CAT 2020 Slot 3 VARC: Automation and the Future of Work

For most of the twentieth century, automation was something that happened to other people's jobs. Assembly-line workers watched robotic arms take over welding and painting; typists watched word processors swallow the pool. The consolation offered to everyone else was that automation ate routine, codifiable tasks and left untouched the messier work of judgment, persuasion and creativity — the work, in short, that required a human. That consolation is now harder to sustain. Machine-learning systems draft legal memoranda, triage radiology scans, compose passable advertising copy and negotiate the first round of customer-service disputes. The line between "routine" and "skilled" work, once thought to track the line between automatable and safe, has turned out to be far blurrier than economists assumed.

The more useful distinction, several labour economists now argue, is not between routine and non-routine tasks but between tasks that can be specified and tasks that cannot. A task is specifiable if its inputs, outputs and success criteria can be stated in advance, however complicated the underlying reasoning. Grading a standardized essay against a rubric is specifiable, even though it once seemed to require a human's literary sense. Comforting a grieving family, or deciding which of two competing organizational priorities should yield in a genuine crisis, is not specifiable in this way — not because it is intellectually harder, but because success itself resists prior definition. On this view, automation's advance is less a march up a ladder of "skill" than a slow colonization of everything that can be written down as a rule, however intricate.

This reframing changes what worries us, and what it should. It is not the neurosurgeon's most difficult judgment calls that are safest from automation, and not the file clerk's simplest ones that are most at risk; it is instead the neurosurgeon's routine post-operative documentation and the file clerk's improvised handling of an irate, non-standard customer that trade places in the risk ranking. Occupations, as a result, are not simply automated or not; they are unbundled. A single job title increasingly survives as a shrinking core of unspecifiable tasks wrapped around a shrinking-still-further set of specifiable ones that software now performs, with the human worker left to supervise, exception-handle and take responsibility for a process she no longer fully executes.

Two consequences follow, and they cut against each other. The optimistic one is that unbundling need not mean disappearance: many jobs retain an irreducible core precisely because so much of what makes them valuable — trust, context, the willingness to be accountable when something goes wrong — cannot be specified in advance and delegated to a system. The pessimistic one is that this residual core is often thin, poorly paid relative to the work it replaces, and unevenly distributed: the professional whose unspecifiable judgment commands a premium and the delivery driver whose unspecifiable improvisation is compensated at minimum wage are both, technically, doing "the part a machine cannot do," yet one arrangement looks like elevation and the other like exhaustion.

What should be resisted is the temptation to resolve this tension by predicting, industry by industry, which jobs will "survive." That prediction game has a poor track record, not because forecasters are careless but because the boundary of the specifiable keeps moving as both the technology and our tolerance for its errors change. The more durable question, and the one this reframing puts within reach, is institutional rather than predictive: who captures the value released when a task is automated, and who bears the cost of the exception-handling that remains? Answering that question well may matter more for the future of work than any inventory of which jobs are, this year, deemed safe.