The basic logic of an automation prediction
Predictions about jobs and automation usually follow a recognisable recipe. First, a job is described as a bundle of tasks, perhaps drawn from an occupational database. Second, someone judges how well current or near-future technology can perform each task. Third, the results are combined into a score for each occupation, and the scores are added up across the workforce to produce a headline such as a share of jobs said to be at risk.
Each step is reasonable on its own terms. The problem is that the result is a statement about technical exposure, and it is often read as a statement about what will happen to employment. Those are different claims.
Where the judgements come from
The task ratings in these exercises often come from a small group of experts, from a statistical model trained on their ratings, or from an analysis of how closely a task resembles things software already does. All of these involve interpretation. Two careful teams can look at the same occupation and reach quite different ratings, because “can a machine do this task” depends on how the task is described and how good is good enough.
Task descriptions in databases may also be dated or generic. A job title can cover a very different mix of duties in different workplaces, so an occupation-level score may not describe any particular worker’s day.
Why outcomes tend to differ
Even when the technical judgement is sound, several reasons commonly explain why real change looks different from the forecast.
Technical ability is not adoption. A system that works in a demonstration may be costly, unreliable in messy conditions, or hard to integrate with existing processes. Employers weigh these against the price of the labour they already have. Adoption is often slower, patchier and more selective than the technical possibility suggests.
Jobs change instead of vanishing. When some tasks in a role are automated, the remaining tasks may grow in importance, or new ones may appear. The job may continue under the same title with a different mix of duties. A forecast that counts only tasks removed will miss that.
New work is hard to see in advance. It is easier to imagine what a technology replaces than what it creates. Roles that grew out of earlier technologies were seldom anticipated in forecasts made before those technologies spread. This does not guarantee that new work will always offset losses, and nothing here is a promise either way, but it is a recurring blind spot in methods that mainly count what disappears.
Other forces move at the same time. Employment is shaped by economic cycles, interest rates, population ageing, trade, regulation and consumer demand. A forecast that isolates one technology can look wrong simply because other forces dominated. Untangling them afterwards is difficult, because several confounding variables operate together, and it is rarely possible to say how much of a change was due to the technology alone.
Human and social factors. Customers may prefer a person for certain services. Rules may require human sign-off. Workers and employers may negotiate how tools are used. These considerations seldom appear in a task-based score.
The timing problem
A forecast needs a date, and dates are where confidence tends to outrun evidence. Technologies often improve in uneven bursts, and the spread into workplaces can take far longer than the first impressive results. Forecasts that offer a precise year for a large shift are, by their nature, guesses with a stamp on them.
Timing matters for interpretation too. If a prediction says a job faces high exposure over a long period, and nothing dramatic happens in the first few years, that does not necessarily show it was wrong. But it also cannot be confirmed by that quiet period. Long-range forecasts are difficult to check, which is partly why they stay in circulation.
Why forecasts are rarely re-examined
Forecasts attract attention when published and little attention afterwards. When a predicted change does not arrive on schedule, there is seldom a follow-up piece, so the original claim lingers in memory while the correction does not. Meanwhile the occasional striking example of a job genuinely transformed gets retold, which is a form of survivorship bias: the cases that fit the narrative are remembered.
There is also a mild incentive problem. A bold forecast is more quotable than a hedged one. Organisations that publish them often benefit from attention, and a range of scenarios is less likely to make headlines than a single number. This is not an accusation of bad faith, only a description of how attention tends to work.
Reading an automation forecast sensibly
Some questions are useful when you meet one.
- Is it measuring exposure or outcomes? “Tasks that could be automated” and “jobs that will be lost” are not interchangeable.
- Who made the ratings, and how? Expert judgement, model output and survey answers each have different limits.
- What does it assume about adoption speed and cost? If those assumptions are missing, the number is hard to evaluate.
- Does it discuss job creation and change in task mix? A one-sided account is less complete.
- Is there a range or a single figure? Ranges are more honest about uncertainty.
- Can it be checked? A forecast with a clear definition and a near-term test is easier to hold to account.
None of this means that technology does not change work. It plainly does, and some occupations have shrunk or been reshaped over the decades. The point is narrower: the leap from a task-level assessment to a confident statement about future employment contains many steps, and each step can move the answer. Treating such figures as scenarios, as discussed in the broader piece on how claims about the future of work are built, is usually a more accurate way to hold them than treating them as predictions.