Across 27 datasets of students learning online, we find an astonishing regularity in the rate they learn. That’s the good news. When given quality practice opportunities with feedback and as-needed instruction, all students learn, making quite similar progress per opportunity. The bad news is we find a wide range in their general performance indicating inequities in the preparation they get prior to course entry. Pursing the hypothesis that these achievement gaps result from opportunity gaps, we have implemented a hybrid human-computer tutoring approach to support disadvantaged math students in getting needed learning opportunities. Results of a quasi-experiment spanning the pandemic year of 2020 show a promising doubling of math learning for students receiving hybrid tutoring relative to demographically matched students in the same urban schools.

The embodiment turn in the learning sciences has fueled growth of multimodal learning analytics to understand embodied interactions and make consequential educational decisions about students more rapidly, more accurately, and more personalized than ever before. Managing demands of complexity and speed is leading to growing reliance by education systems on disembodied artificial intelligence (dAI) programs, which, ironically, are inherently incapable of interpreting students’ embodied interactions. Augmented intelligence systems offer promising avenues by integrating the strengths of omnipresent dAI to detect complex patterns of student behavior from multimodal datastreams, with the strengths of humans to meaningfully interpret embodied interactions in service of consequential decision making to achieve a balance between complexity, interpretability, and accountability for allocating education resources to children. Few such systems exist, however, and there is still much to consider about how, when, and by what criteria we strive to balance human and dAI roles in educational contexts.

Social epistemic pathologies plague our society. We perpetually find polarization, pluralistic ignorance, the spread of fake news, online mobbing, and others. Some scholars attribute these social pathologies to individual irrationality. Fake news spreads because people are not careful consumers of news. Polarization occurs because of irrational attachments to political positions. And so on. In this talk, I will argue that there may be purely social pathologies; epistemic problems that reside entirely at the “group level.” I will do this through a case study of pluralistic ignorance, a phenomenon studied in social psychology, business, philosophy, and political science. When in a state of pluralistic ignorance, every person in a community will disavow a private belief they hold because they think others feel differently. Pluralistic ignorance is consistent with individual rationality, I argue, and it can arise in completely rational communities. I conclude by discussing what the possibility of group level epistemic pathologies means for how we address collective problems around knowledge and belief.

Written science explanations are central to learning and practicing science. However, students struggle to explain their ideas and use supporting data appropriately. Further, it is not realistic for teachers to provide real-time, comprehensive feedback to each student. Natural language processing (NLP) technologies can provide timely, personalized, automated feedback. In this presentation, I will discuss results from our project in which we are working on integrating automated feedback generated by an NLP technology and the support provided by teachers, to help middle school students write science essays. In addition to the automated feedback to students, we also provided summary assessments of students’ essays to teachers, which they used to guide instruction. I will discuss challenges and opportunities for using automated feedback in classrooms, and the synergy between support provided by the technology and the teacher.

The increased use of algorithmic predictions in sensitive do­mains has been accompanied by both enthusiasm and concern. To understand the opportunities and risks of these technologies, it is key to study how experts alter their decisions when using such tools. In this talk, I will present a study of the adoption of an algorithmic tool used to assist child maltreatment hotline screening decisions. By taking advantage of an implementation glitch, we investigate corrective overrides: whether decision makers are more likely to override algorithmic recommendations when the tool misestimates the risk score shown to call workers. We find that, after the deployment of the tool, decisions became better aligned with algorithmic assessments, but human adherence to the tool's recommendation was less likely when the displayed score was misestimated as a result of the glitch. Then, analyzing the effect of adoption and overrides on racial and socioeconomic disparities, we find that the deployment of the tool did not affect disparities with respect to the pre-deployment period. We also observe that the disparities resulting from algorithmic-informed decisions were substantially smaller than those associated with the algorithm in isolation. Together, these results make a case for the value of humans in-the-loop, showing that in high-stakes contexts, human discretionary power can mitigate the risks of algorithmic errors and reduce disparities.

1. Fogliato, Riccardo and De-Arteaga, Maria and Chouldechova, Alexandra, A Case for Humans-in-the-Loop: Decisions in the Presence of Misestimated Algorithmic Scores.

Individuals are often faced with temptations that can lead them astray from long-term goals. We're interested in developing interventions that steer individuals toward making good initial decisions and then maintaining those decisions over time. In the realm of financial decision making, a particularly successful approach is the prize-linked savings account: individuals are incentivized to make deposits by tying deposits to a periodic lottery that awards bonuses to the savers. Although these lotteries have been very effective in motivating savers across the globe, they are a one-size-fits-all solution. We investigate whether customized bonuses can be more effective. We formalize a delayed-gratification task as a Markov decision problem and characterize individuals as rational agents subject to temporal discounting, a cost associated with effort, and fluctuations in willpower. Our theory is able to explain key behavioral findings in intertemporal choice. We created an online delayed-gratification game in which the player scores points by selecting a queue to wait in and then performing a series of actions to advance to the front. Data collected from the game is fit to the model, and the instantiated model is then used to optimize predicted player performance over a space of incentives. We demonstrate that customized incentive structures can improve an individual's goal-directed decision making.

1. Sukumar, S., Ward, A. F., Elliott-Williams, C., Hakimi, S., & Mozer, M. C. (2022). Overcoming Temptation: Incentive Design For Intertemporal Choice. arXiv preprint arXiv:2203.05782.

For about 50 years, since the “social function of intellect hypothesis”, evolutionists have suspected that some form of collective intelligence is the secret of human success. In this talk I will outline a contemporary variant of this view, known as cultural evolutionary psychology or the cognitive gadgets theory. This variant is more fully embedded in cognitive science than any other. Drawing on evidence from comparative, developmental and cognitive psychology / neuroscience, it suggests that distinctively human cognitive mechanisms are assembled from old parts during childhood, and, at the population level, that they have been shaped by cultural selection – a Darwinian process operating on socially inherited, rather than genetically inherited, variant mechanisms. Genetic evolution merely provided a starter kit for human cognitive evolution, tweaking attentional, motivational, and domain-general learning mechanisms to make hominins more malleable by their social environments. I will indicate the kinds of evidence that support this view and contrast the implications of cognitive gadgets with those of the “collective brain” account of human cognitive evolution.

1. Heyes, C. (2020). Psychological mechanisms forged by cultural evolution. Current Directions in Psychological Science, 29(4), 399-404

2. Heyes, C. (2019). Précis of cognitive gadgets: The cultural evolution of thinking. Behavioral and Brain Sciences, 42.

There's a lot of talk about the metaverse, but a surprising lack of research on what happens when people wear immersive VR headsets and come together as avatars. This talk discusses the trials and tribulations of social interaction in VR, focusing on a year-long study which immersed over 400 students in small groups in VR for ten weeks at a time, examining the effects of self-representation and environmental context on nonverbal behavior, speech, performance, and self-report measures longitudinally in a dataset of over 300,000 shared minutes in VR. In addition to discussing findings related to self and context, I discuss what the metaverse is good for--and what it is not.

People are known to judge artificial intelligence using a utilitarian moral philosophy and humans using a moral philosophy emphasizing perceived intentions. But why do people judge humans and machines differently? Psychology suggests that people may have different mind perception models for humans and machines and, thus, will treat human-like robots more similarly to how they treat humans. Here we present a randomized experiment where we manipulated people's perception of machines to explore whether people judge more human-like machines more similarly to how they judge humans. We find that people's judgments of machines become more similar to that of humans when they perceive machines as having more agency (e.g., ability to plan and act) but not more experience (e.g., ability to feel). Our findings indicate that people's use of different moral philosophies to judge humans and machines can be explained by a progression of mind perception models where the perception of agency plays a prominent role. These findings add to the body of evidence suggesting that people's judgment of machines becomes more similar to that of humans motivating further work on differences in the judgment of human and machine actions.

1. Zhang, J., Conway, J., & Hidalgo, C. A. (2022). Why do people judge humans differently from machines? The role of agency and experience. arXiv preprint arXiv:2210.10081.

Technologies offer many possibilities for supporting ambitious learning practices. Ambitious Learning practices refer to pedagogical approaches that are student-centered, collaborative, and require extended engagement with meaningful problems or questions, such as problem-based learning or other forms of collaborative inquiry (Glazewski & Hmelo-Silver, 2019). However, this requires thinking broadly about what functions technologies need to afford for learning in these complex environments (Jeong & Hmelo-Silver, 2016). In this talk, I will present a framework for thinking about affordances of technology for these kinds of complex learning environments.

‘Crowd wisdom’ refers to the accuracy that can be attained by averaging judgments of individuals working alone. However, independence is unusual; people often discuss and collaborate in groups. When does group interaction improve judgment accuracy relative to averaging the group's independent answers? Two large laboratory studies explored the effects of 969 face-to-face discussions on the judgment accuracy of 211 teams facing numeric estimation problems ranging from geographic distances to historical dates to stock prices. Although participants expected discussions to make their answers more accurate, actual effects of group interaction on accuracy were mixed. A novel, group-level measure of collective confidence calibration robustly predicted when discussion helped or hurt accuracy relative to the group's initial independent estimates. When groups were collectively calibrated prior to discussion, with more accurate members being more confident in their own judgment and less accurate members less confident, subsequent group interactions were likelier to lead to increased accuracy. Collective calibration predicts improvement because groups typically listen to their most confident members. When confidence and knowledge are positively associated within a group, the group's most knowledgeable members are more likely to influence a group's discussion.

1. Silver, I., Mellers, B. A., & Tetlock, P. E. (2021). Wise teamwork: Collective confidence calibration predicts the effectiveness of group discussion. Journal of Experimental Social Psychology, 96, 104157.

Talk of group coordination often brings to mind group members synchronizing their behaviors, as with soldiers marching in lockstep or audience members spontaneously clapping in unison. However, effective coordination often crucially involves a division of labor in which different members come to adopt different roles. I will describe three experimental paradigms for exploring the formation of specialized roles in online, interactive groups: “Group Binary Search”, “Battles of the Exes”, and “Find the Unicorn”. In all three paradigms, role specialization can be operationalized as members becoming consistent in their own behavior and differentiated from the other members. The computational models that fit human behavior the best and also perform the best have mechanisms to strategically differentiate a player’s behavior from others. Accordingly, we have developed the CARMI framework which characterizes role specialization processes in terms of: Communication, Adaptation to feedback, Repulsion, Multi-level planning, and Intention modeling.

1. Goldstone, R. L., Andrade-Lotero, E., Hawkins, R. D., & Roberts, M. E. (in press). The emergence of specialized roles Within groups. Topics in Cognitive Science.

2. Andrade-Lotero, E., & Goldstone, R. L. (2021). Self-organized division of cognitive labor. PLoS ONE, 16(7): e0254532.

3. Hawkins, R. X. D., & Goldstone, R. L. (2016). The formation of social conventions in real-time environments. PLoS One, 11(3): e0151670.

4. Roberts, M. E., & Goldstone, R. L. (2011). Adaptive Group Coordination and Role Differentiation. PLoS One, 6, 1-8.

Collective emotions are an elusive research topic due to the difficulty to study them with traditional methods that focus on individual behavior and emotional reactions. Recent approaches to collective emotions leverage the accessibility of social media data to study emotions through their social manifestation in online communication. This is possible thanks to the application of natural language processing methods that measure the subjective expression of emotions in social media text. I will present our work developing LEIA, an AI method for the analysis of social media text that has been trained with more than x million posts with subjective emotion annotations. LEIA's performance in out-of-domain tests is opening up a new range of possibilities for the study of emotions in digital traces. Then I will present how we are applying this kind of method to track the temporal evolution of emotions at the scale of whole societies, validating our measurements against representative surveys of emotional experiences.

1. Pellert, M., Metzler, H., Matzenberger, M., & Garcia, D. (2021). Validating daily social media macroscopes of emotions. arXiv preprint arXiv:2108.07646.

2. Garcia, D., Pellert, M., Lasser, J., & Metzler, H. (2021). Social media emotion macroscopes reflect emotional experiences in society at large. arXiv preprint arXiv:2107.13236.

Collaborative student learning has been shown to lead to significant academic benefits among students, and to improved social skills that are critical for the workforce, such as communication and teamwork. However, these benefits were limited to small face-to-face groups and required the support of human experts who actively monitored and guided the group’s learning. Technological advances now enable globally dispersed teams to collaborate online, from Q&A forums to virtual laboratories. Augmenting these settings with AI technology can scale up the benefits of collaborative group learning to online groups. I will describe challenges to AI in Ed research for supporting this new type of online teamwork, as well as opportunities for combining AI and learning analytics towards supporting students’ learning and teachers’ understanding of how students learn.

1. Shusterman, Einat, et al. "Seeding Course Forums using the Teacher-in-the-Loop." LAK21: 11th International Learning Analytics and Knowledge Conference. 2021.

2. Yogev, Eran, et al. "Classifying and visualizing students' cognitive engagement in course readings." Proceedings of the Fifth Annual ACM Conference on Learning at Scale. 2018.

3. Segal, Avi, et al. "Keeping the teacher in the loop: Technologies for monitoring group learning in real-time." International Conference on Artificial Intelligence in Education. Springer, Cham, 2017.

1. Frey, S., Zhong, Q., Bulat, B., Weisman, W. D., Liu, C., Fujimoto, S., ... & Schweik, C. M. (2022). Governing online goods: Maturity and formalization in Minecraft, Reddit, and World of Warcraft communities. arXiv preprint arXiv:2202.01317.

2. Frey, S., & Sumner, R. W. (2019). Emergence of integrated institutions in a large population of self-governing communities. PloS one, 14(7), e0216335.

The integration of AI into classrooms is filled with promise and, yet, is still in its infancy. In order for AI to enact transformative change for education, it is essential that we use leading theories of learning when designing new AI technologies so as not to fall back on behaviorist and cognitivist approaches to education (e.g., flash cards or behavior management solutions). I will discuss a new model for Research-Practice-Industry Partnerships (RPIP), a cross-sector, co-design method of research and development that encourages the design of new AI that allows for the complexities of the learning environment. Findings from a recent RPIP on the co-design and initial testing of Symphony Classroom—an AI hub, remote control, and AI software platform built for educational settings—led to significant reductions in teacher technostress (i.e., stress related to technology use), and significant increases in time for teaching and learning. This is representative of how we can reshape our design process as well as use insights derived from the learning sciences to transform the educational landscape at scale.

Successful team performance requires individuals to effectively decide how and when to act, with robust decision-making often differentiating expert from novice performance. I will present recent research demonstrating how cutting-edge machine learning and explainable-AI techniques can not only be employed to model and accurately predict human decision-making during team behaviour but can also help identify the information that best explicates expert task performance. Motivated by the increasing need to develop artificial systems capable of safe and robust human interaction, I will also detail how models of human decision-making can be employed to control the decision-making dynamics of interactive artificial agents and create AI systems that can anticipate, prevent, or counteract human performance errors.

Artificial intelligence (AI) and machine learning models are being increasingly deployed in real-world applications. In many of these applications, there is strong motivation to develop hybrid systems in which humans and AI algorithms can work together, leveraging their complementary strengths and weaknesses. In the first part of the presentation, I will discuss results from a Bayesian framework where we statistically combine the predictions from humans and machines while taking into account the unique ways human and algorithmic confidence is expressed. The framework allows us to investigate the factors that influence complementarity, where a hybrid combination of human and machine predictions leads to better performance than combinations of human or machine predictions alone. In the second part of the presentation, I will discuss some recent work on AI-assisted decision making where individuals are presented with recommended predictions from classifiers. Using a cognitive modeling approach, we can estimate the AI reliance policy used by individual participants. The results show that AI advice is more readily adopted if the individual is in a low confidence state, receives high-confidence advice from the AI and when the AI is generally more accurate. In the final part of the presentation, I will discuss the question of “machine theory of mind” and “theory of machine”, how humans and machines can efficiently form mental models of each other. I will show some recent results on theory-of-mind experiments where the goal is for individuals and machine algorithms to predict the performance of other individuals in image classification tasks. The results show performance gaps where human individuals outperform algorithms in mindreading tasks. I will discuss several research directions designed to close the gap.

1. Steyvers, M., Tejeda, H., Kerrigan, G., & Smyth, P. (2022). Bayesian modeling of human–AI complementarity. Proceedings of the National Academy of Sciences, 119(11), e2111547119.

Collaborative problem solving (CPS) has been receiving increasing attention throughout the world because much of the planning, problem solving, and decision making in the modern world is performed by teams with members who often have different expertise. However, students and citizens do not receive systematic science-based education and training on CPS skills in schools and the workforce. CPS is different from other collaboration contexts, such as brain-storming, learning, and decision making, because there is visible accountability on progress on whether attempted solutions achieve the goals. This presentation reports the framework and some results of the Programme for International Student Assessment (PISA, 2015) that included 52 countries that assessed collaborative problem solving skills. This was the first international assessment of CPS. There will also be highlights of follow-up projects after the PISA assessment.

A key question individuals face in social environments is when to innovate alone and when to imitate others. Previous theoretical analyses and simulations have found that the best performing groups exhibit an intermediate balance, yet it is still largely unknown how individuals collectively negotiate this balance. We use an immersive collective foraging experiment, implemented in the Minecraft game engine, to provide unprecedented access to spatial trajectories and visual field data. The virtual environment imposes a limited field of view, creating a natural trade-off between allocating visual attention towards individual search or to look towards peers for social imitation. By analyzing foraging patterns, social interactions (visual and spatial), and social influence, we shine new light on how groups collectively adapt to the demands of the environment through specialization and selective imitation, rather than homogeneity and indiscriminate copying of others. I will present new hierarchical Bayesian modeling results that predict foraging behavior as a function of both individual and social features.

For science to influence policy making it has to be made accessible to stakeholders. This in turn requires the aggregation of disparate pieces of knowledge from the scientific literature into a coherent and concise form, often called a "consensus document". I describe two cases studies of consensus-document creation based on a variety of different techniques. The first case study involves an expert elicitation and describes the process underlying the production of the Debunking Handbook 2020 (https://sks.to/db2020). The second case study involves the production of a Manifesto for Science Communication as Collective Intelligence (https://www.scibeh.org/manifesto/), which was itself created by using a number of voting procedures and machine tools for knowledge aggregation, such as pol.is. I conclude by sketching a few paths to more effective knowledge integration.

1. Lewandowsky, S., Cook, J., Ecker, U., Albarracin, D., Amazeen, M., Kendou, P., ... & Zaragoza, M. (2020). The debunking handbook 2020.

2. Lewandowsky, S., Cook, J., Ecker, U. K. H., & Newman, E. J. (2021). Under the Hood of The Debunking Handbook 2020: A consensus-based handbook of recommendations for correcting or preventing misinformation.

3. Holford, D., Fasce, A., Tapper, K., Demko, M., Lewandowsky, S., Hahn, U., ... & Wulf, M. (2022). A manifesto for science communication as collective intelligence.

4. Lewandowsky, S., Cook, J., Schmid, P., Holford, D. L., Finn, A., Leask, J., ... & Vraga, E. K. (2021). The COVID-19 vaccine communication handbook. A practical guide for improving vaccine communication and fighting misinformation.

I will describe a conceptual framework that my colleagues and I have recently developed for studying collective adaptation: the process of iterative co-adaptation of cognitive strategies, social environments, and problem structures. Going beyond searching for “intelligent” collectives, we integrated research from different disciplines to show how collective adaptation perspective can help explain why similar collectives can follow very different and sometimes counter-intuitive trajectories. In this talk, I will discuss how this perspective helps to explain why successful collectives appear to have a general collective intelligence factor, why collectives rarely optimize their behavior for a single problem, why their behaviors can appear myopic, and why playful exploration of alternative social systems can be useful. I will describe different approaches for the study of collective adaptation, including computational models inspired by evolution and statistical physics. The framework of collective adaptation enables the integration and formalization of knowledge about human collective phenomena and opens doors to a more rigorous, transdisciplinary pursuit of important outstanding questions about human sociality.

Herb Simon (1991) said that all learning occurs inside individual human minds, suggesting that groups and organizations can only learn by the learning of their members or by acquiring new members that have new knowledge. In the past decade we have pursued research that follows in Simon’s footsteps, explaining how collective learning emerges from individual decisions. Furthermore, our need to learn not only with other humans but along with adaptive technology has expanded exponentially since Simon presented his initial ideas. This has created a demand for new models and algorithms that can understand and adapt to human behavior. We increasingly rely on algorithms to augment human abilities and to help coordinate the collective decisions of humans and technology. In this talk, I will present our theory of individual experience-based choice and how this theory is being used to investigate collective human-to-human and human-machine learning. Using snippets of our research on contexts such as social dilemmas, optimal stopping sequential tasks, gridworld navigation tasks, and adaptive cyber defense, I will explain how individual learning models can be used in dyadic, group and network-level interactions. This presentation will illustrate the challenges and potential involved in conducting research on collective human-machine interactions.

Why do people seek the particular information they seek in everyday learning environments, and how can we improve human learning with machine feedback? We examine this question by introducing a new empirical paradigm for studying naturalistic active learning, as well as new computational tools for jointly modeling algorithmic and rational theories of information search. Participants in our task can ask questions and learn about hundreds of everyday items but must retrieve queried items from memory. In order to maximize information gain, participants need to retrieve sequences of dissimilar items. Across several experiments, we find that participants are unable to do this. Instead, associative memory mechanisms lead to the successive retrieval of similar items, an established memory effect known as semantic congruence. These results indicate that there are critical limitations to people’s ability to ask good questions in naturalistic active learning tasks. They also suggest that machines can be used to improve human learning by correcting the associative biases inherent in memory and search.

An ecological perspective on collective behavior examines how collective behavior evolves to adjust to changing environments. Collective behavior is widespread in nature, not only producing the coordinated movement of bird flocks or fish schools, but also regulating activity in natural systems from cells, as in cancer metastasis or embryonic development, to the social groups of many vertebrates. Ant colonies function collectively, and the enormous diversity of more than 14K species of ants, in different habitats, provides opportunities to look for general ecological patterns. Modeling tools from engineering, including dynamical systems, control theory and distributed algorithms, show how local interactions produce the collective foraging behavior of harvester ants in the desert, and generate the trail networks of turtle ants in the tropical forest. These examples suggest how systems with similar dynamics in their surroundings have evolved to show similar dynamics in their collective behavior.

1. Gordon, D. M. (2014). The ecology of collective behavior. PLoS biology, 12(3), e1001805.

2. Pagliara, R., Gordon, D. M., & Leonard, N. E. (2018). Regulation of harvester ant foraging as a closed-loop excitable system. PLoS computational biology, 14(12), e1006200.

3. Davidson, J. D., Arauco-Aliaga, R. P., Crow, S., Gordon, D. M., & Goldman, M. S. (2016). Effect of interactions between harvester ants on forager decisions. Frontiers in ecology and evolution, 4, 115.

4. Chandrasekhar, A., Gordon, D. M., & Navlakha, S. (2018). A distributed algorithm to maintain and repair the trail networks of arboreal ants. Scientific reports, 8(1), 1-19.

5. Chandrasekhar, A., Marshall, J. A., Austin, C., Navlakha, S., & Gordon, D. M. (2021). Better tired than lost: Turtle ant trail networks favor coherence over short edges. PLoS computational biology, 17(10), e1009523.

I’ll discuss large-scale, community-level social coordination. Such co-ordination is the primary driver of individual opinion. The standard view in both academia and the wider culture is that people’s opinions are knowledge driven (the deficit model). The alternative view is that we channel our communities; our opinions reflect those of the people around us (the cultural consensus model). The evidence I discuss is mixed, different subsets supporting each of these views. First, I show that attitudes toward Covid mitigation practices in the US during the pandemic were not predicted by health or risk status, but rather by political ideology. Second, I show that knowledge does predict agreement with the scientific consensus, but that people’s sense of their own understanding is correlated with the opposite, opposition to the scientific consensus. Third, I show that people’s sense of understanding is influenced by what others understand, even when those others share no information of substance. We outsource our judgments to others. Finally, I will also show how to take advantage of this tendency to induce people to bring evidence to bear on policy. I discuss these findings in terms of the narratives that guide communities of knowledge. Those narratives are peculiarly human, even cutting-edge machine learning technology cannot generate them.

1. Sloman, S., & Fernbach, P. (2018). The knowledge illusion: Why we never think alone. Penguin.

2. Geana, M. V., Rabb, N., & Sloman, S. (2021). Walking the party line: The growing role of political ideology in shaping health behavior in the United States. SSM-population health, 16, 100950.

3. Fullerton, M. K., Rabb, N., Mamidipaka, S., Ungar, L., & Sloman, S. A. (2021). Evidence against risk as a motivating driver of COVID-19 preventive behaviors in the United States. Journal of Health Psychology.

4. Sloman, S., Kupor, D., & Yokum, D. (2021). Are voters influenced by the results of a consensus conference?. Behavioural Public Policy, 1-22.

5. Rabb, N., Han, J. J., & Sloman, S. A. (2021). How others drive our sense of understanding of policies. Behavioural Public Policy, 5(4), 454-479.

As social media become major channels for the diffusion of news and information, it becomes critical to understand how the complex interplay between cognitive, social, and algorithmic biases triggered by our reliance on online social networks makes us vulnerable to manipulation and disinformation. This talk overviews ongoing network analytics, modeling, and machine learning efforts to uncover malicious networks of coordinated online actors designed to deceive social media users.

1. Pacheco, D., Hui, P. M., Torres-Lugo, C., Truong, B. T., Flammini, A., & Menczer, F. (2021). Uncovering Coordinated Networks on Social Media: Methods and Case Studies. ICWSM, 21, 455-466.

2. Torres-Lugo, C., Yang, K. C., & Menczer, F. (2022, May). The Manufacture of Partisan Echo Chambers by Follow Train Abuse on Twitter. In Proceedings of the International AAAI Conference on Web and Social Media (Vol. 16, pp. 1017-1028).

3. Sasahara, K., Chen, W., Peng, H., Ciampaglia, G. L., Flammini, A., & Menczer, F. (2021). Social influence and unfollowing accelerate the emergence of echo chambers. Journal of Computational Social Science, 4(1), 381-402.

Over the last decade, it has become painfully clear that online social networks are not just a modernised version of the traditional public square, but rather are best conceived of as a hybrid human – machine intelligence. The fact that we have handed over to such systems a considerable portion of public debate is problematic in many ways. Chief among these is the fact that the opaque means by which discourse is shaped are geared not to social goods, but toward maximising advertising revenue. This raises the question of what we might choose to do if we were to couple these tools with goals such as maximising epistemic success. The talk describes recent research on algorithmic rewiring in the service of enhancing individual and collective accuracy in contexts where the truth might not yet be known and past reliability offers no guidance. The talk describes sample algorithms that dynamically rewire connections in communicating groups and provides initial evidence on their utility from experiment and simulation.

Many of the actions we perform involve coordination with others, be it making music together, having a conversation, or putting up a Christmas tree together. In this talk, I will discuss how individuals in small scale social interactions may benefit from interacting with each other. An obvious benefit of engaging in joint actions is that people can achieve outcomes they could not achieve on their own. This, in turn, may contribute towards expanding the space of action possibilities and action effects that individuals are considering when engaging in problem solving, action planning, or creative activities. Second, individuals can benefit from adopting different perspectives of their interaction partners. This can be seen in tasks where visuo-spatial perspective taking is required, as well as in communicative tasks where the need to describe complex information to multiple partners improves individuals' ability to explain things to others. The effects of engaging in tightly coordinated interactions with others may thus extend from immediate achievements of joint goals to more long-term changes in cognitive skills and mental space.

Research and development of human-machine partnerships has increasingly shifted to evaluation based on the performance of the team, rather than solely the accuracy of the machine or algorithm in isolation. This broader focus raises questions about when and whether one part of the team should, or could, compensate for biases and shortcomings in the other. In particular, I will focus here on the algorithm potentially compensating for biases — cognitive, social, moral, legal, or other — of the human decision-maker. I first show that this type of compensation will occur automatically for a wide range of learning algorithms and system designs; compensatory algorithms are not an unusual edge case, but rather a likely outcome in many situations. I then turn to the ethics of compensatory algorithms. I argue that it can be morally obligatory, permissible, or forbidden for an algorithm to compensate for human biases, depending on the features of the broader context. I finish by taking a close look at one of the most interesting situations: when is it morally permissible to use a compensatory algorithm without disclosing this fact to the user?

As social animals, we routinely share past experiences when we interact with others. Such memory transmission permeates not only our face-to-face interactions, but increasingly, our exchanges on social media across a wide range of social connections. In cognitive-experimental research, although interest in the social transmission of memory can be traced back to Bartlett's seminal treatise in the early 1900s, almost a century of research on memory has almost exclusively focused on individuals working in isolation. Drawing upon this body of research on individual memory, we investigate in laboratory experiments how individual memory constraints shape the performance of the group, and in turn, how collaborative remembering by a group reshapes the memory of each member. In this context, we ask how collaborative recall changes memory representations not only in terms of what we remember but also how we organize these memories. We know from individual memory research that memory organization is important for driving retrieval and learning, motivating questions about the ways in which social influences can shape these fundamental processes. More generally, theory and data from my lab are aimed at elucidating cognitive mechanisms that underlie memory enhancement as well as forgetting in shared remembering, the influence of the structure of the social network on memory propagation, false memory transmission, and the cascading effects of these changes on the emergence of collective memory.

1. Rajaram, S. (2011). Collaboration both hurts and helps memory: A cognitive perspective. Current Directions in Psychological Science, 20(2), 76-81.

2. Choi, H. Y., Blumen, H. M., Congleton, A. R., & Rajaram, S. (2014). The role of group configuration in the social transmission of memory: Evidence from identical and reconfigured groups. Journal of Cognitive Psychology, 26(1), 65-80.

3. Rajaram, S. (in press). Collaborative Remembering and Collective Memory. Chapter to appear in M. J. Kahana &. A. D. Wagner (Eds.), Handbook on Human Memory. Oxford University Press.

Why are collectives more emotional than an array of separated individuals? This question has occupied thinkers throughout history, and with the rise of social media it is even more pressing than before. Historically, the majority of attention has been given to amplification driven by the way emotions are communicated and shared between people via emotional interactions. While emotional interactions are indeed a major driver for group emotionality, I argue and empirically show that further understanding of collectives’ emotions requires a broader view which integrates two additional processes: how emotions change and are being changed by the social infrastructure in they are communicated, and how these processes impact and being impacted by people’s perception of others’ emotions. I propose an infrastructure, perception, interaction framework that contributes to a more comprehensive understanding of group emotionality. I then provide empirical evidence for some of the connections in this framework and argue that framework should improve our ability to predict a variety of group emotionality occurrences and to find ways to regulate these emotions when necessary.

Running, swimming, or flying through the world, animals are constantly making decisions while on the move—decisions that allow them to choose where to eat, where to hide, and with whom to associate. Despite this most studies have considered only on the outcome of, and time taken to make, decisions. Motion is, however, crucial in terms of how space is represented by organisms during spatial decision-making. Employing a range of new technologies, including automated tracking, computational reconstruction of sensory information, and immersive ‘holographic’ virtual reality (VR) for animals, experiments with fruit flies, locusts and zebrafish (representing aerial, terrestrial and aquatic locomotion, respectively), I will demonstrate that this time-varying representation results in the emergence of new and fundamental geometric principles that considerably impact decision-making. Specifically, we find that the brain spontaneously reduces multi-choice decisions into a series of abrupt (‘critical’) binary decisions in space-time, a process that repeats until only one option—the one ultimately selected by the individual—remains. Due to the critical nature of these transitions (and the corresponding increase in ‘susceptibility’) even noisy brains are extremely sensitive to very small differences between remaining options (e.g., a very small difference in neuronal activity being in “favor” of one option) near these locations in space-time. This mechanism facilitates highly effective decision-making, and is shown to be robust both to the number of options available, and to context, such as whether options are static (e.g. refuges) or mobile (e.g. other animals). In addition, we find evidence that the same geometric principles of decision-making occur across scales of biological organisation, from neural dynamics to animal collectives, suggesting they are fundamental features of spatiotemporal computation.

I will present a research program that conceptualizes collective-level phenomena as psychologically-grounded, interactively-constructed, and dynamic in nature. Using experiments that involve conversational interactions in social networks, I will show how collective phenomena (e.g., collective memory and collective beliefs) can emerge from micro-level local dynamics (i.e., socio-cognitive mechanisms). Taking such an approach, I claim, offers a more sophisticated view of the processes that characterize human communities. But while this approach can help the scientific effort to model collective phenomena, it will further complexify this effort, for reasons that will be made clear.

It is well known that networks generated by common mechanisms such as preferential attachment and homophily can disadvantage the minority group by limiting their ability to establish links with the majority group. This has the effect of limiting minority nodes' access to information. We present the results of an empirical study on the equality of information access in network models with different growth mechanisms and spreading processes. For growth mechanisms, we focus on the majority/minority dichotomy, homophily, preferential attachment, and diversity. For spreading processes, we investigate simple vs. complex contagions, different transmission rates within and between groups, and various seeding conditions. We observe two phenomena. First, information access equality is a complex interplay between network structures and the spreading processes. Second, under certain circumstances, there is a tradeoff between equality and efficiency of information access (e.g., when the number of links between groups is small and information is transmitted asymmetrically). Our findings can be used to provide recommendations for the mechanistic design of social networks with information access equality.

1. Wang, X., Varol, O. & Eliassi-Rad, T. Information Access Equality on Generative Models of Complex Networks. Applied Network Science 7, 54 (2022). https://doi.org/10.1007/s41109-022-00494-8

My research group has been analyzing K-12 students learning behaviors when they collaborate to solve problems in computer-based STEM learning environments. In parallel, we have also been studying how adults train in simulation-based augmented training environments. In this presentation, I will discuss how students collaborate to co-construct knowledge by creating a shared understanding of the problem domain. We are interested in studying how a group’s prior knowledge distribution impacts the processes they employ to construct domain knowledge, the nature of their social interactions and metacognitive problem-solving strategies they use when they work on computational modeling tasks in open-ended learning environments (OELEs). Our study is based on a high school kinematics curriculum with inquiry and modeling tasks in 1D and 2D motion. By applying AI and Machine Learning methods to analyze the multimodal data that includes student discourse, video, and students’ actions logged in the OELE, we gain a better understanding of the processes that students working in pairs employ for co-constructing knowledge and building computational models in kinematics. We adopt a case study approach to show that the prior knowledge distribution within dyads has an impact on groups’ knowledge co-construction processes, i.e., how they integrate their science and computational knowledge to build their models, their social interactions, and their metacognitive problem-solving processes. We also show that over time, some of the differences are mitigated as students’ gain knowledge in both domains, and thus close the knowledge differential between partners. However, some difficulties persist over time. Therefore, this work suggests that designing student supports based on groups’ distribution of prior knowledge may lead to students adopting more effective knowledge co-construction processes.

Learning Engineering is the application of evidence-based principles from the learning sciences combined with computational and design methodologies in the computer sciences to create engaging and effective learning experiences, support the difficulties and challenges of learners as they learn, and come to better understand learners and learning. Computational and design methods, inherent to learning engineering, offer the means to not only enhance and personalize the learning process, but to aid in the orchestration of learning. To that end, it is important to consider ways to support learning at multiple levels based on the needs of the learner. Bloom's Taxonomy, for example, is a hierarchical model that categorizes learning objectives into levels of complexity, from basic recall of knowledge and comprehension to advanced analytical, evaluative, and creative processes. One common misconception of taxonomies such as Bloom's, is that educators should focus immediately or even exclusively on higher levels. For example, it has been suggested that jumping straight to collaborative exercises, skipping the foundational processes that allow the learner to construct the knowledge and skills, is a more productive use of instructional time. However, those knowledge and skills are necessary to participate in creative and collaborative tasks. Such an approach potentially leaves developing learners behind, without the requisite foundational knowledge and skills to engage in more complex tasks inherent to analysis, evaluation, and creation. My own work has focused primarily on the development of AI and tutoring technologies that support learners' development of strategies to understand, describe, and apply knowledge, i.e., lower, foundational levels within learning taxonomies. I will describe the roles of learning engineering in the development of these technologies. I will also discuss the multiple facets of technology development, potential directions for future work, methods to scaffold learning for students, and support educators in the orchestration of learning.

The amount of information available to individuals today is enormous and rapidly increasing; however, human cognition is limited in the amount of information it can process at once. One promising solution to this problem is through combining human and machine intelligence at scale to accelerate sensemaking and innovation. I’ll discuss our exploration of architectures breaking up thinking into components that can be externalized and augmented by other humans or machines outside an individual’s brain, kickstarting a virtuous cycle where everyone can build on others’ work instead of starting from scratch. Specifically, I’ll focus on our latest thread of research aimed at harnessing the trillion hours of online information seeking and sensemaking that people engage in every year, and the challenges and opportunities for research and deployment.

People sometimes get stuck in creative writing. Can AI nudge people out of it? Consider the task of writing as many attributes for a given concept, e.g. "penguins", in a short time. I may write "live on ice, eat fish, ..." but then cannot think more. AI can nudge me by showing a new word close in embedding space to an old word I have written, e.g. "shark" which is close to "fish". I may then be inspired to think of predators and write a new attribute "eaten by seals"; note I do not have to use the nudge word. Or AI can nudge me by showing a completely random word, e.g. "Dota" which makes me think of computer games and then write "are main characters in Club Penguin". More generally, there can be more than two types of nudge. Which nudge type is more useful for a particular writer is unknown and may change over time. I will explain how this is a multi-armed bandit problem, and how the bandit can optimally tradeoff exploration (estimating the usefulness of each nudge type) and exploitation (giving the writer the most useful nudge type) adaptively.

Understanding human social dynamics requires understanding the forces that shape how we perceive and assort with others. A driving force in our interconnected, cosmopolitan world is social identity—those affiliations through which we structure our social lives. I'll discuss how considerations of identity influence models of collective dynamics, and present both modeling and empirical research on how individuals signal their identities in diverse populations. In particular, I'll focus on how covert signals of identity enable assortment and coordination in diverse communities.