← Wiki

The Direction of Technology

ai-and-jobstechnology-and-laborinstitutionspolitical-economy

The Argument

Daron Acemoglu and Simon Johnson argue in Power and Progress (2023) that there is nothing automatic about new technologies bringing widespread prosperity. Whether they do is “an economic, social, and political choice.” Technology is shaped by what powerful people want and believe — and throughout 1000 years of history, the default direction has been to serve elite interests, not broad prosperity.

This challenges the “productivity bandwagon” — the assumption among economists and technologists that productivity gains automatically translate to shared benefit. Acemoglu and Johnson show this assumption is historically false: medieval agricultural technology enriched lords while peasants saw zero net wealth gains over 200 years. Industrial Revolution productivity surged while worker wages stagnated for a century. Digital technology boosted corporate profits while wage growth flatlined.

Machine Intelligence vs Machine Usefulness

Acemoglu distinguishes two directions for AI:

Machine Intelligence (the current path)

Focuses on replicating human cognitive abilities through large datasets and pattern recognition. Enables worker displacement. Creates “so-so” productivity gains. “Many algorithms are being designed to try to replace humans as much as possible. We think that’s entirely wrong.” This path supercharges surveillance, labor substitution, and emotional manipulation.

Designs AI to amplify human capabilities rather than replicate them. “Don’t think of your labor as a cost to be cut. Think of your labor as a human resource to be used better, and AI would be an amazing tool for it.” Example: AI tools that expand what home health care workers can do, making their services more valuable without reducing the sector’s workforce.

So-So Technologies

A critical concept: not all automation is progress. “So-so technologies” displace workers without generating meaningful productivity or service quality gains. Self-checkout kiosks transfer work to customers while eliminating cashier jobs — no net efficiency gain. Automated call centers frustrate customers while reducing service quality. These technologies represent the worst outcome: destroyed livelihoods without created value.

The contrast: genuinely productive technologies create “marginal productivity gains” that compel firms to hire more workers and increase compensation. Software tools that aid car mechanics in precision diagnosis increase worker marginal productivity — a different dynamic from industrial robots installed to replace people. See Enabling vs Replacing Technologies.

New Tasks: The Key Variable

Acemoglu and Restrepo’s research (2020) found that for every robot added per 1,000 US workers, wages declined by 0.42%. But when technology creates genuinely new tasks — roles that didn’t exist before — the result is productivity gains AND wage gains AND employment gains simultaneously. The historical pattern: shared prosperity came specifically when technology created new tasks for humans, not when it automated existing ones.

The Expanding Work Frontier tracks the same distinction: AI agents that enable workers to do things they never could before (review every contract in a portfolio, build custom demo environments, monitor supply chains in real time) follow the prosperity-generating path. AI that simply replaces call center workers follows the impoverishing path.

The 100-Year Lesson

The Industrial Revolution’s productivity gains took 100 years to reach workers — not because the technology improved slowly, but because institutional power shifted slowly. Workers secured gains through organized struggle: unions, democratic governance, regulatory frameworks. Post-WWII prosperity was the exception, not the rule — enabled by specific institutional conditions (strong unions, progressive taxation, public investment) that had to be fought for.

The implication for AI: the question is not “how capable will AI become?” but “who controls how AI is deployed, and toward what ends?” This is Polanyi’s The Double Movement applied to the AI era — market-driven deployment provokes social protection responses, and whether those responses block progress entirely (see The Technology Trap (Concept)) or channel it constructively depends on institutional design.

Counter-Arguments

Noah Smith challenges several claims: real wages rose during AI’s commercialization (2012-2024), with production workers outpacing managers. Employment reached record highs; inequality flatlined. Smith finds zero historical examples of governments successfully directing firms toward specific technology directions before those technologies were invented. The “menu of technologies” the authors propose may be impractical — societies can demand redistribution of technology’s gains without prescribing which technologies to develop.

Acemoglu’s own earlier research (2022) found positive or neutral employment effects from IT capital investment, complicating the book’s automation narrative.