Will AI take my job? What the evidence shows

You have probably seen two kinds of headline. One says a large majority of workers have jobs exposed to artificial intelligence and that tens of millions of roles will be displaced. The other says careful studies of actual labour markets find almost no effect on employment or wages. Both are reporting real findings from serious researchers, which is confusing until you notice that they are not measuring the same thing. *The question resists an answer because AI acts on tasks rather than jobs, and a job is a bundle of tasks, so the alarming headline figures measure task exposure, which is the broadest possible quantity, while what actually happens to a role depends on whether each exposed task resolves into automation or augmentation, and that is settled by how work is organised and what employers decide rather than by what the technology can do.*

This guide separates the three quantities that get reported as one, sets out what the measured evidence currently shows and where it actually conflicts, explains why the pressure is concentrated where it is, distinguishes measurements from projections, and ends with the uncertainties that remain open. It does not predict what will happen to your job, because nobody can do that honestly. It should let you read the next headline correctly.

Three different things get counted as one

Almost all the apparent disagreement between studies dissolves once you notice that three distinct quantities are being reported under the same heading.

Task exposure means a task that an AI system could plausibly touch or accelerate. This is the broadest measure, it produces the biggest numbers, and it is the source of the figures people find most alarming. When a study reports that a large majority of workers have exposed tasks, it is measuring this and only this. Importantly, exposure says nothing about whether anything changes, because a task can be exposed and remain entirely human for reasons of cost, regulation, reliability, or simple inertia.

Automation means the system performs the task with little human involvement. This is what people picture when they hear that AI takes jobs, and it is a much narrower category than exposure. It is also where AI agents, systems that take actions rather than only produce text, matter most, since automating a task end to end generally requires acting on the world rather than drafting something for a person to use.

Augmentation means the system assists a person who remains in the loop. The task still happens, the human still does it, and the time it takes changes.

The critical fact is that the same exposed task can resolve into either automation or augmentation, and which one it becomes is not determined by the technology. It depends on how reliably the tool performs in that specific context, what verification the work requires, what the law demands, what the tooling costs to implement, and what the organisation chooses. This is why a study reporting very large task exposure and a study reporting negligible employment effects can both be correct at the same time. One is measuring what could be touched, the other is measuring what actually changed.

What the aggregate evidence shows so far

Set the projections aside for a moment and look at what has been measured in actual labour markets. The picture is more muted than the discourse suggests, and it is worth stating plainly.

Studies using administrative records and large surveys have generally found little evidence of economy-wide job loss or wage decline attributable to AI, despite rapid adoption. Analysis tracking AI exposure against unemployment through 2025 found no clear relationship. A study linking survey-reported AI use to national administrative records across a set of highly exposed occupations found essentially no effect on earnings or hours. Research covering a period in which more than a third of workers reported using generative AI found small positive wage effects and no statistically significant declines in job openings or employment in exposed occupations.

The measured productivity effects are similarly modest and revealing. One careful measurement found average time savings of a few percent of working hours, with a meaningful portion of that consumed by the time spent reviewing and correcting AI output, and a notable share of workers acquiring new AI-related tasks that did not previously exist. That texture matters: adoption so far looks uneven, partly self-cancelling, and accompanied by new work rather than only the removal of old work.

Gains are also not evenly distributed across kinds of work, and much of the recent change comes specifically from generative AI rather than from the wider field of machine learning that has been automating routine work for decades. Reported productivity improvements are substantial in areas like customer support, software development, and content production, and considerably smaller on tasks requiring deeper reasoning or judgment. This variation is not noise; it maps onto which tasks are the kind AI currently handles well.

Where the evidence actually conflicts

It would be dishonest to present the null findings as the whole story, because they are contested, and the disagreement is worth understanding on its own terms rather than resolving prematurely.

Alongside the studies finding little aggregate effect, other analyses attribute meaningful ongoing job losses to AI, with estimates of tens of thousands of net positions per month in a large economy, concentrated among younger and entry-level workers. These conclusions come from different methods, cover different windows, and rest on different approaches to attribution, which is where the difficulty lies. Isolating AI's effect from ordinary economic conditions, interest rates, sector-specific cycles, and post-pandemic normalisation is hard, and reasonable analysts reach different conclusions from overlapping data.

Two things can be said without overreaching. First, aggregate nulls and concentrated effects are not contradictory: an economy can show no measurable overall displacement while specific segments experience real pressure, because the aggregate averages across a labour market where most work is unaffected. Second, the direction of the disagreement is mostly about magnitude and attribution rather than about whether anything is happening. Almost nobody working with the data claims the effect is zero everywhere, and almost nobody credible claims broad displacement has already occurred.

Where the pressure is real: the entry level

The most consistent finding across otherwise divergent studies concerns who is affected, and it is the part of this literature that deserves the most attention.

Multiple independent analyses identify pressure in the entry-level segments of highly exposed occupations, particularly among younger workers and recent hires. Evidence points to reduced hiring at the start of career ladders, especially where tasks are automatable rather than complementary to what humans add. Descriptions of employer behaviour capture the pattern as slower substitution: organisations reduce entry-level hiring before eliminating existing headcount, because not filling a role is far easier than removing one.

There is a coherent explanation for why this segment specifically, and it is more useful than the observation alone. Work divides roughly into codifiable knowledge, which can be written down and transmitted explicitly, and tacit knowledge, which is acquired through experience and resists articulation. Entry-level work is disproportionately composed of codifiable tasks, since that is precisely what can be handed to someone new: structured research, first drafts, routine analysis, standard documentation. Experienced work is disproportionately tacit, involving judgment about which problem to solve, when the usual answer does not apply, and what the client actually means.

AI is strong on codifiable tasks and weak on tacit ones. So the same technology can substitute for junior work while complementing senior work, which is consistent with the observed pattern of employment pressure at the entry level alongside stable or rising wages in exposed occupations that reward experience. Research connecting patent types to labour demand supports the underlying mechanism, finding that innovations which augment workers increase demand for labour while those which automate reduce it. AI is producing both at once, in different parts of the same occupation.

This raises a real long-term concern that is worth naming without dramatising: if the tasks people traditionally used to build tacit expertise are the ones being automated, the pipeline that produces experienced workers may narrow. Nobody knows yet whether new routes into expertise will emerge, as they have after previous transitions, or whether this one is different. It is an open question rather than a settled worry.

The constraints that slow full replacement

Theoretical capability consistently exceeds observed use, and the gap is instructive. Analyses find that AI could in principle assist with the large majority of tasks in some professional fields, while measured usage sits far below that.

Several constraints explain the gap. Verification cost is the first: when output must be checked by a person, the work does not disappear but changes shape, and the time spent reviewing offsets part of the time saved. This is a direct consequence of the reliability problems covered elsewhere, since a system that is usually right but confidently wrong sometimes requires exactly the human checking that limits how much it displaces. Legal and regulatory requirements are the second, since many decisions must be made or signed by a qualified person regardless of what a machine could produce. Implementation cost is the third, because integrating a capability into workflows, systems, and processes is slow and expensive work that has little to do with the model itself. Organisational inertia is the fourth and is routinely underestimated: institutions change slowly even when the case is clear.

These constraints are not permanent, and some will weaken. But they mean the translation from capability to labour-market effect is slower and more uneven than capability alone suggests, which is one reason predictions keyed to model releases have consistently run ahead of measured outcomes.

Projections, and how much weight to give them

Much of the most-quoted material in this area consists of projections rather than measurements, and the distinction should be kept sharp.

Widely cited forecasts estimate large numbers of roles displaced and larger numbers created over the coming years, netting to substantial job growth. These are scenario models built on assumptions about adoption rates, capability improvement, and organisational response, and they are useful for seeing structure rather than for prediction. Their track record deserves stating: forecasts of technological employment effects have historically been wrong in both directions, generally overestimating displacement speed while underestimating the emergence of work that did not previously exist and could not have been named in advance.

There is one signal in the current data that is a measurement rather than a forecast and is worth more than most projections: the labour market is repricing for AI-related skills well ahead of any aggregate displacement. Job postings mentioning AI skills have grown sharply, and mentions of newer agent-related capabilities have grown faster still. That is employers acting on their expectations with real money, and it is consistent with the augmentation-first pattern the usage data shows.

Reading your own situation

The framework above supports a more useful question than the one in the title. Rather than asking whether AI can do your job, decompose it.

List what you actually spend time on, in tasks rather than in job title. For each, ask three things. Is it codifiable, meaning could it be described precisely enough that someone following instructions could do it, or does it depend on judgment built from experience? Does it require verification, meaning would a person need to check the output before it could be used, which keeps a human in the loop? And is it complementary, meaning does doing it faster increase the value of the rest of your work rather than removing the need for you?

Tasks that are codifiable, unverified, and self-contained are the ones most exposed. Tasks that require judgment, carry accountability, or make your other work more valuable are the ones that tend to be augmented rather than replaced. Most jobs contain both, which is why the realistic expectation for most people is a change in the composition of their work rather than its disappearance, with the mix shifting toward the parts machines handle poorly.

Two observations follow that are worth holding lightly, since they are inference rather than measurement. The value of verification and accountability appears to be rising, because someone must be answerable for output that a system produced and cannot itself stand behind. And the value of knowing what to ask for, which problem is worth solving and what a good answer looks like, appears to rise as the cost of producing answers falls.

The honest uncertainties

Several things remain unknown, and treating them as settled in either direction is the main failure mode in public discussion.

The speed is unknown. Capability has improved quickly while labour-market effects have appeared slowly, and whether that gap closes suddenly or stays wide is not established. The distribution is unknown: aggregate stability can coexist with severe concentrated disruption, and averages conceal exactly the cases people care about. The composition of new work is unknown, since the jobs created by a technology are usually not describable in advance, which makes the created side of every projection much softer than the displaced side. And the trajectory of capability is itself uncertain, as covered in discussions of what these systems can and cannot do and how far current approaches extend toward general capability.

What can be said is narrower and more reliable than the headlines. Task exposure is broad. Measured aggregate displacement to date is limited and contested. Pressure at the entry level is the most consistent finding across studies. Augmentation currently outweighs automation in observed usage. The constraints slowing full substitution are real but not permanent. And the labour market is already repricing for the skills involved.

The short version

The question is hard to answer because AI acts on tasks while employment is organised into jobs, and a job is a bundle of tasks. Three different quantities get reported as one: task exposure, meaning work an AI could touch, which produces the largest and most alarming numbers and implies nothing on its own; automation, where the system performs a task with little human involvement; and augmentation, where it assists a person who stays in the loop. The same exposed task resolves into automation or augmentation depending on reliability, verification requirements, regulation, implementation cost, and employer choice rather than on capability. Measured aggregate effects on employment and wages have so far been limited in most large studies, though other analyses attribute meaningful ongoing losses to AI, and the disagreement concerns magnitude and attribution rather than whether anything is happening. The most consistent finding across studies is concentrated pressure at the entry level, with reduced hiring at the start of career ladders, which is explained by entry-level work being disproportionately composed of codifiable tasks while experienced work is disproportionately tacit, so the same technology substitutes for junior work and complements senior work. Widely quoted forecasts of displacement and creation are scenario models rather than measurements, and such forecasts have historically misjudged both speed and the emergence of new kinds of work.

The idea to hold onto is that AI does not take jobs, it takes tasks, and what happens to a job depends on which of its tasks are exposed, whether each becomes automated or augmented, and what that does to the value of the work that remains, which is why the honest answer to the headline question is a decomposition rather than a yes or no.

Common questions

Will AI take my job? No honest answer to that question exists at the level of a whole job, because AI acts on tasks rather than roles and every job is a bundle of tasks. The useful version is to ask which of your tasks are exposed, and for each whether it will be automated, meaning done with little human involvement, or augmented, meaning done faster with you still in the loop. Tasks that are codifiable, require no verification, and stand alone are most exposed. Tasks requiring judgment, carrying accountability, or making your other work more valuable tend to be augmented. Most jobs contain both, so the realistic expectation for most people is a change in the composition of their work.

How many jobs will AI replace? Nobody knows, and the widely quoted figures are projections rather than measurements. Prominent forecasts estimate large numbers displaced alongside larger numbers created, netting to job growth over the coming years, but these are scenario models resting on assumptions about adoption speed, capability, and organisational response. Forecasts of technology's employment effects have historically erred in both directions, usually overestimating how fast displacement happens and underestimating work that did not exist before and could not have been named in advance. Measured effects to date are considerably more modest than the projections, though this is contested.

Why do studies about AI and jobs contradict each other? Mostly because they measure different quantities. Task exposure counts work an AI could touch and produces the largest figures while implying nothing about job loss. Automation counts tasks actually performed without human involvement. Augmentation counts tasks where a person remains in the loop. A study reporting very high exposure and one reporting negligible employment effects can both be correct. The remaining disagreement, over analyses attributing ongoing job losses to AI versus those finding null effects, comes from different methods, time windows, and approaches to separating AI's influence from ordinary economic conditions, which is difficult in practice.

Which jobs are most at risk from AI? The pattern in the evidence is less about whole occupations than about task composition, but entry-level positions in highly exposed fields show the most consistent pressure across studies. The explanation is that entry-level work is disproportionately made of codifiable tasks, the kind that can be described precisely and handed to someone new, such as structured research, first drafts, and routine analysis. Experienced work is disproportionately tacit, involving judgment learned through practice. Since AI is strong on codifiable work and weak on tacit work, the same technology can substitute for junior roles while complementing senior ones in the same occupation.

Is AI creating jobs as well as destroying them? The evidence suggests both are happening, though the created side is harder to measure. Direct evidence includes workers acquiring new AI-related tasks that did not previously exist, and a sharp rise in job postings requiring AI-related skills, which is employers acting on expectations with real budgets. Historically, new work created by a technology has been difficult to anticipate, which is why the creation side of any forecast is softer than the displacement side. What is measurable now is that the labour market is repricing for AI skills well ahead of any aggregate displacement, which is consistent with augmentation being more common than replacement so far.

Why hasn't AI replaced more jobs already, given how capable it is? Because theoretical capability and actual deployment are separated by several practical constraints. Verification cost is significant: when output must be checked by a person, the work changes shape rather than disappearing, and reviewing consumes part of the time saved. Legal and regulatory requirements mean many decisions must be made or approved by a qualified person regardless of what a machine could produce. Integrating a capability into existing workflows and systems is slow and expensive work unrelated to the model itself. And organisations change slowly even when the case is clear, a pattern visible in how often agent deployments fail for operational rather than technical reasons. These constraints are real but not permanent, which is one reason predictions tied to model capability have run ahead of measured outcomes.

What skills matter most as AI becomes more capable? The evidence supports a general direction rather than a specific list, and it should be held lightly. Work that requires judgment built from experience, rather than following describable procedures, has been more complementary to AI than substitutable by it. Accountability appears to be rising in value, since a person must be answerable for output that a system produced and cannot stand behind itself. Knowing what to ask for, which problem is worth solving and what a good answer looks like, becomes more valuable as producing answers gets cheaper. And practical understanding of what these systems can and cannot do reliably is itself increasingly in demand, which is visible in how job postings have changed.