Artificial intelligence is reshaping the global economy in ways that are both profound and uncertain. While some economists predict only modest gains—around a 0.9% rise in GDP over the next decade—others forecast transformative growth, with AI potentially adding between $17 trillion and $26 trillion annually to global output by 2045 and automating up to half of today’s jobs. Yet even before these changes fully unfold, perceptions of AI’s future are already influencing decisions across society: guiding career paths, shaping government policy, and fueling massive investments in semiconductors and data center infrastructure. To understand AI’s true economic impact, researchers are turning to rigorous methods like natural experiments and randomized controlled trials. These studies compare groups with and without access to AI tools under controlled conditions, measuring outcomes such as productivity, job satisfaction, and learning. However, two major limitations hinder their usefulness. First, AI evolves so rapidly that findings from one year may be outdated by the next. For example, early studies showed AI helped call-center workers handle queries 15% faster and software developers complete 26% more tasks. But today’s AI systems can handle three times as many customer-support tasks independently as they could just a few years ago—making past results less relevant. Second, controlled studies often miss broader systemic effects. They assume stable organizational structures, but in reality, firms may reorganize work, replace human roles with AI, or shift responsibilities. Evidence suggests that employment among younger workers—especially in AI-prone fields like customer service and software development—has declined since 2022. While it’s unclear how much this trend is driven by AI, it underscores the risk that controlled experiments may underestimate negative impacts on certain groups. To address these gaps, researchers need more than data—they need imagination grounded in economic theory. One promising approach is “social science fiction,” a term coined by economist Jean Tirole. This involves speculative but rigorous scenarios based on core principles of human behavior and market dynamics. For instance, early models predicted that self-driving cars could worsen urban congestion because passengers would tolerate longer commutes when they can work or relax in transit. Such insights highlight the need for policies like congestion pricing to correct unintended consequences. Another key insight is that automation may not always reduce labor demand. As AI handles routine tasks, the value of uniquely human skills—such as creative oversight, judgment, and emotional intelligence—could rise. This could boost wages and opportunities in certain roles, potentially softening the blow of job displacement. But it could also deepen inequality between those who can adapt and those who cannot. Beyond theory, real-time data is essential. Benchmarks measuring AI performance on standardized tasks—like solving math problems or diagnosing diseases—are useful but often fail to reflect real-world complexity. A medical AI might excel on textbook cases but struggle with ambiguous patient communication. Better benchmarks are needed to capture practical performance. Meanwhile, tracking actual AI usage patterns offers valuable clues. Data shows that AI chatbots are widely used in software development, signaling early adoption in that sector. Other indicators—such as employment trends, job openings, firm profitability, and expansion rates—can reveal how AI is affecting the economy in real time. Yet descriptive data alone cannot prove causation. Researchers must still isolate whether AI drives improvements or simply correlates with already high-performing organizations. To overcome these challenges, researchers can adapt methods used in scaling pilot programs. By simulating the conditions of large-scale implementation—such as budget constraints or bureaucratic hurdles—they can anticipate how AI might perform when adopted broadly. Cost trends are especially critical: as AI computation becomes cheaper, its reach expands. In low-income regions like Sierra Leone, where internet access is costly, AI tools are becoming dramatically more affordable relative to web browsing—potentially democratizing access to knowledge. Ultimately, predicting AI’s economic impact requires combining rigorous data, imaginative modeling, and an understanding of human behavior. While technology evolves rapidly, human responses tend to follow predictable patterns—how we trust machines, respond to incentives, and adapt to change. By blending these insights, policymakers and researchers can prepare for a future where AI reshapes not just productivity, but the very fabric of work and society.