Generative AI at Work (2023)
The 2023 NBER working paper Generative AI at Work, by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, provided some of the first rigorous empirical evidence about how generative AI affects worker productivity. Studying the staggered rollout of a generative AI conversational assistant to 5,179 customer support agents at a large software company, the authors found that access to the tool increased productivity (measured in issues resolved per hour) by 14 percent on average. The gains were radically uneven across the skill distribution: novice and low-skilled workers improved by 34 percent, while experienced and highly skilled workers saw minimal impact. The finding suggested the AI was effectively capturing and disseminating the tacit knowledge of the best workers — helping newer employees move down the experience curve faster. If the pattern generalized, AI could compress skill inequality rather than widen it, though subsequent evidence on junior hiring declines complicated the optimistic interpretation.
The Substrate Dependency Problem — Contrarian ^ Opus
There is a parallel reading that begins from the material conditions required to sustain this compression effect. The study captures a moment when AI assistance costs could be absorbed into enterprise software budgets, but the exponential compute requirements for maintaining and improving these systems create a fundamental instability. As the models grow more capable, their operational costs scale faster than the productivity gains they enable—especially for the low-margin, high-volume work where skill compression matters most. The 14% productivity gain becomes meaningless if maintaining the AI assistant costs 20% more year-over-year.
Read through the lens of organizational power dynamics, the compression mechanism itself becomes a form of deskilling rather than upskilling. The novice workers aren't developing expertise—they're becoming dependent on an AI crutch that captures and crystallizes the knowledge of experts who can now be more easily replaced. The company gains the most: it can hire cheaper workers, achieve similar outcomes, and maintain the threat of further automation. The compression isn't democratizing skill; it's commoditizing it. When the next model arrives that doesn't just surface expert knowledge but synthesizes it autonomously, both the novices who never developed real expertise and the experts whose knowledge has been extracted become equally redundant. The study documents not skill development but skill dissolution—a process where human capability becomes progressively externalized into systems that workers access but never truly possess.
In the AI Story
The study's design enabled causal inference in a way most AI productivity research has not. Because the tool was rolled out in a staggered fashion across teams, the authors could compare productivity changes in teams that received the tool at different times, controlling for general trends and seasonal effects. This quasi-experimental design — closer to a randomized trial than the correlational analyses that dominate the literature — gave the findings unusual empirical weight.
The mechanism the authors identified was pedagogical: the AI was surfacing the responses and approaches of the company's most effective agents, making this tacit knowledge available in real-time to agents who had not yet developed it through years of experience. The AI functioned less like a replacement for human judgment and more like an automated mentor, compressing the training timeline for skill development. Novices gained most because they had the most to learn. Experts gained least because they had already developed the capabilities the AI was now propagating.
The finding had major implications for the Great Decoupling debate. If AI generally compressed skill distributions rather than widening them, the technology could narrow rather than widen inequality — a dramatic reversal of the pattern established by previous digital technologies. But the generalization had limits. The customer service context provided tight feedback loops, clear quality metrics, and well-defined tasks — conditions that may not hold across all work domains. Moreover, separate data began showing entry-level hiring declining sharply in AI-exposed occupations by 2025, suggesting that AI's effect on the existing workforce (compression) diverged from its effect on the future workforce (pipeline collapse).
The paper became one of the most cited economics papers on AI and work by 2025, shaping both the academic debate and policy discussions. It functioned as empirical ballast for the augmentation-over-automation argument — evidence that AI could function as augmentation in practice, not merely in principle, producing broadly distributed gains rather than concentrating them.
Origin
Danielle Li is an associate professor at MIT Sloan; Lindsey Raymond completed her PhD at MIT and is now at Stanford. The study was supported by the anonymized Fortune 500 software company that provided the data, which allowed the researchers unusual access to granular performance data combined with AI tool usage data.
The working paper was released through NBER in April 2023, revised through 2024, and published in the Quarterly Journal of Economics in 2025. It entered the public discourse at the moment when AI workforce effects were becoming the central policy question, giving it disproportionate influence on both academic and public debates.
Key Ideas
14 percent average productivity gain. Access to the AI assistant increased resolved-issues per hour across the full workforce.
34 percent gain for novices. Workers with less experience saw dramatically larger productivity improvements than experienced workers.
Minimal gain for experts. Experienced workers showed almost no productivity improvement from the tool.
Compression mechanism: tacit knowledge diffusion. The AI surfaced the approaches of top performers, making them available to novices in real time.
Optimistic generalization limited. The customer service context may not generalize, and the effect on existing workers diverges from the effect on pipeline hiring.
Appears in the Orange Pill Cycle
Temporal Horizons of Compression — Arbitrator ^ Opus
The disagreement hinges fundamentally on timeframe. For the immediate present (2023-2025), Edo's reading dominates—the empirical evidence is unambiguous that AI tools are compressing skill gaps and helping novice workers achieve expert-level performance faster. The 34% productivity gain for novices isn't theoretical; it's measured and real. Score this 80% to Edo's framing. The augmentation is happening now, with clear benefits to workers on the lower rungs.
But extend the timeline to 2030 and the contrarian view gains weight. The substrate dependency problem becomes acute as model costs compound while margins compress. More critically, the organizational dynamics shift: once companies realize they can achieve expert outcomes with novice workers plus AI, the incentive to maintain either expert positions or advancement pipelines evaporates. Here the split is 70% contrarian—the commoditization dynamic likely dominates the compression benefit.
The synthesis requires holding both temporal realities simultaneously. AI creates a genuine but temporary window of skill compression that benefits existing workers, while simultaneously undermining the structural conditions for skill development in the future. The right frame isn't compression versus commoditization but compression-toward-commoditization—a process where the immediate benefits to workers serve as the mechanism for their eventual displacement. The Brynjolfsson study captures the first movement accurately; it cannot see the second movement because it hasn't happened yet. Both readings are correct at their respective horizons. The challenge is that workers experience the short-term while organizations plan for the long-term, creating an alignment problem that no amount of empirical evidence about current productivity can resolve.
Further reading
- Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. Generative AI at Work. NBER Working Paper, 2023.
- Noy, Shakked and Whitney Zhang. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 2023.
- Peng, Sida, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. 2023.