As large language models revolutionize how we gather and share information, this new era of AI-driven collective intelligence brings powerful benefits but also significant challenges, from shrinking diversity to misinformation risks.
Perspective: How large language models can reshape collective intelligence. Image Credit: Stokkete / Shutterstock
In a research paper published in the journal Nature Human Behaviour, researchers examined how large language models (LLMs) are transforming not only how information is aggregated but also how it is accessed and transmitted online, presenting unique opportunities and challenges for collective intelligence (CI). They highlighted that CI refers to the ability of individuals to collectively act in ways that surpass the intelligence of any single individual, including experts. CI is crucial for the success of groups and organizations, enabling outcomes beyond individual capabilities. The authors called for interdisciplinary collaboration to explore the benefits, risks, and policy considerations related to these models in addressing complex problems. The paper stressed the need for further research into the delicate balance between leveraging LLMs and maintaining human-driven contributions to CI.
Do LLMs Reshape CI?
Recent advancements in LLMs have significant implications for CI, influencing civic deliberation, elections, and everyday interactions. Researchers synthesized interdisciplinary perspectives to identify how LLMs could reshape CI, highlighting their capabilities to enhance diversity, individual competence, and idea generation. Crucially, CI is supported by three key elements: the diversity of participants, the competence of individuals, and effective aggregation mechanisms. LLMs could either promote or undermine these elements depending on how they are integrated into collective processes. CI, the collective ability to act intelligently, relies on both surface-level diversity (e.g., demographics) and deeper functional diversity (e.g., different ways of solving problems), individual competence, and effective aggregation mechanisms.
LLMs can facilitate larger, more engaged groups by breaking down language barriers and providing writing assistance. They can also accelerate idea generation by serving as an instant source of ideas and enhancing individual creativity. Ultimately, LLMs represent both a tool for improving CI and a product of collective contributions, necessitating careful examination of their impacts. However, there are concerns that relying too heavily on LLMs for idea generation may homogenize thought processes, reducing the functional diversity that is essential for high-quality collective decision-making.
LLMs: Enhancing Deliberation, Risks
The text discusses the role of LLMs in facilitating CI through deliberative processes and collaboration while highlighting potential risks. It acknowledges that while many individuals may be reluctant or unable to participate meaningfully in deliberative democracy due to cognitive limitations and disillusionment with the processes, LLMs could serve as cognitive aids, prompting individuals to articulate their views more clearly and assisting in deliberative support. LLMs could also act as facilitators, managing the flow of conversations, summarizing opinions, and prompting participants to consider new angles or refine their arguments. They could engage with participants, manage discussions, and help clarify objectives by summarizing diverse opinions.
However, the paper also outlines risks associated with LLMs. For instance, their use could discourage contributions to traditional knowledge commons, as people may prefer to rely on LLMs for information instead of engaging with original content. This could potentially lead to a decline in the quality and diversity of contributions. This trend could significantly weaken the ecosystem of open platforms such as Wikipedia, where CI thrives on user-generated content.
Additionally, reliance on LLMs can create consensus illusions by favoring commonly represented views, which may obscure minority opinions and limit functional diversity. It can lead to premature convergence of ideas and reduce the quality of collective decision-making, as individuals may converge too quickly on similar solutions without exploring diverse perspectives.
Moreover, LLMs pose risks in spreading misinformation. They can generate plausible but incorrect information and may be misused in disinformation campaigns, making challenging or verifying their outputs difficult. The opaqueness of LLM-generated content can also exacerbate these risks, as it may not be immediately clear which sources or datasets the models are drawing from. The authors recommend addressing these challenges through open access to LLMs, increased computational resources for diverse research, and greater oversight of LLM use to mitigate risks while enhancing their benefits for CI. In particular, they advocate for "truly open LLMs," where models are transparent about their training data and for public funding of independent research into LLMs' societal effects.
The discussion emphasized the potential of LLMs to enhance deliberative processes and collaborative efforts but also underscored significant risks related to misinformation, reduced contributions to knowledge commons, and the propagation of illusory consensus. To balance these challenges and opportunities, recommendations focused on promoting transparency, accessibility, and oversight in LLM development and deployment. The paper suggests that third-party audits and model tracking systems could help manage some of these risks.
Generative AI's Future Impact
In addition to the recent advancements in LLMs, significant progress has been made in other forms of generative artificial intelligence (AI), such as image, video, and audio generation systems. Examples include mid-journey for images, sora for videos, and seamlessM4T for audio.
While these systems are not yet sufficiently developed to have an immediate impact on CI, they hold the potential to influence it in the future. Generative AI applications in these domains could support creativity by enabling faster prototyping and design, similar to how LLMs facilitate idea generation. Currently, most applications of generative AI in image, video, and audio generation are primarily for entertainment purposes, and empirical evidence regarding their broader implications is lacking.
However, tools like DALL·E, mid-journey, and stable diffusion can accelerate the creation of visual designs and prototypes, encouraging divergent thinking and aggregating non-linguistic styles from various designers. The integration of these tools into CI processes will require careful consideration of their effects on creative diversity and the potential for misuse, especially in generating misleading or harmful content.
Despite the positive aspects, the implications of image-generating AI are only partially beneficial. The widespread adoption of these applications may discourage individuals from contributing to image-based knowledge commons, similar to how LLMs can reduce contributions to text-based platforms. Additionally, malicious actors could exploit multimedia-generating AI to create false or misleading content, potentially leading to more severe consequences given the powerful impact of visual and auditory information.
While generative AI's potential to reshape CI remains uncertain, advocating for open models, increased computational resources for independent researchers, and enhanced oversight of generative AI usage is essential. The authors argue that the risks and benefits associated with these technologies must be thoroughly examined before they are widely integrated into CI frameworks. Considering their societal impacts and acceptability, a nuanced evaluation of these technologies' strengths and weaknesses is crucial for understanding their implications.
Conclusion
To sum up, the rapid adoption of LLM applications like chat generative pre-trained transformer (ChatGPT) transformed the online information environment. However, the impact on individual information search, reasoning, and decision-making remained uncertain. It was essential to anticipate potential collective effects proactively.
As CI tools, LLMs enhanced CI and posed risks to society's problem-solving capabilities. The paper not only explores these risks but also provides a roadmap for future research, particularly in maintaining diversity and improving oversight in the age of LLM-driven CI. This paper synthesized perspectives from diverse researchers in academia and industry, motivating further investigations into the relationship between LLMs and CI.