups before tinkering begins

    by Francis Mollica, Andrew Perfors, Brett K. Hayes, Yoshihisa Kashima

    While the target paper’s proposal dovetails nicely with prominent perspectives on cultural evolution, there are also substantial gaps. In the commentary, we will argue that the target paper’s proposal does not adequately take the role of cooperative selection into account.

  • Open Access

    A rational cascade from teaching explains strategy discovery in development: The case of early addition

    by Merrick Giles, Andrew Perfors, Francis Mollica

    Without being taught, children learning addition discover strategies that are faster and less error-prone than those in their repertoire. Leading explanations of strategy discovery propose metacognitive mechanisms where children iteratively rewrite their existing strategies. Departing from this tradition, we propose that children make discoveries through a combination of reconstructive imitation during teaching, followed by rational inference under basic performance pressures. We instantiate our proposal with a model of Bayesian program induction with a prior for algorithmic simplicity and a likelihood that first encodes a pressure of mimicry (teaching phase), and then a pressure for speed and/or accuracy (individual practice phase). Regardless of whether the pressure is to respond faster or more accurately, the model robustly captures the long-observed developmental trajectory of strategy discovery.

  • Open Access

    The Impact of Social Learning on Learning Traps

    by Yanjun Liu, Alisha Shaikh, Ben R Newell, Andrew Perfors, Brett Hayes, Matthew Young

    People often fall into learning traps where a false belief about the structure of the environment leads to under-exploration of rewarding options. Five studies (N = 1342) examined whether observation of decisions made by another learner facilitated escape from a trap or increased trap prevalence. In studies of horizontal social learning (Experiments 1a, 1b, 2), participants first had private experience with a task where approach of different category members could lead to gains or losses. We identified whether participants had learned an optimal two-dimensional categorization rule or fallen into the trap of using a one-dimensional rule. Participants then observed the approach decisions of another learner using the same or a different category rule. A substantial proportion of those who had initially fallen into the trap shifted to the optimal rule after observing use of an alternate rule. This escape from a learning trap was found following observation of both optimal (Experiment 1a) and sub-optimal rules (Experiment 1b, 2). In vertical social learning (Experiments 3, 4), social observation of approach decisions took place before private learning. Such observation led to reduced trap prevalence when the social demonstrator modeled an optimal rule but increased trap prevalence when they modeled a one-dimensional rule. These effects generalized to a learning environment with the same underlying feature structure but different feature-category mappings. These results both offer a novel approach to remediating learning traps as well as highlighting how poorly coordinated social learning can intensify traps.

  • When a helpful bias is unhelpful: Limitations in reasoning about random and deliberately misleading evidence.

    by Manikya Alister, Keith J. Ransom, Andrew Perfors

    Social information aids learning: By making assumptions about other people's knowledge and intentions, people can draw strong and accurate inferences from limited data. In this study, we systematically tested people's ability to reason from information providers with different intentions. The task was an adaptation of Shafto et al.'s (2014) rectangle game, where learners guessed a rectangle's size and location based on provided clues. We examined reasoning based on information from four types of providers: a helpful provider, a provider who sampled randomly, and two misleading providers (who could mislead but not lie). We also varied whether people were given a cover story describing the provider in advance or whether they could infer how helpful a provider was based on what the provider shared. Participants learned efficiently from helpful providers, aligning closely with the predictions of a normative Bayesian model, even without a cover story. However, while people usually recognized unhelpful providers, they struggled to identify and respond appropriately to misleading strategies. Overall, our results suggest a helpful bias: In our task, participants assumed helpful intent unless given explicit feedback, and even then, they did not fully adjust in line with Bayesian predictions. People also struggled to overcome this bias when learning from randomly sampled information, especially when they had experience being an information provider themselves (Experiment 3). (PsycInfo Database Record (c) 2026 APA, all rights reserved).

  • Open Access

    Core Vocabulary Reveals Differences Between Human Word Prediction and Large Language Models

    by Andrew Wang, Simon De Deyne, Meredith McKague, Andrew Perfors

    The question of which words are the most central or important to a language has been explored in various ways. In this study, we propose definitions of core vocabulary that are based on how language is learned, represented, and processed from psychological perspectives, and test these on a word prediction task. We aim to (1) compare core vocabulary based on word frequency in natural language, word association network centrality, and age-of-acquisition in terms of how well they are guessed in word prediction contexts, and (2) investigate the extent to which word prediction in large language models aligns with humans, and if there are systematic differences between them, whether these can be captured by core vocabulary measures. Across two experiments, 867 participants completed a task which involved guessing target words that were missing from sentence contexts. Word frequency-based core words were easier to guess overall, but when the predictability of words was taken into account, word association- and acquisition-based core words were easier to predict, indicating that people have a preference for simpler or more basic words over and above distributional predictability. Additionally, we show that there are systematic deviations between the predictions of a large language model (BERT) and humans, and that these are best captured by word association-based coreness. The findings suggest that distributional relationships between words in text is not all there is to human word prediction, but that people may also rely on factors like communicative usefulness and multimodal or extralinguistic information. Implications for core vocabulary, word prediction, and the relationship between large language models and human cognition are discussed.


  • Open Access

    Misinformation Spread in Transmission Chains

    by Toby Prike, Bradley Walker, Nicolas Fay, Andrew Perfors, Yoshihisa Kashima, Ullrich K. H. Ecker

    We examined peer-to-peer transmission of misinformation and corrections, an understudied mechanism of misinformation spread. Participants (n = 120 per wave; total N = 600) viewed and wrote summaries of four news articles (conditions: control, misinformation-only, weak correction, strong correction). These summaries were passed down four subsequent waves of participants, each wave summarizing for the next. Summaries were coded for the presence of misinformation and correction details, and misinformation reliance. Initially (position one), both weak and strong corrections reduced misinformation reliance, although strong corrections were more effective. However, both misinformation and correction details were lost during transmission, leading to increased misinformation reliance in the correction conditions. By position five, there was no significant difference in misinformation reliance between the weak correction and misinformation-only conditions, showing that corrections can “lose out” to misinformation during peer-to-peer transmission and highlighting the importance of using detailed corrections to prevent misinformation reliance re-emerging.

  • Open Access

    Reasoning from social information enables efficient human intelligence

    by Manikya Alister, Andrew Perfors

    Humans have a remarkable capacity for efficient reasoning. We achieve a level of general intelligence far beyond that of any other species, and although artificial agents now exhibit many impressive, human-like abilities, they still require orders of magnitude more data than a human encounters in a lifetime. What, then, underlies our extraordinarily powerful yet efficient cognitive abilities? We examine one possibility: how people use social information to learn and reason in everyday life. We suggest that social reasoning directs us toward relevant information, enables inferences about others’ minds that support learning from minimal data, and provides cues for quickly assessing the accuracy of information. We also discuss how this necessary reliance on social reasoning leaves us vulnerable to exploitation, particularly in modern information environments such as social media. We conclude by considering large language models as an emerging challenge for social reasoning: the same cognitive tendencies that make us efficient social learners may leave us uniquely vulnerable to exploitation by these systems, yet theories of social reasoning also offer our best framework for anticipating and mitigating these risks.