Future Labour Markets

The framing under test is usually said in one breath, and it contains two claims with very different evidence behind them. That automation restructures work is about as well established as anything in labour economics: the task framework that replaced skill-biased technical change is the field's dominant paradigm, and the industry-level accounting behind it is not seriously disputed. That we can see how is where the record is bad — not because the data are thin, but because the same data support opposite readings depending on the country, the unit of analysis and the decade. American commuting zones say robots destroy jobs. German administrative records say total local employment is unchanged. French firm data say adopters grow while their industry shrinks. All three are competently estimated and all three are about robots.

What survives that disagreement is not an employment level. It is a distributional shape, and it is the most replicated result in the subject: cohort-neutral and incumbent-destructive. AT&T mechanised more than half the US telephone network between 1920 and 1940 and later cohorts of young women were absorbed by clerical and service work, including categories that did not exist before — while the operators who held the job were, a decade on, more likely to be in lower-paid work or out of work entirely. Dutch robot adopters gain hours and value added while the routine blue-collar workers inside them lose earnings. German robot adoption leaves the local employment count intact by reassigning incumbents onto other tasks. Ridesharing entry left high-earning taxi incumbents alone and pushed low-earning ones out. Four settings, four data sources, one shape.

This brief owns the displacement half of the task model — employment, wages, the labour share, task reallocation, and the record of who bore the adjustment. The productivity half — output per hour, the firm-level effect sizes, and whether national accounts can see any of it — belongs to AI Driven Productivity, which reads the same literature from the other end. Where a source here also carries a productivity number, the number is named and handed over rather than interpreted.

Established State the replicated shape first, because it is the one finding here that does not depend on which country you look at. Feigenbaum and Gross reconstruct AT&T's mechanisation of telephone switching across more than half the US network between 1920 and 1940, using genealogy-based census linking to follow individuals rather than occupation counts. Future cohorts' overall employment was not reduced — clerical and service work absorbed the young women who would have become operators, including newly created job categories. Incumbent operators bore the cost: a decade after mechanisation they were more likely to be in lower-paid occupations or out of work. Published in the Quarterly Journal of Economics. Cohort-neutral, incumbent-destructive is the cleanest statement of the actual distributional structure in this literature, and almost nobody frames the modern results as a replication of it.

Established The anchor identification is narrower than its reputation, and the narrowness is the point. Acemoglu and Restrepo use industrial-robot adoption across US commuting zones 1990–2007, instrumenting local exposure with robot penetration by industry in other advanced economies. The published Journal of Political Economy version reports that one more robot per thousand workers reduces the employment-to-population ratio by about 0.2 percentage points and wages by about 0.42%; the NBER working paper gives ranges of 0.18–0.34 points and 0.25–0.5%, and the non-technical summary translates this to roughly 5.6 workers displaced per robot in a local labour market. Established Two things the paper does not claim are routinely attributed to it. It is not an aggregate forecast: the design is cross-sectional across local labour markets, so if robots raise national output and that output is spent everywhere, the general-equilibrium offset is differenced out by construction, and the authors say so. And it is not about AI: the exposure measure is International Federation of Robotics installations — welding arms, paint robots, pick-and-place. Handwave Extending the coefficient to software is the step where popular coverage does its work by assertion.

Established The same research group, using an AI-specific measure, finds a null — and the interest runs against the finding. Acemoglu, Autor, Hazell and Restrepo's vacancy-based work reports rapid growth in AI-related vacancies 2010–2018, concentrated in AI-exposed establishments, and those establishments reducing hiring in non-AI roles — alongside no detectable relationship between AI exposure and employment or wage growth at the occupation or industry level. Frontier That is the task framework's own principal authors testing their framework on the next technology and reporting nothing at the aggregate. A null from the people whose research programme would benefit from a positive carries more weight than a null from a sceptic, and it is weighted accordingly here. Frontier The window closes in 2018, which is before generative AI, so it is evidence about the machine-learning wave rather than about what is being deployed now.

Frontier Outside the United States the sign of the employment effect changes, and it changes in ways that are not attributable to robots. The German study on administrative Institute for Employment Research data finds robots destroy manufacturing jobs while total local employment is not reduced, the losses offset by service-sector gains, and incumbent manufacturing workers largely protected — staying with their employers on reallocated tasks rather than being separated. The adjustment margin is within-firm task reassignment, not job destruction. Handwave The obvious mechanism is institutional — works councils, sectoral bargaining, short-time work, a firing-cost structure unlike America's — and that mechanism is not identified by anything in this literature. It is plausible, universally asserted, and untested. Frontier This brief could reach only the journal landing page for the German study; the substantive figures circulating from it are the widely reported ones and are carried here as reported rather than verified against the full text.

Established The Dutch and French firm panels show why country-level summaries of this literature are unsafe: firm-level and industry-level effects carry opposite signs. Dutch employer–employee data 2009–2020: robot-adopting firms gain value added and hours; their competitors lose; and inside adopting firms, blue-collar workers in routine tasks lose earnings and employment while other workers at the same firm benefit. Established In France, only 598 of 55,390 manufacturing firms adopted robots, but those firms were 20% of manufacturing employment. Adopters expanded their own employment while their labour share fell, and industry employment fell overall, because adopters took share from non-adopters. That is the cleanest available demonstration that a study measuring one level of aggregation will mislead about the other, and almost no study in this field reports both.

Established The cross-country panel finds no employment effect at all, and its productivity number belongs to the other brief. Graetz and Michaels, on IFR data for 17 countries, 1993–2007, across 14 industries: robot densification contributed about 0.36 percentage points to annual labour productivity growth against a mean of 2.4% — roughly 15% of aggregate productivity growth — raised value added, raised wages by an order of magnitude less than productivity, raised total factor productivity, and had no significant relationship with total hours worked. It did reduce the employment share of low-skilled workers. Frontier The estimates are as reported; the reading that robots are employment-neutral is contested, because the panel is short, industry-coarse and ends before 2008. Frontier Pointing the other way on AI specifically: across 16 European countries 2011–2019, employment shares rose in occupations more exposed to AI, concentrated among younger and more educated workers, with little wage relationship. Opposite sign to the robot literature, on a newer and more contested exposure measure.

Frontier The quantitative claim that does the most work in this field is a decomposition, not an employment count. Acemoglu and Restrepo attribute 50% to 70% of changes in the US wage structure over four decades to relative wage declines of groups specialised in routine tasks in rapidly-automating industries. Published in Econometrica, and contested by everyone who prefers markups, institutions or unions as the primary driver; the decomposition is only as good as the exposure measure underneath it. Frontier A newer extension argues automation preferentially targets high-rent tasks, so a displaced worker loses the rent attached to the task as well as the task, and estimates that this accounts for 52% of the rise in between-group inequality since 1980, with rent dissipation contributing about a fifth of that. Established The same paper carries a productivity claim — that the resulting inefficiency offsets most of automation's productivity gains — which is AI Driven Productivity's half and is deliberately not developed here. Established Restrepo's own survey states the evidentiary position plainly: there is “wide qualitative support for the implications of task models and the displacement effects of automation” alongside acknowledged “shortcomings of the existing literature”. Qualitative support is not pinned-down magnitudes, and the survey says so.

Established On the newest form of work, the official statistics disagree with each other, and the disagreement is the finding. Abraham, Haltiwanger, Sandusky and Spletzer show that household survey and administrative tax data paint different pictures of self-employment growth — surveys show little of what a gig surge implies, tax records show considerable growth — and the mismatch is one-directional: a large and growing fraction of people with self-employment activity in administrative data have none recorded in survey data, while the reverse mismatch is smaller and not growing. Frontier Katz and Krueger's reconciliation using the Contingent Worker Supplements 1995–2017 lands on a modest upward trend in alternative work during the 2000s, considerably weaker than the mid-2010s narrative — but it reconciles survey instruments that the previous finding identifies as the unreliable half. Established The clean natural experiment is ridesharing's effect on taxi driving, in tax records: entry accelerated dramatically; entrants were disproportionately young, female, White and US-born; many combined platform income with wage employment; low-earning incumbent drivers exited at higher rates while high earners were largely unaffected; and in heavily regulated cities the earnings decline was reduced or absent. That is the replicated shape again, in a fourth dataset.

Established The long-run churn evidence is more pessimistic than its authors' reputation suggests. Autor, Chin, Salomons and Seegmiller reconstruct new work 1940–2018 from patents, census occupation codes and job descriptions, and find the majority of current employment is in job specialties introduced after 1940, with the locus of new work shifting from middle-paid production and clerical roles before 1980 to high-paid professional and, secondarily, low-paid service roles after it. Frontier The uncomfortable half of the same paper is its headline: augmentation and automation innovations are positively correlated across occupations, but automation's demand-eroding effects have intensified over the last four decades while augmentation's demand-increasing effects have not. That is a genuinely pessimistic reading from authors usually read as optimists. Frontier Their 2026 follow-up separates new work from more work: new work draws younger and more educated entrants, commands a persistent wage premium that decays as the expertise diffuses, and emerges in response to specific local demand shocks — which is to say the offset arrives, but it arrives for someone else and at a place and time nobody can pick in advance.

Established The management layer, not the machine, is where the newest displacement mechanism sits, and it has a measured literature of its own. Kellogg, Valentine and Christin’s review in the Academy of Management Annals consolidates a decade of field studies into six mechanisms by which software directs labour — restricting and recommending what a worker may do, recording and rating what they did, and replacing and rewarding on the strength of the score — and argues that the object under contest is no longer the task but the control system. Established The term is younger than the practice: Lee and co-authors coined it in 2015 from ride-hail dispatch, where allocation, pricing and deactivation were already fully software-mediated. Frontier The OECD’s cross-country review finds the practice concentrated where work is already heavily instrumented — warehousing, logistics, platform-mediated services, contact centres — and spreading into salaried office work through productivity telemetry, while reporting no representative incidence estimate for any member economy. A management technology that governs an unknown share of the workforce is being regulated before it has been counted.

Established Where monitoring has been evaluated with a credible design it moves the measured number, and not always the outcome the measurement was justified by. The cleanest positive is theft monitoring in restaurant chains, where staggered installation of transaction-surveillance software reduced employee theft and raised revenue per outlet — a real gain from observation alone. Established The cleanest negative is American trucking, where the federal electronic-logging mandate of December 2017 replaced a paper hours-of-service record with a tamper-resistant device: compliance with driving-hour limits improved sharply and crash rates did not fall, with operations evaluations reporting increases for the smallest carriers. Frontier Levy’s ethnography of the same mandate supplies the mechanism — the device measured the rule rather than the fatigue, and pressure displaced into the parts of the job it could not see. Handwave Generalising either result is the unexamined step: both are single-sector, and the monitoring reviews report that effects turn on the stated purpose of the monitoring and on whether the worker can see what is recorded. Monitoring reliably changes behaviour on the monitored margin and unreliably changes anything else.

Established The bargaining-power claim here is an information claim, and it is documented rather than theorised. Rosenblat and Stark’s driver study set the asymmetry out precisely: the platform sets fares, allocates work, scores performance and can deactivate, while the driver sees a price and a rating and cannot see the allocation rule, the comparison set, or the basis of a deactivation. Frontier Dubal’s Columbia Law Review argument extends this to pay — that personalised, opaque, per-job variable rates are a wage practice rather than a pricing practice, and belong under the heading of algorithmic wage discrimination — a legal characterisation resting on platform examples rather than an estimated wage effect, and carried here as contested. Established The contestability record, by contrast, is concrete: a Bologna court held in 2020 that a delivery platform’s reliability ranking was indirectly discriminatory because it scored a legally protected absence identically to an unexplained one, and Dutch courts applying the General Data Protection Regulation’s Article 22 right against solely automated decisions ordered platforms in 2021, and on appeal in 2023, to explain and to reinstate. Every one of those wins turned on disclosure, which is the variable the employer controls.

Frontier Does the US pattern generalise, or do institutions set the sign? The dominant position treats task displacement as the master variable and the American estimates as the general case; the European position treats automation as close to employment-neutral where labour is protected. Frontier What would settle it is an out-of-sample replication of the local-labour-market design in economies with different institutions, showing the same sign. So far it does not replicate in Germany. Handwave What would settle the mechanism is exogenous variation in labour-market institutions holding robot exposure fixed. Nobody has it, and until somebody does, “works councils explain the German result” is a story rather than a finding.

Frontier Is the whole action at the firm level, with aggregation hiding it? France gives adopters growing while their industry shrinks; the Netherlands gives adopters gaining while competitors and their own routine workers lose. Both are consistent with a world in which automation reallocates employment between firms and leaves regional counts nearly untouched, which would make the commuting-zone design the wrong instrument for the question everyone asks of it. Frontier This is the most under-explored hypothesis in the field and arguably the most likely to be right; what would settle it is a single design reporting firm, industry and regional effects together, which no fetched study does.

Frontier Is reinstatement slowing, so that this time the offset does not arrive? The evidence is the intensifying demand-erosion and flat demand-augmentation over four decades. Speculative What would settle it is another two decades of new-work data, which means the question is not answerable inside the horizon in which it is being asked. Frontier A weaker and testable version is available now: whether the wage premium on new work is decaying faster than it used to, which the 2026 new-work paper measures and which could be tracked forward.

Speculative Is AI categorically different because it automates prediction rather than tasks? Agrawal, Gans and Goldfarb argue prediction is an input to decisions, so whether AI destroys jobs depends on whether the decision is also automated — an organisational choice, not a technological one. Frontier As a framework this reframes the entire exposure-measurement problem; as an empirical claim it is untested, and the AI-exposure literature is consistent with it without discriminating in its favour. Frontier The competing live hypothesis is straightforward complementarity: European employment shares rose in AI-exposed occupations, and Hampole and co-authors find reduced labour demand in exposed tasks offset by productivity-driven increases at AI-adopting firms. Both readings fit the 2011–2019 data, which mostly predates generative AI.

Frontier Is anything measurable happening at all? A minority position — held by technology-aligned commentators and some macroeconomists — holds that this literature measures adoption rather than effect. Its evidence is the AI-vacancy null, the persistence of full employment through the entire robot era, and the failure of the 47% forecast. Frontier It is very hard to distinguish “no effect” from “offsetting effects” without a structural model, which is exactly what the sceptics reject, so the position is difficult to falsify by design.

Speculative Two minority readings deserve stating because they are raised constantly in seminars and almost never in print. The first: the measured robot effect is partly the China shock wearing a robot costume. Robot adoption and import competition hit the same US commuting zones in the same window, and the instrument — foreign robot adoption — is not obviously orthogonal to foreign export capacity. Nothing in the fetched corpus tests this directly; the concern is structural, and what would settle it is robot variation orthogonal to trade exposure. Speculative The second: gig work is a statistical artefact of the tax system rather than a labour-market transformation, with the survey/administrative divergence driven by 1099 issuance thresholds and platform reporting rules changing rather than by work changing. The divergence is one-directional in exactly the way that reading predicts. It is a coherent interpretation of a published measurement finding that almost nobody states out loud.

Established One hole is recorded rather than papered over. This pass could not verify a Japanese industrial-robot labour-market study through the available sources, in the country with the deepest robot stock per worker in the world. Frontier What was verified for Japan is service-sector AI evidence on taxi dispatch, which is a productivity result and sits in AI Driven Productivity. The obvious natural laboratory for automation and employment is a gap in this brief, and naming it stops its absence reading as an absence of effect.

Frontier Does algorithmic management belong in this brief at all, or is it a working-conditions subject wearing a displacement costume? The case for it is mechanical: task allocation, scoring and automated deactivation are the channel through which a measured productivity gain is converted into a smaller headcount or a lower effective hourly rate, and no robot or AI-exposure design observes that channel. Frontier The case against is that no study reached here estimates an employment or earnings effect of algorithmic management itself, with an exposure measure, on linked data. That study does not exist, and it is cheaper to build than it looks: deactivation logs and warehouse quota records are administrative data already held by firms rather than measurements that have to be constructed. Speculative What would settle it is a staggered-adoption design on scheduling or scoring software inside one sector, reporting hours, separations and earnings for incumbents.

Frontier Is the productivity–autonomy trade-off real, or an artefact of which side gets measured? The managerial claim is that discretion costs throughput; the occupational-health claim, running back to the job demands–control literature, is that high demand with low control is the configuration that produces strain. Frontier Both can hold at once; the empirical question is the exchange rate, which nobody has estimated. Throughput studies rarely measure turnover, injury or error; strain studies rarely have output data. Speculative The sharpest available test is an appeal right, because it is cheap to randomise — give one arm of a workforce human review of algorithmic scheduling and scoring decisions and measure output, turnover and appeal rates against a control. Frontier Research on contesting automated content decisions suggests a contest process raises perceived fairness independently of whether the appeal succeeds, which would make procedural design a cheap lever and, for the same reason, a manipulable one.

Established The exposure variable underneath the robot literature is a count of boxes. IFR data record installed units by industry and country, with no capability, price or utilisation information. A 1994 welding arm and a 2015 collaborative robot count the same. Frontier Every coefficient in the robot literature is therefore an effect per unit, in a technology whose per-unit capability has changed by orders of magnitude across the sample window, and no study in the fetched corpus adjusts for it.

Frontier The AI exposure measures are worse, because they are constructed rather than counted. Every AI-labour result rests on an index built by natural-language processing over job descriptions, patent text or task inventories, and different indices produce different occupational rankings. Frontier None is validated against realised automation — that is, against whether the occupations it ranks as exposed subsequently automated. The measure is doing enormous work and has never been scored. Handwave Comparing results across AI studies as though their exposure variables measured the same thing is the routine and unexamined step.

Established Aggregation is a first-order defect, not a technicality. The French panel shows firm-level and industry-level effects with opposite signs from the same technology in the same country in the same years. Established A study that reports only one level cannot tell you the other, and almost no study reports both. Two competent papers can disagree entirely because they aggregated differently, and readers will attribute the disagreement to the countries.

Established The statistical system cannot currently size a form of work that is directly observable in tax records. The household-survey and administrative pictures of self-employment diverge, and the divergence grows. Frontier If the same system is used to detect task-level restructuring — a far subtler object than a second income stream — the confidence attached to those detections should be lower than it is. This is the bottleneck this brief regards as the most consequential, because it is upstream of everything else and is a budget line rather than a discovery.

Frontier Publication incentives cut in an unusual direction here. The task framework is the field's dominant paradigm and its principal authors are prolific. That is not evidence of error. But it means the null results in this brief disproportionately come from the same authors testing their own framework on new technologies — which raises those nulls' weight considerably, and lowers the weight of the positive findings that are the framework's flagship. Established A field whose most reliable nulls are self-administered is in a better epistemic position than one with no nulls at all, and a worse one than a field with adversarial replication.

Frontier Algorithmic management has no exposure variable, and the field is proceeding as though it does. The robot literature counts boxes; the AI literature builds an index over job descriptions. For management software there is no register of installations, no standard taxonomy of what a system does, and no survey instrument that distinguishes a scheduling optimiser from a scoring system from an automated termination rule. Established The most systematic cross-country review of the practice reports it by sector and by case, not by measured incidence. Handwave Treating “algorithmically managed” as a binary worker characteristic — which is how it is asked, where it is asked at all — assumes workers can identify a control system they are by construction not shown.

Established Nothing on this map produces a result this brief waits on, and no typed depends-on edge is claimed. The evidence base here is better identified than most of this category — genuine instruments, administrative panels, a randomised natural experiment in Brazil's neighbours' methodological tradition, and a century-long historical control. What it waits on is measurement infrastructure and one policy instrument, and all three are recorded as typed requirements in section 13 because a legislature or a statistical agency could choose to supply them.

Established The first dependency is a statistical system that can size platform and self-employment work. Two official instruments disagree about the level and the disagreement is growing in one direction. Until that is fixed, every claim about how work is being restructured rests on a system that cannot count a category of workers who file tax returns saying they exist. Frontier This is not a hard research problem; it is a linkage and instrument-design problem with a budget attached.

Established The second is an occupational classification that records tasks rather than titles. The entire task framework is estimated by mapping occupation codes onto task inventories built for other purposes, and the mapping is the weakest link in every downstream estimate. Frontier Where a classification does record tasks, the within-occupation heterogeneity turns out to be large enough to move headline risk estimates by a factor of five — which is precisely the difference between the 47% and the 9% versions of the automation-risk literature.

Frontier The third is a policy instrument that follows the incumbent worker rather than the local labour market. If the replicated shape is cohort-neutral and incumbent-destructive, then aggregate employment statistics are structurally incapable of detecting the harm, and area-based adjustment programmes are targeted on the wrong unit. Speculative No fetched source evaluates an instrument of this design at scale, so its efficacy is an open question rather than a known quantity — but the mismatch between what the evidence measures and what policy targets is not open at all.

Established What this brief explicitly does not wait on is the productivity question. Whether automation has raised output per hour, whether firm-level gains aggregate, and whether the national accounts could see them are taken as inputs here and adjudicated in AI Driven Productivity. The two briefs share a literature and split it at the model's two terms.

Established The highest-value experiment is also the cheapest: report firm, industry and regional effects from one design. France and the Netherlands each estimate two of the three levels and find them inconsistent; nobody estimates all three together. The data to do it exist in at least four countries with linked employer–employee registers. A single paper doing this would resolve the field's largest interpretive dispute without collecting any new data.

Established Second: score the AI exposure indices against realised automation. Take the published indices, freeze their rankings at their publication dates, and measure rank correlation with subsequent occupational employment change. Frontier This has been done once, adversarially, to the 47% index — giving a rank correlation of 0.26 — and never to any of the indices currently in use. It requires no new instrument and it would tell every reader of this literature how much of the variation their measure explains.

Frontier Third: find robot variation orthogonal to trade exposure. The standing objection to the anchor design is that robot adoption and import competition landed on the same American places in the same years, and the foreign-robot instrument is not obviously independent of foreign export capacity. Speculative A design that separates them — a technology-specific subsidy, a differential tariff on capital equipment, a supplier bankruptcy — would either retire the objection or retire a decade of estimates.

Established Fourth: re-run the telephone-operator design on a modern displacement. The method is linked individual records tracking incumbents for a decade after a technology arrives. Administrative registers in the Nordic countries, the Netherlands and Germany make this feasible on robot adoption now, and on generative-AI adoption from roughly 2023. Frontier It is the only design in the corpus that measures the thing the evidence says actually happens, and it has been run once, on a technology that finished displacing people eighty years ago.

Frontier Fifth: exploit the 1099 reporting-threshold changes as a measurement experiment. If platform-work statistics move when reporting rules move and not when work moves, the fringe reading of the gig divergence is right and a large literature is measuring tax compliance. Speculative The threshold changes are dated, legislated and exogenous to individual behaviour; nobody in the fetched corpus has used them this way.

Speculative And one experiment that would be decisive and will not be run: randomise the institution. Exogenous variation in firing costs, works-council coverage or short-time work eligibility, holding automation exposure fixed, is what the German–American divergence needs and what no country will supply. Handwave Every institutional explanation in this brief is an inference from a cross-country contrast, and saying so is more useful than repeating the inference.

Established The machinery that produces the German result, whatever produces it, operates inside firms rather than through labour markets. The measured adjustment margin is within-firm task reassignment: an incumbent stays with the employer and moves onto different work. That requires an employer with the scale to have other work, an information system that knows which worker can do it, and a contract structure under which retaining is cheaper than separating. Handwave Whether the German institutions cause this or merely accompany it is the unidentified step, and it is the step every policy recommendation drawn from the German case depends on.

Established The competing margin, visible in the French data, operates through firm boundaries. Adopters expand and take share; non-adopters contract. The displaced worker is not displaced by a robot at their own workstation but by a robot at a competitor's. Frontier That has a direct engineering implication for policy: a subsidy that raises adoption at the margin increases the reallocation, so an instrument justified by its effect on adopters can be employment-destroying at the industry level while every firm-level evaluation of it reports success.

Established The measurement machinery is the part actually missing, and its specification is not mysterious. Linked employer–employee registers, occupational coding that records tasks, and administrative-to-survey record linkage that resolves the self-employment divergence are three specific, costed, buildable systems. Several countries have one or two. Frontier The country with the best identification in this literature — Germany — has them because a social-insurance system needed them for other reasons, which is a useful reminder that research infrastructure is usually a by-product of administration rather than a research investment.

Frontier On the technology itself, the engineering constraint that mattered for robots was fixturing and part presentation, not the arm. The occupations that mechanised are those whose physical environment could be made regular enough for a machine with no perception to work in it. Speculative If that is right, the correct predictor of automation is not task content but environment regularity, and no exposure index in use measures it. Handwave This is stated as a hypothesis worth testing rather than a finding; nothing in the fetched corpus operationalises it.

Within this map: AI Driven Productivity, the other half of the same task model and the owner of every productivity number named here; Digital Economies, which owns the platform market structures that ridesharing and gig work sit inside; Innovation Ecosystems, where the policy instruments aimed at the displaced places are evaluated; Future Education Systems, which carries the human-capital response the reinstatement effect assumes; Human-AI Integration and Cognitive Enhancement, on the augmentation margin; Robotics in Infrastructure and Autonomous Supply Chains, which describe the deployments this brief measures the labour consequences of; and AI Governance, where the regulatory response is assessed.

Outside it: labour economics and the task-based framework; the economics of technology adoption; public finance, which supplies the tax records that turn out to be the better instrument; historical demography, which supplies the linkage methods behind the telephone-operator design; and industrial relations, which supplies the institutional variation nobody has managed to exploit.

Established The binding institutional constraint in this subject is statistical capacity, and it is unusually easy to state. Two official instruments in the same country disagree about how many people are self-employed, in one direction, by a growing margin. Established That is not a subtle inference problem: it is a linkage failure between a household survey and a tax system that both belong to the same government. Until it is fixed, the labour statistics cannot size a directly observable form of work, and every finer claim about task-level restructuring inherits the failure.

Established Several sources here are interested parties and it is load-bearing to say which way the interest runs. The most prominent outcome-scoring of the 47% forecast comes from a technology-industry-aligned think tank whose institutional position is that automation anxiety is overblown — the interest runs with the finding, which lowers its weight even though the exercise is real. Established Conversely, the AI-vacancy null and the intensifying-automation result both come from authors whose research programmes would be better served by the opposite answer, and both are weighted up for it. Frontier The German-robots result comes from a literature with a professional and national stake in the proposition that European labour institutions work; that does not make it wrong, and it is the reason this brief separates the robust finding from the unidentified mechanism.

Frontier The occupational classification is an institution, not a dataset, and it is optimised for something other than this question. Occupation codes exist to run censuses, price wage schedules and administer immigration systems. The task inventories mapped onto them were built for vocational guidance. Established The single largest divergence in the automation-risk literature — roughly 47% of employment at high risk against roughly 9% — is produced by allowing task heterogeneity within occupations rather than treating the occupation as the unit. A five-fold range in the field's headline number is a fact about a classification system, not about technology.

Frontier And the adjustment institutions are targeted on the wrong unit if the replicated shape is right. Trade adjustment assistance, regional development funds and area-based active labour market programmes all attach to places. The measured harm attaches to individuals who held a specific job at a specific moment, and who are, in the aggregate, invisible because the next cohort's employment is fine. Speculative No fetched source evaluates an incumbent-following instrument at scale, so this is an argument about design rather than a claim about efficacy.

Established The regulatory instruments arrived before the measurements, and they are converging on disclosure rather than prohibition. The European Union’s platform-work directive, adopted in 2024 with transposition due at the end of 2026, requires that workers be told which decisions are taken or supported by automated systems, bars certain inferences from being processed at all, and requires human review of significant decisions including deactivation. Established Sub-national instruments run the same way: California’s warehouse-quota statute of 2021 requires a quota to be disclosed in writing and forbids quotas that prevent legally required breaks, and Ontario’s employment statute obliges employers above a size threshold to publish an electronic-monitoring policy. Frontier None of these is an outcome standard. Each creates a record where none existed, which is the precondition for the evaluation this brief says has never been run.

Established The distributional finding is the one most often dropped from both sides of this argument. The optimists' best historical case — telephone operators, where later cohorts were absorbed without measurable employment loss — is also the case that shows incumbents in lower-paid work or out of work a decade later. The offset operates across cohorts, not within careers. Established Saying “new jobs replaced the old ones” is true of the labour market and false of the workers, and both halves come from the same study.

Frontier The income story and the employment story have different signs and different geographies, which is the deepest problem with the framing. Autor and Salomons, on four decades of cross-country industry data measuring automation as industry productivity movements common across countries, find that automation displaces workers in the industry where it occurs and that those losses are reversed by indirect gains in customer industries and induced aggregate demand — employment reallocated, not destroyed. Established But the labour-share losses are not recouped elsewhere: own-industry labour share falls and stays fallen. An analyst watching employment counts sees restructuring with no net damage; an analyst watching factor shares sees persistent one-way erosion. Both are looking at the same automation, and the framing presumes a single legible restructuring.

Established Inside the firm the incidence is sharper than any regional average shows. Dutch data give routine blue-collar workers at adopting firms losing earnings and employment while their colleagues at the same firm gain. Frontier A regional statistic that nets those two against each other reports approximately nothing, which is what several country-level studies report. The absence of an aggregate effect is compatible with large, concentrated, and entirely real harm.

Frontier The regulation result in the taxi data cuts against a common intuition and is worth stating plainly. In heavily regulated cities the earnings decline for incumbent drivers was reduced or absent. Speculative Whether that is protection worth its price — the same regulation restricted entry for the disproportionately young, female and immigrant entrants the platforms brought in — is a distributional judgement the data cannot make. Both effects are in the same study and pointing at only one of them is the standard move on both sides.

Speculative And the fringe reading deserves an ethical hearing, not just an empirical one. If a substantial part of the measured gig-work transformation is a change in tax reporting rather than a change in work, then a decade of policy argument about precarity has been conducted partly on an artefact — and the workers it was conducted about are neither more nor less precarious than before. Frontier That would not make precarity unreal. It would mean the statistical system's inability to count is doing political work.

Established The terminal position is a declared tie on the level and a clear answer on the shape. On employment levels the evidence genuinely does not converge: American commuting zones show losses, German administrative data show none, French firms show adopters growing while their industry shrinks, a 17-country panel shows no relationship with hours, and European AI exposure shows employment shares rising. Every one of those is competently estimated, and picking a side requires ignoring most of the measurements. Established On distribution the evidence does converge, across a century and four independent designs: cohort-neutral, incumbent-destructive. The field's least contested finding is about who bears the cost, and its most contested is about whether there is a cost to bear.

Established The change of question that follows is not a compromise. “How many jobs will automation destroy?” is answerable only at a level of aggregation where the answer is approximately zero and the question has stopped meaning anything. “Which incumbents, at which firms, lose which rents, and over what horizon?” is answerable with existing data in at least four countries, and it is the question the replicated evidence is actually about. Frontier The field has spent four decades estimating the first and has answered the second by accident.

Frontier The long-run risk is not mass unemployment; it is a permanently mis-specified measurement system. If the harm is concentrated in incumbents and the statistics are cohort-level, then a society can run a large and continuous transfer from specific workers to consumers and shareholders while its instrument panel reads normal. Established That is not a hypothetical: the labour-share erosion is measured, persistent, not recouped, and invisible in the employment series everyone watches. Speculative A civilisation that automates for a century under those instruments will not notice the distributional history it is writing until it reads it backwards.

Speculative And the deepest open question is whether new work can be produced deliberately. The evidence says new work emerges in response to specific local demand shocks, carries a wage premium that decays, and draws younger and better-educated entrants. Nothing in it says how a government causes new work to appear, where, or for whom. Handwave A policy recommendation that reduces to “the economy will invent new jobs” is an observation about the past stated as an instrument, and this brief does not pretend otherwise.

These horizons track data availability, statistical reform and technology diffusion rather than capability:

• 10 yr: Frontier The generative-AI cohort becomes measurable. Mass adoption dates from 2023, administrative registers publish with a two-to-four-year lag, and the telephone-operator design needs a decade of follow-up — so the first credible incumbent-tracking study of AI displacement lands late in this window, not early. Frontier Expect the aggregation dispute to be settled before the level dispute, because it requires no new data. Speculative Expect the self-employment measurement divergence to persist, because fixing it requires two agencies to agree and neither is penalised for the disagreement. Frontier The European transposition deadline for disclosed and human-reviewable algorithmic management falls inside this window, which turns a policy argument into a cross-border research design.
• 25 yr: Speculative If the reinstatement slowdown is real, its signature over this horizon is a falling wage premium on new work rather than a falling employment level — a measurable prediction that distinguishes the pessimistic reading from the optimistic one. Speculative The plausible split is that identification improves wherever linked administrative data exist and stays absent for cross-country questions, because 17-country panels cannot be fixed by adding countries. Handwave Forecasting which institutions survive to condition the outcome is asserting a political result, not extrapolating a measurement.
• 50 yr: Speculative At this horizon the interesting object is not employment but the labour share, which has fallen in own-industry terms without recouping for four decades already. Handwave Extrapolating that trend is a straight line through a series with two structural breaks in it, and nobody should present it as a forecast. Speculative The one durable claim available is compositional: whoever holds the automatable job at the moment of automation bears the cost, and that has been true of every episode in the record.
• 100 / 250+ yr: Handwave Beyond useful forecasting. The single datum at that horizon is the mechanisation of telephone switching, which took twenty years, absorbed the next cohort, and damaged the incumbents. Handwave One episode in one industry in one country is a story rather than a base rate, and the honest thing to do with it is refuse to annualise it.

• Depends on Nothing on this map. This brief waits on no result another brief produces: the identification here is better than most of this category, and what is missing is measurement infrastructure and one policy instrument rather than a discovery. No typed depends-on edge is claimed.
• Requires (not on this map) Four things a government could choose to supply and has not. First, a statistical system that can size platform and self-employment work: household surveys and administrative tax records in the same country disagree about the level, the mismatch is one-directional, and it is growing — a large and rising share of people with self-employment activity in tax data have none recorded in survey data, while the reverse mismatch is smaller and flat. A statistical system that cannot count a category of workers who file returns saying they exist cannot be trusted to detect task-level restructuring, which is a much subtler object. Second, an occupational classification that records tasks rather than titles: the entire task framework is estimated by mapping occupation codes onto task inventories built for vocational guidance, and allowing task heterogeneity within occupations rather than treating the occupation as the unit is what moves the headline automation-risk figure from roughly 47% of employment to roughly 9% — a five-fold range that is a fact about a classification system rather than about technology. Third, an adjustment instrument that follows the incumbent worker: the replicated distributional finding is cohort-neutral and incumbent-destructive, so aggregate employment statistics are structurally incapable of detecting the harm and place-based adjustment programmes are targeted on the wrong unit. Fourth, a contestable record of algorithmic employment decisions: where scheduling, scoring, task allocation and deactivation are software-mediated, neither the worker nor the statistician can see the decision rule, and the disclosure duties now being legislated create the written record that is the only realistic route to an exposure variable for the management layer. None of the four is a research result; all four are line items, and the country with the best identification in this literature has its registers because a social-insurance system needed them for other reasons.
• Enables In principle every assessment on this map that assumes a workforce — the deployment briefs, the education and adjustment questions, the distributional arguments in the rest of Category IX — rests on knowing who bears the cost of a technology transition. No typed enabling edge is claimed, and the reason is a finding rather than modesty: the incumbent-following instrument this brief identifies as missing has never been evaluated at scale, so the enabling relationship has never been measured.
• Adjacent Labour economics and the task framework; the economics of technology adoption; public finance, which supplies the tax records that turn out to be the sharper instrument; historical demography, which supplies the record-linkage methods; industrial relations, which supplies the institutional variation nobody has exploited; and within this map AI Driven Productivity, Digital Economies and Future Education Systems.

Frontier “47% of jobs are at risk from automation.” This is the most-cited number in the subject and the most misquoted, and it has now been overtaken by its own date. What Frey and Osborne produced was a subjective hand-labelling of a small subset of occupations, propagated across the full occupational set by a Gaussian-process classifier over nine engineering-bottleneck task variables, with occupations scoring above 0.7 pooled as “high risk” and their employment shares summed — a statement about occupations whose task bundles look technically susceptible, over an unspecified horizon of “perhaps a decade or two”. Frontier Scored as a forecast at year eight it fails on level and on rank order. The US added roughly 16 million jobs after 2013; insurance underwriters, near the top of the risk ranking, grew 16.4% between 2013 and 2021 while recreational therapists, ranked least automatable, declined 8.9%; and the rank correlation between computerisation risk score and subsequent job loss was 0.26. Frontier That scoring comes from a technology-industry-aligned think tank whose interest runs with the finding, and is weighted down accordingly — but the method critique is independent: decomposing the scores to task level by linear programming gives compensation and benefits managers 9.1% against their published 96%, finds a judge's 40% score to be entirely procedural tasks with sentencing untouched, and reports current automation levels predicting future vulnerability at a correlation of only 0.23. Established The honest verdict: it was a technical-susceptibility index misread as an employment forecast, and scored as what it actually said it is close to untestable, which is its own indictment. The task-based rework allowing heterogeneity within occupations gives roughly 9% and is the standard counterweight.

Established “The robot studies tell us what AI will do.” They do not. The exposure measure is installed industrial units — welding arms and pick-and-place machines — and the same authors, using an AI-specific vacancy measure, find no relationship with employment or wage growth at the occupation or industry level. Handwave Transferring a coefficient estimated on physical capital in manufacturing to software in services is the step the argument does by assertion, and nobody in the original paper endorses it.

Established “Multiply the robot coefficient by the national robot stock.” This is not a permitted operation. The design is cross-sectional across local labour markets, so any national general-equilibrium offset is differenced out by construction; the authors are explicit that local and aggregate effects are different objects. Established The published estimates support no aggregate employment forecast of any sign.

Handwave “German-style outcomes are available on request, via labour-market institutions.” The German result is real in its own data: total local employment not reduced, incumbents largely retained on reassigned tasks. The institutional explanation for the divergence from the United States is asserted rather than identified — there is no design anywhere in this literature with exogenous variation in labour institutions holding automation exposure fixed. Frontier A policy recommendation built on the contrast is building on a correlation between two countries.

Established “New jobs will replace the old ones, as they always have.” True as an aggregate claim and false as a reassurance, and the same study establishes both. Later cohorts were absorbed after telephone switching mechanised; the operators themselves were, ten years on, in lower-paid work or none. Frontier The strongest modern version of the worry is not that the offset stops arriving but that it has been arriving more weakly: automation's demand-eroding effects have intensified over four decades while augmentation's demand-increasing effects have not — a result published by the researchers usually cited for the optimistic reading.

Established “Official statistics show the gig economy transformed work.” Official statistics disagree with each other about its size, in one direction, by a growing margin. Frontier The careful reconciliation of the survey instruments lands on a modest upward trend during the 2000s, and the instruments it reconciles are the ones the measurement literature identifies as the unreliable half. Speculative The reading nobody says out loud — that the divergence tracks 1099 issuance thresholds and platform reporting rules rather than work — is consistent with the one-directional shape of the mismatch and has never been tested against the dated threshold changes that would test it.

Speculative “The commuting-zone results are really the China shock.” Carried because it is the objection raised most often in seminars and addressed least often in print: robot adoption and import competition hit the same American places in the same window, and the foreign-robot instrument is not obviously orthogonal to foreign export capacity. Frontier Nothing in the fetched corpus tests it directly, so it is neither established nor dismissed here — it is a live structural concern about the field's most-cited design, and the experiment that would settle it is specified in section 6.

Frontier “We can see how automation is restructuring work.” The framing itself is the last misconception. We can see that it restructures, in a given institutional setting, at a stated level of aggregation, in a known decade. Established The sign of the employment effect is not a property of robots; it is a property of robots interacted with labour-market institutions, firm boundaries, aggregation level and period. Established The one thing that is legible across all of them is who pays: not the next cohort, and not the average worker, but the person holding the job when the machine arrives.