Almost every confident claim about automation and employment rests on one of three kinds of study, and they answer different questions. Conflating them is how the same data supports both "half of all jobs are at risk" and "nothing has happened."
Learning to tell them apart is more useful than any individual figure.
A related workplace category is employee monitoring software, which should be evaluated separately from broad claims about automation and job displacement.
The three kinds of study
Exposure studies estimate what share of tasks in an occupation a technology could plausibly perform. They are a measure of technical potential, not of what will happen. They say nothing about cost, regulation, organisational inertia, customer preference, or whether anyone wants the automated version.
This is where the large alarming percentages come from, and they are almost always quoted without the word "exposure."
Administrative and labour-market studies look at what actually happened to employment, earnings and hours. Slower, narrower, and the only category that can tell you about the present.
Forecasts project forward from assumptions. Useful for thinking; not evidence.
When you see a number, ask which of the three it is. Most public argument consists of exposure figures being answered with administrative findings, or vice versa, with neither side noticing.
What the administrative data currently shows
Studies using administrative records and large surveys find little evidence of economy-wide job loss or wage decline despite rapid adoption.
The Budget Lab at Yale finds no clear relationship between AI exposure and unemployment through August 2025. Humlum and Vestergaard link survey-reported ChatGPT use to Danish administrative records across eleven exposed occupations and find essentially zero effects on earnings or hours through 2024.
A survey of nearly 750 corporate executives likewise finds little evidence of near-term aggregate employment declines, though larger companies anticipate workforce reductions while smaller firms expect modest gains.
Consistent with a task-based view, current systems are characterised as primarily augmenting human labour rather than automating it outright.
That is the honest present-tense position. It is not a prediction, and it does not rule out change later.
Why "tasks, not jobs" matters
The most useful analytical move in this literature.
A job is a bundle of tasks. Technology rarely removes the bundle; it removes some tasks and changes the value of the rest. What happens to the job depends on what is left and whether it is enough to constitute a role.
Three outcomes are possible from the same starting point:
The job disappears — the remaining tasks are too few to justify the role.
The job changes — the mix shifts, the title stays, and the skill requirements change underneath it. This is the most common historical outcome and the least visible in statistics.
The job expands — automation reduces cost, demand rises, and more people do the work than before. Counter-intuitive, and it has happened repeatedly.
Which one occurs depends on demand elasticity, not on the technology. That is why technical capability tells you so little about employment.
The distribution matters more than the total
Aggregate stability can conceal substantial disruption.
Employment can hold steady while composition changes completely — some occupations shrinking, others growing, and the people affected not being the same people. "No net job loss" and "nothing happened to anyone" are different claims.
One review notes higher exposure in higher-income economies and disproportionate exposure among women and mid-education workers in some contexts.
Ask of any aggregate figure: who is inside the average.
What to be sceptical of
A precise percentage of jobs at risk. Almost always an exposure estimate presented as a forecast.
Any claim that this time is categorically different. It may be. That claim has also been made about every previous general-purpose technology, and the pattern of task substitution and demand response has held more often than not. Uncertainty is the honest position.
Any claim that nothing will change. The administrative data covers a short window during early adoption, and absence of evidence at this stage is weak evidence of absence.
Vendor research on either side. Companies selling automation and companies selling reskilling both have positions.
Executive expectations quoted as outcomes. What leaders anticipate is a fact about expectations, not about employment — and expectations in this area have a poor track record.
What follows for an individual
Not "learn to code," and not "AI will take your job."
Identify which of your tasks are structured and checkable. Those are the ones most susceptible to assistance or substitution, and the evidence on where AI performs well is fairly clear. See AI tools at work.
Notice where your value is judgment, relationship or accountability. Those have proven durable, and they are the parts of a role that do not appear in a task inventory.
Watch the entry level of your field. Displacement tends to appear first as fewer junior roles rather than as redundancies — which is invisible in employment statistics and highly visible to anyone trying to enter.
Do not restructure a career around a forecast. Build capability that survives more than one scenario. See skill half-life.
What follows for an organisation
Do not plan headcount on projected automation. The measured gains so far are concentrated, modest in aggregate, and heavily dependent on workflow redesign that most organisations have not done.
Do plan for role composition changing. The mix of tasks in a job changes faster than the job disappears, and job descriptions and hiring criteria lag it badly.
Watch your own junior pipeline. If assistance is closing the gap between a new hire and an experienced one, the reason for hiring juniors changes — and if you stop, you have removed your own supply of senior people in five years. This is the least discussed consequence and the most consequential.
The summary worth carrying
Exposure is not displacement. Tasks change faster than jobs. Aggregate stability can hide substantial redistribution. And the administrative evidence to date shows far less employment effect than the discourse implies — while covering a window too short to settle anything.
Anyone who tells you the answer with confidence is telling you about their position, not about the data.
For broader public guidance and background, consult the National Institute of Standards and Technology.