Automating a task is not the same as automating a job, and the difference decides the effects

The unit that gets automated is smaller than the unit people count

A job title covers a collection of activities, many of which have nothing to do with the one the title names. A radiologist reads images and also consults, supervises, decides what to image next, and takes responsibility. A translator translates and also negotiates meaning with a client, judges register, and stands behind the result.

Technology arrives at the level of activities. When one activity becomes cheap, the job does not vanish; its composition changes, and the remaining activities take up more of the time. Whether that makes the job better, worse, more numerous or fewer in number depends on things outside the technology entirely.

This is the single most common error in forecasts about employment here. Counting jobs where some activity can be automated and calling the total at risk assumes the bundle disappears with the activity, which historically it usually has not.

Substitution and complementarity produce opposite effects

If a tool does the same thing a worker does, it competes with them and pushes down what that work is worth. If it makes the worker more effective at something they still need to do, it raises the value of their time. The same tool can do both, to different workers, in the same industry.

Which effect dominates depends partly on demand. Where cheaper output creates more demand for the whole activity, employment can rise even as each unit takes less labour. Where demand is fixed, the same efficiency reduces the labour required. Economists have studied this pattern across earlier waves of automation and it is not predictable in advance from the technology alone.

There is a further complication: the tasks that become cheap are not distributed evenly across a workforce. Effects concentrate on whoever spent most of their time on the affected activity, which is frequently people earlier in their careers.

This wave points at different work from the last one

Previous automation displaced physical and routine clerical activity, and the standard framework held that non-routine cognitive work was insulated by its unpredictability. Systems that produce text, images, code and analysis do not respect that boundary, which is why the earlier framework has needed revision.

What has not changed is that anything requiring physical dexterity in unstructured surroundings remains difficult and expensive to automate. The result is an unfamiliar ordering in which some highly credentialled cognitive work is more exposed than skilled manual trades, which reverses the assumption of the last several decades.

This ordering is a description of current capability, not a law. It has already shifted once and could shift again in either direction, and treating it as fixed would repeat the mistake made by the framework it replaced.

The measurable effects so far are modest and hard to read

Aggregate employment statistics move slowly and reflect a great many things at once, which makes attributing any change to a specific technology extremely difficult. Studies of individual tasks in controlled settings have found meaningful productivity effects, often larger for less experienced workers, and generalising from those to an economy is a long step.

Adoption is also slower than announcement. Organisations have to redesign processes, establish who is accountable, and satisfy themselves about quality before anything changes at scale, and that has historically taken far longer than the technology itself. The gap between capability and deployed effect is measured in years.

Anybody quoting a confident figure for jobs affected is quoting a model with assumptions in it. The assumptions are usually the interesting part and they are usually not stated.

The distributional question is separate from the aggregate one

Even where the total amount of work is unchanged, who does it and what they are paid can change substantially. Historical episodes of automation produced long transitions in which aggregate employment recovered and specific communities did not, and the average concealed the part that mattered most to the people in it.

That makes the policy questions here about transition rather than about totals: what happens to somebody mid-career whose main activity became cheap, who pays for retraining, and whether the gains show up as wages or as returns to whoever owns the systems. These are political questions with a technical trigger.

The one defensible general claim is that the pace matters more than the destination. A change absorbed over decades is a different social event from the same change absorbed over years, and reasonable people disagree sharply about which this will be.