The Learning Machine: Artificial Intelligence, Labour, and Expertise.
Festival of Ideas · Judge Business School · University of Cambridge · March 24, 2026
Robert Solow's 1987 observation — that "you can see the computer age everywhere except in the productivity statistics" — reappears in the AI era with striking fidelity.
Two competing interpretations organise the debate. Optimists project cumulative GDP gains of 6.1% over ten years. Pessimists put the figure at 0.9%. The gap is not about the technology — it is about assumptions: task coverage, diffusion speed, and whether AI augments or substitutes human labour.
General-purpose technologies extract their productivity gains only after a costly reorganisation of firm structure and work practices. We may be in the trough of that curve now.
Brynjolfsson, Rock & Syverson — The Productivity J-CurveGoldman Sachs: +6.1% cumulative GDP over ten years as AI diffuses across the economy.
Acemoglu: +0.9% — gains confined to a narrower set of tasks than the hype suggests.
Real gains, structurally lagged. Intangible investment is the binding constraint — invisible to standard accounting.
The paper's original analytical contribution. AI does not dissolve the expertise-scarcity-wage chain — it relocates the scarcity node.
The central question is not whether AI can perform a task. It is what kind of knowledge the task requires — and therefore whether scarcity, once eroded, can reconstitute elsewhere. AlphaGo's Move 37 marks the empirical boundary of Tier 3.
Codified, written down, reproducible. AI learns it directly. Scarcity erodes. Wage premium falls.
Polanyi's surplus — partially circumvented by scale. Premium compresses but survives at the judgment frontier.
Reinforcement learning surpasses human knowledge. Premium shifts to oversight and governance of what has gone beyond knowing.
Neil Lawrence's information-theoretic argument: the ratio between human internal processing and external communication is 2.5 trillion to one.
Human experience is structurally unextractable — not merely hard to articulate, but impossible to fully externalise. This is a permanent architectural feature of human cognition, not a temporary technological limitation.
We know more than we can tell.
Michael Polanyi, The Tacit Dimension, 1966Lawrence asks what remains irreducibly human. Weeks asks what happens to the wage of what remains irreducibly human. These are different questions with different answers — and conflating them produces the wrong policy conclusion.
The expertise-scarcity-wage chain has organised thinking about human capital for 250 years. AI does not dissolve it — it relocates the scarcity node.
"A man educated at the expence of much labour and time... may be compared to one of those expensive machines."
Adam Smith, The Wealth of Nations, 1776When the scarcity node moves — from the human expert to the platform infrastructure through which AI is accessed — the distributional consequences are profound. The gains flow not to the worker whose expertise was replicated, but to the owners of the infrastructure that did the replicating.
Knowledge lives in the hand and eye. Scarcity is absolute.
Division of labour makes knowledge explicit. Specialisation creates both value and vulnerability.
AI codifies cognitive work at scale. The scarcity node relocates.
The shift from AI as tool to AI as actor. Agentic systems do not merely respond — they plan, coordinate, and execute across multiple tasks autonomously.
The Model Context Protocol acts as a coordination layer — automating the Coasian glue between tasks. This is invisible to standard occupational exposure indices and systematically underestimated by existing measurement frameworks. The unit of displacement is no longer the task. It is the role.
Chatbots and reasoning models. Responds and plans, but does not act independently.
Tool use and multi-agent coordination. Plans, executes, delegates across systems. Where we are now.
Autonomous actors with minimal human oversight. Sets objectives, manages resources across extended horizons.
The future is not determined by the technology. It is determined by decisions being made now, in boardrooms, legislatures, and research labs.
There is one word missing from much of this discussion: choice.
Daron AcemogluSystems that replace human judgment, or systems that amplify it? The design philosophy chosen today locks in the labour market outcomes of tomorrow.
Substituting existing tasks versus creating entirely new categories of human capability. The latter generates new demand; the former compresses wages.
General-purpose models concentrate returns in a handful of firms. Domain-specific systems distribute capability more broadly across sectors.
Zero-sum races accelerate deployment without governance. Cooperative frameworks allow the gains to be distributed rather than captured by incumbents.
Productivity gains from agentic AI depend critically on infrastructure access. Firms with MCP connectivity and vertical integration capture the premium. Those without do not.
The expertise-scarcity-wage chain persists — but the scarcity node has moved from human skill to platform architecture. This structural divide is invisible in aggregate statistics, decisive in competitive outcomes, and largely absent from current policy frameworks.
AI does not dissolve the expertise-scarcity-wage chain. It relocates the scarcity node — from human expertise toward platform and infrastructure access.
The Learning Machine — Central ArgumentVertical integration + MCP access. Agentic AI multiplies firm-specific knowledge. Premium captured.
No infrastructure integration. Agentic gains unavailable. Productivity gap widens.
Existing frameworks address applications. Infrastructure access asymmetry is unregulated and growing.
The Learning Machine: Artificial Intelligence, Labour, and Expertise.
Festival of Ideas · Judge Business School · University of Cambridge · March 24, 2026
