Machine Usefulness
Machine usefulness is the design philosophy Acemoglu and Johnson propose as the inclusive alternative to the prevailing AI research agenda. Where current frontier AI investment concentrates on systems that perform tasks without human involvement — autonomous agents, end-to-end replacement — machine usefulness targets systems that extend, augment, and support human capability without displacing the human from the loop. The distinction is not about technological capability but about the target of optimization. A useful machine makes a mediocre radiologist into a good one; an autonomous machine makes the radiologist redundant. Both are technically possible. The second is currently overinvested relative to the first.
In the AI Story
The concept draws empirical support from studies showing that augmentation-oriented AI deployments produce larger productivity gains for less-skilled workers than for highly-skilled ones — compressing wage distributions rather than widening them. Erik Brynjolfsson and colleagues' 2023 study of customer service AI found fourteen percent productivity gains concentrated among the lowest-performing workers, with negligible effects on top performers. This distributional profile is the opposite of what current AI deployment typically produces, and the difference lies in design choices rather than technical capacity.
Applied at industry scale, the machine usefulness agenda would redirect research subsidies, restructure venture capital incentive schemes, and reform procurement policies to favor augmenting technologies. The automation tax is one component. Public research funding for human-AI collaboration interfaces, rather than for ever-larger foundation models, is another. Government procurement requirements that specify augmentation goals rather than replacement goals is a third.
The concept challenges a foundational assumption of the contemporary AI industry: that the goal of research is artificial general intelligence capable of replacing human labor across domains. Acemoglu and Johnson argue this goal was never economically justified — the social returns to augmenting a global workforce of four billion are higher than the returns to replacing them — and has persisted as a cultural artifact of the research community's specific aesthetic preferences, not as a response to actual demand.
The usefulness framework also reframes the developer-in-Lagos case. Useful AI systems would be designed around the specific constraints she faces: unreliable power, intermittent bandwidth, local language, limited capital. Current AI is designed for Silicon Valley infrastructure and speaks English. The democratization claim is true of the model but false of the deployment environment, and machine usefulness is the design discipline that would close the gap.
Origin
The concept was developed across Acemoglu and Johnson's 2023 book and extended in their 2024 MIT Shaping the Future of Work publications. It builds on Daniel Susskind's parallel work on task redesign and on empirical studies from Brynjolfsson and others.
Key Ideas
Augmentation and replacement are design choices. The same underlying AI capability can be shaped toward either goal, and the choice is made by researchers, funders, and deployers — not by the technology itself.
Augmentation compresses wage distributions. Empirical evidence consistently shows augmenting AI delivers larger gains to less-skilled workers, producing inclusive rather than extractive distributional profiles.
Current AI investment is skewed. Venture capital, research subsidies, and corporate priorities currently favor replacement over augmentation by orders of magnitude.
Redirection requires institutional action. Shifting the AI research agenda toward usefulness requires policy intervention — tax reform, public research direction, antitrust enforcement — not merely cultural exhortation.
Debates & Critiques
Critics including some at OpenAI and Anthropic argue the distinction between augmentation and replacement is unstable because augmentation technologies evolve into replacement technologies as capabilities grow. Acemoglu's response is that the trajectory is not deterministic; intermediate deployment decisions, shaped by institutional incentives, determine whether a given system lands on one side or the other.
Appears in the Orange Pill Cycle
Further reading
- Acemoglu and Johnson, Power and Progress, Chapter 9 (2023)
- Brynjolfsson, Li, and Raymond, 'Generative AI at Work,' NBER Working Paper (2023)
- Acemoglu, 'The Simple Macroeconomics of AI,' Economic Policy (2024)
- MIT Shaping the Future of Work Initiative publications