Economists measure artificial intelligence's contribution to economic growth through three distinct frameworks: forward-looking productivity projections built on task-based models, current-state exposure analyses of what work AI can actually touch, and investment tracking that follows the money flowing into AI infrastructure. No single method yields a definitive number, because AI behaves less like a conventional product and more like a general-purpose technology that resists the categories of national accounting.

Introduction: The Challenge of Measuring AI's Economic Impact

The core difficulty is that standard economic statistics were designed for goods and services with clear prices and quantities. As Intereconomics researchers explain, within national accounting frameworks many AI systems function as bundles of different categories of investment, which makes them hard to measure directly. As a result, much of the preliminary evidence on AI's diffusion rests on proxies—measures of AI exposure or AI skill demand—rather than on observed use of the technology itself.

The stakes of this measurement problem are considerable. A review of the empirical evidence on AI and productivity by the International Center for Law & Economics notes that the absence of a pronounced productivity surge in official statistics may reflect measurement constraints rather than a lack of real economic gains. That possibility—that the statistics are wrong, not the technology—frames why economists have developed parallel approaches rather than waiting for GDP data to speak clearly.

Deconstructing AI's TFP Growth Projection

The clearest illustration of how a forward-looking projection is actually built comes from the Penn Wharton Budget Model, whose September 2025 analysis by Alexander Arnorm follows a framework developed by the MIT economist Daron Acemoglu. The model projects generative AI's long-run effect on Total Factor Productivity—the portion of output growth not explained by more labor or capital—by multiplying two terms: the GDP share of tasks impacted by AI and the percentage cost savings from AI.

Each term decomposes further. The GDP share of tasks impacted equals the GDP share of tasks exposed to AI multiplied by the share of exposed tasks worth automating; the cost-savings term equals the percentage labor cost savings multiplied by labor's share of income from exposed tasks. A third component, the generative AI adoption rate, scales the long-run estimate to a timeline, defined relative to the universe of tasks that can be profitably automated.

The framework's logic is worth stating plainly: AI only moves productivity where it touches economic activity, and only to the extent that it saves cost. A technology exposed to half the economy but yielding trivial savings produces little; large savings on a sliver of tasks produce little as well. The five components below show how the projection is assembled.

Penn Wharton Budget Model: Components of the TFP Projection
ComponentWhat It CapturesHow It Enters the Calculation
GDP share of tasks exposed to AIHow much economic activity AI could touchFirst factor of the "GDP share of tasks impacted" term
Share of exposed tasks worth automatingWhether automation is profitable, not merely possibleSecond factor of the same term
Percentage labor cost savings from AICost reduction on labor in exposed tasksFirst factor of the cost-savings termLabor's share of income from exposed tasksHow much of those tasks' income goes to laborSecond factor of the cost-savings term
Generative AI adoption rateHow quickly profitable automation actually spreadsMultiplies the long-run change to produce the projected TFP change

Readers can use this decomposition to interrogate any AI productivity forecast: the projected TFP change is simply the product of these five quantities, and disagreement about the projection is disagreement about the inputs.