The means of prediction and the production function of AI

Behavior of Generative AI While Interpreting Risk-Sensitive Forecasts

The adoption of Generative AI in business has been increasing recently. This increase in demand is evident by the huge investments coming up in data centers and IT infrastructure across the world to support LLM models. But the core question that remains unaddressed in the decision science domain is whether LLM models are good enough to interpret and take business level decisions on inventory, planning and operations based on forecasts. A core tenet is that the forecasts itself are not always necessarily accurate. So, it is important to investigate whether generative AI understands the uncertainty associated with the forecast, on which its decision is based. When a group of humans take a decision, if all of them have the same understanding of the uncertainty, even though expressed in different forms or wordings, the final decision always remains the same. This is how normal decision theory works. This paper investigates whether generative AI models give consistent decisions, understand the risks of uncertainty and remain committed to the decision regardless of changes in phrasing.

From Frontier to Public Good: Governing the AI Stack as Economic Architecture for Inclusive Productivity, Trust, and Resilience

Artificial intelligence is a general-purpose technology, and like the technologies before it, its value surfaces only through complementary investment in skills, data, organizational redesign, and institutions. History is suggestive here: the size and distribution of the gains turned at least as much on the institutions built around a technology as on the technology itself. AI governance, by contrast, is still treated largely as content moderation of generative outputs. This paper takes a different view. Governance is itself a critical complement, and the distribution of AI's productivity gains is, we hypothesize, partially endogenous to how governance is designed rather than fixed by capability alone. We treat AI as a full stack, from foundational machine learning, computer vision, and reinforcement learning through the foundation-model and agent layers down to compute, hardware, and a material base of minerals and energy; within that stack, compute may be the most governable layer under present institutional conditions. The paper sets out a fullstack operational definition of AI governance, a complementaryinvestment framing, a measurable enabling condition that weighs innovation lift against compliance drag, a three-layer architecture coupled through conformity assessment and mutual recognition, and an operational measurement framework. It is conceptual and formal rather than empirical: an analytically structured framework built for later empirical testing.

Artificial Intelligence and the Brave New World in Finance

What do AI companies deserve? Shapley attribution and optimal tax theory

AI Expectations and Outcomes

Evaluating the accuracy of predictions about AI use and outcomes is important given their potential influence on policy and decision making. In this paper, we study whether businesses’ expectations about their use of AI have translated into the desired outcomes. First, we use the U.S. Census Bureau’s Business and Trends Outlook Survey for 2023–2026 to compare predictions of AI use with actual AI use to see how well businesses have been anticipating their true rate of AI adoption. We find a learning curve: AI adoption initially happened slower than expected, was followed by a short period of growth that was faster than expected, and more recently has been close to expected rates. Next, we see if the motivations for adopting AI translated to associated outcomes by combining data about the motivations of early AI adopters from the Census Bureau’s 2019 Annual Business Survey with detailed data on output, inputs, and total factor productivity from the BEA-BLS Integrated Industry-Level Production Account. We find some evidence that stated motivations for using AI are linked to related changes in production processes; importantly, use cases for AI are associated with increased intensity of R&D use. This suggests that even if the link between motivations and outcomes is murky at this point, structural change may be in the planning process but not yet observed in the outcome data.

AI winters as legitimacy crises. From the history of technological promises to governance models of generative AI

In the debate on artificial intelligence, the term “AI winter” is often used as shorthand for a technical failure or a temporary decline in market interest. This article proposes a different interpretation. We argue that the classical AI winters were primarily legitimacy crises, in which the system of justifications linking technical promise, funding, commercialization, and social acceptability collapsed. Drawing on the history of AI, legitimacy studies, and the sociology of expectations, we propose a hierarchical interaction model of four legitimacy gaps. In this model, the capability gap functions as the technical substrate of AI promises; the institutional assessment and commercialization gaps mediate whether those promises are credited, funded, and productized; and the governance gap operates as a meta-condition of legal, moral, and political authorization. On this basis, we reinterpret the first and second AI winters. We argue that the episodes later grouped under the first AI winter can be read primarily as a crisis of capability and assessment, whereas the second was a crisis of product, brand, and commercial ecosystem. We then argue that the contemporary boom of generative models does not herald a simple repeat of past winters. A more likely scenario is a regulatory-economic cooling driven by compliance costs, disputes over training data, infrastructure concentration, information manipulation, algorithmic verification, and increasing documentation requirements. In this scenario, generative AI becomes contested not only as an automation technology, but also as epistemic infrastructure involved in producing, verifying, ranking, and stabilizing public truth. In the final section, we formulate four implications for governance: promise restraint, evidence of deployment, documentation obligations, and greater infrastructural pluralization. This perspective shifts the debate from the question of whether AI works to the question of under what conditions its development remains politically and ethically legitimate.