A Collectivist, Economic Perspective on AI (2507.06268v1)

Abstract: Information technology is in the midst of a revolution in which omnipresent data collection and machine learning are impacting the human world as never before. The word "intelligence" is being used as a North Star for the development of this technology, with human cognition viewed as a baseline. This view neglects the fact that humans are social animals, and that much of our intelligence is social and cultural in origin. A related issue is that the current view treats the social consequences of technology as an afterthought. The path forward is not merely more data and compute, and not merely more attention paid to cognitive or symbolic representations, but a thorough blending of economic and social concepts with computational and inferential concepts, in the service of system-level designs in which social welfare is a first-class citizen, and with the aspiration that a new human-centric engineering field will emerge.

• The paper introduces a tripartite framework that integrates computational, economic, and inferential thinking to enhance AI system design with a focus on social welfare.
• It critiques current approaches to AI uncertainty and advocates for collective strategies that address data provenance and market design challenges.
• It proposes novel market models, employing mechanisms like Stackelberg games and incentive-compatible contracts, to improve recommendation systems and protect user data.

Economic Perspectives on AI

This paper addresses the limitations of viewing AI solely through the lens of cognitive intelligence, advocating for a more comprehensive approach that integrates economic and social concepts with computational and inferential methods. It posits that AI development should prioritize social welfare and system-level designs, moving beyond the individualistic perspectives prevalent in classical AI. The author suggests that embracing uncertainty and understanding social dynamics are crucial for creating AI systems that enhance the human condition.

Uncertainty and Collectives in AI

The paper critiques the perception that LLMs truly quantify uncertainty and reason inductively, arguing that their capabilities are limited by the human-generated data they are trained on. Real-world uncertainty management requires a deeper understanding of epistemic considerations, information asymmetries, and the provenance of data. The paper emphasizes the role of collectives in mitigating uncertainty, using the example of markets for produce reducing individual foraging uncertainty. The paper claims that technology needs to be designed with collectivist principles in mind to aid adaptive decision-making on a large scale.

Multi-Way and Multi-Layered Markets

The author expands on the idea that the internet should be viewed as a place where economic interactions occur, and not merely as a large collection of text and images. The paper then critiques the existing internet markets for their inability to reward creators, value data as an economic good, create trust, and disincentivize socially harmful behavior. It suggests incorporating microeconomics and mechanism design principles.

Recommendation Systems

The paper discusses recommendation systems as examples of collectivist ML systems, but notes their limitations as microeconomic entities due to the absence of monetary transactions. It introduces a three-way market for recorded music (Figure 1), involving musicians, listeners, and brands, where incentives are aligned through direct payments to artists when their music is used by brands. This design contrasts with traditional online music platforms, where revenue is generated through subscriptions or advertising, providing weak incentives for compensating musicians.

Figure 1: A three-way market for recorded music in which aligned incentives are present.

Data Markets

The paper analyzes a three-layered market structure (Figure 2) consisting of users, a platform providing services, and third-party data buyers. The platform provides a service to the user in exchange for a payment. The platform sells the data that it obtains to third-party data buyers. It highlights the issue of user privacy loss when platforms sell user data to third parties. It then proposes that platforms should provide formal guarantees of privacy by adding noise to the data. The overall system is then modeled as a Stackelberg game to find its equilibria. This requires specifying utility functions for the various players.

Figure 2: A three-layer market in which a platform provides services, and also sells data to third-party buyers.

Computational, Economic, and Inferential Thinking

The paper introduces "inferential thinking" and "economic thinking" as thinking styles that are complementary to computational thinking when designing systems that operate in social environments. It argues that inferential thinking involves characterizing populations and understanding the underlying mechanisms by which data arise, while economic thinking focuses on the design of incentives and the achievement of social welfare or revenue.

Computation and Inference in Database Design

The paper contrasts computation and inference in a stylized database problem. It posits that inferential thinking refers to the design and analysis of algorithms that can extract value, and that it requires consideration of the underlying population, the set of possible queries, and the sampling operator.

Figure 3: (a) A database in which a randomized algorithm Q provides privacy guarantees while ensuring that the privatized response $\tilde{y}$ to a query $x$ is not too far from the true response'' $y$. (b) An inferential perspective, in which the database is assumed to arise from an underlying population, under a sampling operator $S$. The goal is to ensure that $y$ is not too far from thetrue response'' $y^*$. (c) The privacy-preserving-inference problem, where we make predictions while providing a privacy guarantee for the individuals who were in the original database. Here, we want $\tilde{y}$ to be close to $y^*$.

Inference and Incentives

The paper discusses the intersection of inferential and economic thinking, particularly in scenarios where data suppliers have strategic interests in the outcome of data analysis. It describes the design and analysis of incentives as a major part of economic thinking. The paper then introduces the concept of statistical contract theory for settings involving sequential play, Stackelberg equilibria, and information asymmetry. It discusses the design of contracts using e-values to control false positives and false negatives, even in the presence of strategic data.

Bias and Local Knowledge

The paper emphasizes the importance of heterogeneity in collectivist AI systems, arising from differences in participant types, goals, and the locality of data and knowledge. It argues that local knowledge and biases are critical for solving local problems, and that these biases should be mitigated using methodologies like prediction-powered inference (PPI).

AI Education

The paper suggests that academia has not caught up with the tripartite blend of computational, economic, and inferential thinking. It describes how the field of machine learning blends computation and inference, econometrics blends economics and inference, and algorithmic game theory blends computation and economics (Figure 4). The author argues that a tripartite blend is needed, and that academic curricula should reflect this blend.

Figure 4: Three core thinking styles that have come together in pairwise blends as academic disciplines.

Conclusion

The paper concludes by advocating for a more nuanced consideration of issues like fairness, privacy, and transparency through the integration of computational, economic, and inferential perspectives. It calls for the development of modular, transparent design concepts for AI systems, drawing inspiration from the engineering disciplines. Finally, it recognizes the essential contributions of cognitive science, social psychology, the humanities, and behavioral economics in shaping human-centric technology.