Labour economics of AI

The unit matters because the arithmetic changes with it. Say a tool halves the time a task takes. If that task is the whole job, the job needs half the hours. If it is three of eight tiles weighted by hours, the saving is whatever those three tiles were worth, and the other five tiles are untouched. Most real weeks are the second case. The habit is to ask "which tiles?" before "which jobs?", and to refuse to answer the second until someone has done the first.

Substitution is the tool doing the task instead of the person. The hours leave the person, and where they go is a decision someone makes: reassigned to other tiles, or removed from the payroll, or filled with tiles that did not exist before. Complementarity is the tool making the person more productive at the task. The hours stay and produce more, and a task that was a bottleneck can stop being one, which raises the value of the tiles around it. The same tool on the same job usually does both, tile by tile, and a sentence about "AI and this job" that does not say which tile it means is not yet a sentence.

Exposure is the word for "a tool could plausibly touch this tile". It is a statement about the tile, not about what happens next. A highly exposed task can be complemented rather than substituted, and a substituted task can free hours into work that was previously starved of them. Exposure is where the analysis starts. It is not where it ends.

Two people hold the same job title and different bundles. One spends most of the week on the three exposed tiles; the other spends most of it on the five plain ones and picks up the exposed three for an hour a day. The tool arrives. For the first person, the exposed tiles were the job, and the outcome depends entirely on what management does with the freed hours. For the second, the tool takes an hour a day off tasks that were never the point, and the week gets better. Same title, same tool, opposite experiences, and a job-level statistic would average them into a number that describes neither.

This is also why the experiments in Part 3 report different effects for novices and for experienced workers on the very same task. The tool's contribution to a task is measured against what the person could already do. Where the person was already good, there is little to add; where they were new, the tool carries expertise the person has not yet built. Whether that narrows the gap between them for good or only for now is one of the contested questions, and the page leaves it contested.

Displacement is the arrow everyone sees. A machine, or a model, takes over a task that a person used to do, and the share of tasks performed by labour falls. It is the mechanism behind every "jobs at risk" headline, and it is real; it is also only half of the paper. Reinstatement is the arrow that the headline leaves out: the creation of new tasks in which people have a comparative advantage over the machine. New tasks are hard to count because they do not exist yet when the forecast is made, which is why forecasts systematically see the first arrow more clearly than the second.

The paper's contribution is to put both on the same line and make the net a question about their relative sizes rather than a slogan. That is why the page cites the 2019 Journal of Economic Perspectives paper, "Automation and New Tasks: How Technology Displaces and Reinstates Labor", 33(2) 3-30, and not the 2018 American Economic Review paper by the same authors, which is a different model with a different argument. Getting the citation right is part of getting the claim right: the claim is two arrows and an open net.

For generative models specifically, the honest position is that displacement is being measured tile by tile (Part 3) and reinstatement is, by its nature, not yet countable. So the net is not known, and the page says so rather than picking a side.

Robert Solow, reviewing a book in the New York Times Book Review on 12 July 1987, is widely reported to have written that you can see the computer age everywhere but in the productivity statistics. The line is reported here, not read from the microfilm, and it names a pattern that has recurred with every general-purpose technology since: adoption is visible long before its effect on measured output is. Three reasons, none of them exotic. Firms adopt the tool before they reorganise the work around it, and the gain is in the reorganisation. The complementary investments, in training, in data, in new processes, are spent before they pay. And the statistics that would show the gain are built around the old tasks, so a new task that the tool created is measured badly or not at all.

The consequence for tonight is a warning in both directions. Early task-level gains, like those in Part 3, do not automatically show up in output statistics, so their absence in the aggregate numbers is not evidence that the task gains are illusory. And the task gains do not automatically become aggregate gains either, because the reorganisation between the tile and the economy has not happened yet. Both readings are open. Someone who tells you which one is right has more information than the statistics contain.

Whether a tool that complements a task raises or lowers the wage of the person doing it is contested, and the contest is not a failure of research; it is that the answer depends on things the task-level studies do not measure. If the tool raises the output of novices most, as the support-desk study found, one reading is that the wage gap between novice and expert narrows because the expert's edge is now partly in the tool. Another reading is that the expert's remaining edge, on the tasks the tool cannot do, becomes scarcer and more valuable, and the gap widens. Both are coherent, both have supporters, and the same experiment is compatible with each because the experiment measured a tile and the wage is paid for the bundle.

The professional norm is therefore to say what is measured and what is inferred, in separate sentences. "Novices resolved 34 percent more issues per hour with the assistant" is measured. "Therefore novice wages will rise" is inferred, and the inference needs an argument about the labour market that the study did not make. Keeping the two apart is most of what it means to read this literature well.

Brynjolfsson, Li and Raymond, "Generative AI at Work", NBER Working Paper 31161 (2023; later published in the Quarterly Journal of Economics, 2025). Customer-support agents, a generative assistant, one task: issues resolved per hour. Up 14 percent on average, 34 percent for novice and low-skilled workers, minimal for experienced and highly skilled workers. Read at source; primary. The unit is a rate per hour, and the headline is the gradient by experience as much as the average.

Noy and Zhang 2023, Science 381(6654). Professional writing tasks, done with and without a model. Reported as about 40 percent faster with about 18 percent higher quality. The figures are reported here from the abstract as reproduced by a secondary index; the article itself was not opened in this build, so the tier is secondary and the page says "reported". The unit is time on a task and a graded quality score, both at the task.

Dell'Acqua et al. 2023, "Navigating the Jagged Technological Frontier", Harvard Business School Working Paper 24-013. Consultants given a model, on a set of tasks the authors chose to sit inside the model's capability and one task chosen to sit outside it. Inside: 12.2 percent more tasks completed, 25.1 percent faster, more than 40 percent higher quality. Outside: 19 percentage points less likely to produce a correct solution. Read from the working paper; primary. The two-sided result is the reason the study matters more than its good numbers: the same tool, the same people, and the sign flips with the task.

What the three share is the design. Each randomised who got the tool, so the difference is the tool and not the kind of person who chooses to use one. What the three also share is the unit. Every figure above is a task in an experiment. Not one is a job in an economy, and not one of the papers claims otherwise.

"19 percentage points less likely to produce a correct solution" is a gap between two rates: the share of the tool group who got it right, minus the share of the control group who did. A percentage point is one hundredth of the whole, subtracted. "19 percent less likely" would be a proportion of the control group's rate, multiplied. They are only the same number when the base rate is 100 percent, which it never is.

Make it concrete with an invented base rate, used only for the arithmetic and not the paper's own figure: if the control group got 80 of 100 right, 19 points takes the tool group to 61 of 100. Nineteen percent of 80 is about 15, which would take the tool group to about 65. The misquote understates the finding by four people in a hundred on that base rate, and by a different amount on any other base rate, which is exactly why the unit has to travel with the number. The paper says points. Say points.

A task study holds the job fixed and changes the tool. It cannot see what the firm does with the freed hours, because in the experiment the firm did nothing yet; it cannot see whether the task still exists as a human task a year later, because the experiment ran for weeks; and it cannot see reinstatement at all, because a task that does not exist cannot be randomised. These are not gaps a better experiment would close. They are the boundary of what a task experiment is, and the papers state their boundary plainly.

The open question therefore has a precise shape: given that these tiles got faster or better, what happens to the bundles that contain them, and to the people who hold the bundles? Answering it needs data on jobs over time, which is a different kind of evidence with different problems, and the results so far do not agree with each other. The professional stance is to carry the task figures with their units and their tiers, to carry the employment question as a question, and to notice when someone joins the two with a "therefore" that no study supplied.