Collective Intelligence

Collective Foraging
Harnessing Human and Machine Collective Intelligence in Open-Ended Medical Diagnostics
A Cognitive–Computational Approach to Social and Collective Decision-Making
How Large Language Models Reshape Collective Intelligence

Collective Foraging

Foraging—locating and extracting resources from the environment—is a ubiquitous behavior across species. Whether bees searching for pollen, squirrels searching for nuts, or humans searching for information online, foragers have to develop effective decision-making strategies to reduce uncertainty. Moreover, foraging usually takes place in complex social settings. Some evolutionary scientists argue that the selective pressures of collective foraging were crucial to the evolution of complex human cognition. To better understand how humans forage socially, ARC researchers explored the decision-making mechanisms at work in an immersive social foraging task and a natural collective field setting.

Wu et al. (2025) studied foraging in a version of the game Minecraft, where participants could relocate freely in a virtual arena and collect rewards in groups of four. Participants foraged across different resource distributions (clustered vs. dispersed resources), revealing how the distributions impacted two processes: personal exploration and social exploitation. Combining visual field data with high-resolution spatial tracking showed that social learning is immensely flexible: With clustered resources, individuals used area-restricted search and turned their attention to others, especially when they themselves were unsuccessful. When resources were highly dispersed and social information was not valuable, individuals reduced their reliance on social learning.

Inspired by these findings, Schakowski et al. (2026) investigated social foraging behavior in ice fishing competitions in Finland (Figure 1). Ice fishing is an important tradition in the Nordic countries, and competitions are regularly held to preserve this cultural heritage. During competitions, roughly 50 ice fishers were equipped with GPS tracking devices and headcams across 10 days, enabling ARC to study spatial search and record foraging success. Similar to the laboratory-based results, foragers were found to use social information to locate resources, especially when they were unsuccessful themselves.

Taken together, the two studies show that social learning in foraging is selective rather than indiscriminate: People turn to others when their own search fails, not by default. This selectivity may help explain how groups avoid maladaptive herding. More broadly, the studies highlight the potential of collective foraging as a paradigm for understanding adaptive social decision-making in ecologically relevant settings.

Key References

Harnessing Human and Machine Collective Intelligence in Open-Ended Medical Diagnostics

Diagnostic errors are among the most pressing issues in medical practice, causing an estimated 795,000 deaths and permanent disabilities in the United States alone each year. Reducing diagnostic errors is essential to improving patient outcomes worldwide. With this in mind, ARC researchers explored using the wisdom of crowds to boost decision accuracy in open-ended medical diagnostics.

The wisdom of crowds has been mostly applied to relatively simple, low-dimensional decision tasks, such as binary classifications. But scalable applications in open-ended domains, which have a much larger problem space (e.g., general medical diagnostics with over 60,000 possible diagnoses), are largely lacking because it is challenging to integrate unstandardized inputs from different crowd members (e.g., free-text diagnoses from clinicians). Kurvers et al. (2023) therefore developed an automated approach for harnessing the wisdom of crowds in general medical diagnostics. It leverages semantic knowledge graphs, natural language processing, and the SNOMED CT medical ontology to automatically identify the intended diagnosis from unstructured text. The authors applied their method to 1,333 medical cases, each independently rated by 10 diagnosticians on a medical crowdsourcing platform. The approach substantially increased diagnostic accuracy: Individual diagnosticians achieved 46% accuracy, while pooling the decisions of 10 diagnosticians increased accuracy to 76%. Importantly, the general approach of this study could be applied to many domains where structured expert knowledge is available or can be gathered.

In another study, Zöller et al. (2025) investigated the potential of combining human diagnoses with the diagnoses of five state-of-the art large language models (LLMs). The authors applied the automated approach developed by Kurvers et al. (2023) to identify diagnoses in free-text responses from both humans and machines. The hybrid collectives of physicians and LLMs outperformed individual physicians, physician collectives, single LLMs, and LLM ensembles (Figure 2). These results held across a range of medical specialties and professional experience and are a consequence of the different kinds of errors that arise from the complementary contributions of humans and LLMs.

Both studies highlight the potential for collective human and machine intelligence to improve accuracy in complex, open-ended domains like medical diagnostics.

Key References

A Cognitive–Computational Approach to Social and Collective Decision-Making

Many decisions are embedded in social contexts, such as sharing news on social media or deciding whether to jaywalk at a busy intersection. Individuals are embedded within social groups that both influence and are influenced by individuals. These processes of reciprocal social influence shape the dynamics of social systems.

Although social systems are inherently dynamic, few studies take a dynamic approach to investigating how cognitive processes shape and are shaped by collective dynamics over time. Instead, social cognition studies typically look at static, often simulated, sources of social information, leaving them unable to accommodate the complexity of multiple individuals interacting dynamically and repeatedly in real time.

Tump et al. (2024) argued for two points as a promising path forward for describing, understanding, and predicting collective outcomes. First, researchers should use dynamic cognitive–behavioral experiments in which participants repeatedly interact in real time and respond to the unfolding social environment created by the decisions of others. Second, researchers should model the cognitive processes of dynamically interacting individuals. The paper highlights this approach by applying two highly influential cognitive models, evidence accumulation (Figure 3) and reinforcement learning, to interactive group experiments, showing how they can simultaneously reveal how individual cognition drives social systems, how social systems drive individual cognition, and the dynamic feedback processes between both layers.

Key Reference

How Large Language Models Reshape Collective Intelligence

Since the public release of ChatGPT in late 2022, large language models (LLMs) have transformed the information environment. Instead of using a search engine to look up information or asking a colleague to help brainstorm ideas, many people now turn to an LLM. The implications of this transformation do not stop with individuals; LLMs are poised to reshape the collective intelligence of teams, markets, and societies at large.

Burton et al. (2024) formed an interdisciplinary team of 28 researchers from across industry and academia to identify how LLMs will affect collective intelligence, and how this process can be shaped in the public's interest. LLMs can expand collective intelligence by lowering barriers to participation, via translating and providing writing support, accelerating idea generation, mediating deliberation with real-time guidance, and synthesizing diverse viewpoints at scales far beyond what human groups alone can manage. Yet some of the same properties that make LLMs powerful pose new risks: LLMs may reduce incentives for individuals to contribute to open knowledge commons (e.g., Wikipedia or Stack Exchange), narrow the diversity of thought that underpins effective collective problem-solving, create illusions of consensus by obscuring minority perspectives, and contribute to the production and dissemination of false or misleading information.

These benefits and risks call for careful stewardship. Without regulations and guidelines for LLM development and usage, LLMs could inadvertently erode the foundations of collective intelligence upon which societies—and LLMs themselves—have been built. To prevent this, the authors call for open models, expanded public research infrastructure, and third-party oversight of real-world LLM deployment. Such measures could ensure that these new systems strengthen, rather than replace, human collective intelligence.

Key Reference