AI changes the work. Who gets the gains?
Series date follows the editorial schedule. First published ; updated .
Automating part of a job does not tell us who benefits. The worker might finish earlier. Their employer might cut staffing. Lower prices might bring in customers who could not previously afford the service. All three can happen in the same industry.
That is the question behind this series: can we deliberately build AI systems that make people more capable, and how would we know when we have succeeded? I favour amplification as a design objective. I do not think it is a law of technological progress.
This revises the earlier version of this overview. Its accounting analogy promised too much: that routine work would disappear, people would move into judgment, and more jobs would follow. The evidence supports competing mechanisms, not that assured sequence. The unsourced numerical illustrations and mixed employment forecasts have been removed.
A job is more than its easiest demonstration
Consider an accountant using software to reconcile transactions. The software can reduce one kind of effort without settling who handles an unusual expense, explains the accounts to a client, or accepts responsibility for an error. That is an illustration of task decomposition, not a claim that every accountant's career followed the same path.
In his 2015 essay on workplace automation, David Autor explains how machines can substitute for some tasks while complementing others. Productivity, demand, and labour supply interact. The survival of an occupation's name therefore tells us little about whether particular workers kept their hours, wages, or bargaining power.
AI also reaches tasks we would call judgment: classification, diagnosis, writing an argument. Calling something judgment is not a permanent reservation for human labour. We have to examine the actual task and its consequences.
Displacement is part of the story
Daron Acemoglu and Pascual Restrepo's 2019 framework distinguishes the displacement of labour from the creation of new tasks in which people have a comparative advantage. Automation can raise productivity while reducing labour demand. New tasks can pull the other way.
It matters which force is stronger. Saying that technology created new work before does not show that this transition will be quick, painless, or beneficial to the same people who lose work. Their framework and historical decomposition are a way to reason about those forces, not an estimate of the number of AI-era jobs.
An aggregate gain can coexist with a very bad outcome for a particular worker.
Cheaper work does not automatically mean more paid work
Imagine a small bookkeeping business, not a reported case. An assistant drafts client explanations faster. The owner could serve more clients, spend longer checking difficult accounts, shorten the working day, or reduce paid hours. Software capability alone cannot choose between those outcomes.
Extra demand depends on customers wanting and paying for the extra service. Even if volume rises, the labour required per unit may fall enough that total hours decline. The earlier version of this essay called on the Jevons paradox as if lower cost guaranteed more jobs. That skipped both demand and the organisation of work.
There is a similar gap in the bubble argument. A useful product does not establish that every investment in it will earn a return. Equally, an investment loss does not show that the product has no useful applications. I am making neither an investment forecast nor a claim that new infrastructure inevitably pays for itself.
Better than what?
A combined workflow can improve on unaided human performance and still be worse than a feasible automated alternative. It can also produce excellent results while demanding so much review that the total cost rises. The first numbered essay separates those comparisons.
For the hypothetical bookkeeping service, count accepted explanations, corrections and total staff time, including checking. Ask clients whether the result helps them understand their accounts. Do not count draft paragraphs as the outcome. Where autonomous handling would violate an obligation, compare only lawful alternatives rather than running an unsafe experiment to complete a chart.
And ask who receives the benefit. A productivity test is not a substitute for a decision about pay, workload, access, or accountability. Those are choices people make around the tool.
Keep a way to learn the work
If software supplies the first answer every time, a beginner may get less practice constructing one. That is a design risk to test, not proof that all assistance causes cognitive decline. We can measure assisted output and later independent performance separately.
In the hypothetical business, a trainee could make an initial interpretation before reading the draft, then discuss the difference with a qualified reviewer. Whether this teaches useful judgment needs checking on new cases. Simply leaving a human in the approval chain does not establish that learning happened.
The remaining essays examine routing, novice performance, learning, code, creativity and skills before putting the pieces into a trial plan. The aim is practical: decide which work AI should do, give people a contribution worth making, and keep enough evidence to change the arrangement when it fails them.
Sources and scope
Autor (2015), Why Are There Still So Many Jobs?, explains substitution, complementarity and demand. Acemoglu and Restrepo (2019), Automation and New Tasks, distinguishes displacement from reinstatement. Neither establishes that generative AI will create more jobs than it removes. The bookkeeping example and all five illustrations are explanatory designs, not measured outcomes.