Reading List

The third vintage of the most-cited paper in this field, now with ADP payroll data through June 2026 — and it changes its own headline measure. Earlier versions led with a regression estimate adjusting for firm shocks (13%, then 16%). This one leads with the simpler descriptive number that needs no modeling choices: employment of 22-25 year olds in AI-exposed occupations stands 19% below where it would be had it kept pace with less-exposed peers, up from 15% on the same measure a year ago. Experienced workers show no comparable gap, and Fact 1 remains that there is no economy-wide displacement — the ADP sample grew about 6%. The most useful thing here is the authors arguing against themselves. Education is the one control that bites (the gap attenuates from -18pp to -9pp), and they present the two estimates as bracketing a range rather than picking the flattering one, because generative AI substitutes best for exactly the codified knowledge schooling produces. They also concede the magnitude is ADP-specific: the ACS gap is -2.2pp with a confidence interval spanning zero against -13.2pp in ADP, though the two agree closely within white-collar work. Adjustment runs through hiring, not separations or pay.

The NY Fed's director of research uses the inaugural post of a new commentary series to state how his own shop reads its AI evidence: the labor-market effect so far "has more to do with changing skill requirements than eliminating jobs." The one numeric table, from the August 2025 Regional Business Surveys, shows Second District adoption climbing quickly — service firms from 25% using AI in 2024 to 40% in 2025, manufacturers from 16% to 26%, with 44% and 33% expected within six months — alongside the finding that firms report very few AI-driven layoffs and "overwhelmingly intend to retrain workers rather than fire them." The counterweight Athreya does not soften: firms anticipate more reductions in hiring plans ahead, especially for college-educated workers. Worth reading as institutional interpretation rather than new measurement. Everything quantitative is restated from earlier work, the survey is regional rather than national and so is not comparable to the Census BTOS series jobsdata plots, and the retraining chart carries no labeled values. His closing speculation is the more interesting half: that specialization leaves each of us exposed to a sudden collapse in the value of the one skill we sell, and that AI's demand-side effects may ultimately matter more than its absorption into production.

A quietly destabilising paper for anyone tracking AI's labor effects through professional profile data — which is most of us. Using monthly vintages of Revelio Labs data from 2020 to 2026, the authors find 19.7% of established US LinkedIn users go back and rewrite the title or description of a job they have already left. The records are "not fixed historical snapshots, but mutable accounts that workers revise over time." Edits cluster around job changes, and AI-related language in them jumps sharply after 2022. Two of the four authors are at Revelio, which makes this a vendor documenting its own limitation rather than an outside attack. The implication worth sitting with: an exposure measure built on profile text partly captures workers relabelling old work in new vocabulary, not the work itself changing.

The assumption that AI-written books are slop buyers will ignore turns out to be wrong, and the mechanism of harm is not the one people expected. Across 14,419 self-published Amazon genre-fiction titles matched to daily sales, quarterly selling titles grew 19.2-fold while revenue grew only 8.9-fold — the market added books far faster than it added money, and revenue per title fell across most genres. Books with detected AI text win a growing share of the scarce top-rank positions. Human-authored books lose the most ground precisely in genres where AI has diffused furthest. None of the books disclose AI use. The conclusion is that generative AI "can reshape a creative market through scale rather than quality" — dilution, not replacement.

Stanford's economic policy institute weighs the apocalypse framing against the data and finds it wanting. The headline number: since 2022 unemployment among the most AI-exposed quintile of workers rose 0.77 points — while the least-exposed quintile rose 0.85. Exposure has not translated into job loss. Software developer postings are growing faster than the average occupation, and firms that adopted enterprise AI added 10% headcount over the following two years. The one genuine soft spot is new graduates, where unemployment hit 5.6%, though the brief notes entry-level declines only become visible in 2024, which complicates blaming a late-2022 model release.

Not findings but plumbing, and useful for it. Three Fed economists catalogue the public indicators worth watching as AI diffuses, ordered the way general-purpose technologies actually arrive: capabilities and costs first, then firm investment and adoption, then productivity and labor. The note is candid that reported adoption overstates real usage — intensity “remains shallow even where reported adoption is broad” — and that the absence of an aggregate signal in 2026 says little about 2030. A good map of which series to trust.

A monthly tracker across five lenses — talent supply, demand, equilibrium, work content, matching — built on online professional profiles rather than payroll. Its headline finding independently replicates Canaries on non-ADP data: early-career workers (22-25) in the most AI-exposed occupations are down 13% relative to the least exposed since pre-ChatGPT, versus about 4% for all ages. Demand for the most-exposed roles is down 42% relative to the least exposed. The firm-side picture cuts the other way: AI-adopting firms grow headcount 27% more than non-adopters (though they were already growing faster before adopting), gains concentrate in senior roles (+31% vs +6% junior), and more AI-exposed firms see fewer layoffs, not more. Two series no one else publishes monthly: a within-occupation activity-mix dissimilarity index (+8.4pp yoy, with most change happening inside occupations rather than between them) and a matching-efficiency series (5.05 postings per hire, +264% yoy).

The second entry in OpenAI's Work at the Frontier series asks not whether AI can do a task, but who ends up doing it. Using 800,000+ work-related messages from US ChatGPT users whose occupations were linked via ChatGPT Business role data, each message is classified to a single O*NET detailed work activity and compared with the sender's stated occupation. The result they call task crossover: 16.8% of all work-related messages, and 43.5% of occupation-specific messages, concern tasks historically associated with another occupation. Cross-occupation work is the majority of occupation-specific messages in five of eight groups (customer experience 77%, design 75%, HR 69%, legal 56%, marketing 53%). The structurally useful finding is that borrowing and lending are independent: designers draw 35.2% of their messages from other fields while design tasks are only 1.7% of everyone else's; engineering is the inverse (18.5% borrowed in, 7.4% traveling out). Marketing leads both directions. Two tasks travel universally — calculating financial data and troubleshooting computer applications rank top-three in all seven non-native occupation groups. The authors are explicit that this is descriptive evidence about the division of labor, not an employment or productivity estimate, but the measurement implication is sharp: if AI changes which workers perform which tasks, statistics built on existing job descriptions will progressively diverge from how work is actually organized.

FRED incorporated 137 data series on US generative AI adoption from the Bick-Blandin-Deming Real-Time Population Survey (category 8) — the quarterly module fielded since August 2024 that underpins the GenAI Adoption Tracker and nearly every major adoption synthesis (Fed Board, Stanford DEL, IMF, OECD). The series cover usage rates (overall, work, nonwork), reported time savings, GenAI-assisted work hours, and adoption compared with the PC and internet, broken out by industry and occupation. Latest FRED readings: 43.4% of employed adults used GenAI for work and 57.9% of working-age adults used it anywhere in Q1 2026; GenAI-assisted work hours reached 6.3% in Q2 2026, up from 4.1% in Q4 2024; reported time savings hit 2.2% of work hours. FRED hosting matters beyond convenience — it makes the field's cleanest worker-side adoption series citable, vintage-controlled, and API-accessible alongside official government statistics.

Google completes the usage-data triad: 15M Gemini interactions mapped to O*NET and ATUS. Diffusion broad but shallow — usage spans occupations covering 88% of US employment yet median task saturation is 21%; end-to-end automation is <10% of non-routine cognitive use. Adds household lenses no other dataset covers: 86% of conversational use is non-work, government/civic queries over-index ~20x, household time savings worth $15-149B/yr. Reviewed by Coyle and Autor.

Monthly refresh of the Canaries Dashboard (anonymized ADP payroll data, 4.6M workers, 730+ occupations; balanced five-year firm sample), with data through June 2026 and a new gender decomposition published July 22. At the aggregate level the story softened: every AI-exposure quintile now shows employment growth since ChatGPT, though the most-exposed grows just +1.1%/yr annualized vs +2.0%/yr for the least-exposed (most-exposed is -0.2% on a year-over-year basis). The early-career divergence persists but eased slightly: workers 22-25 in the most-exposed quintile are contracting at 3.5%/yr annualized (from 3.8%/yr in the April reading), and -4.3% over the past year. The new gender analysis shows early-career women in the most-exposed quintile contracting at 4.5%/yr vs 2.5%/yr for men, driven primarily by occupational composition — 43.8% of early-career women work in the most-exposed occupation category vs 32.4% of men. The lab frames results as correlation, not causation: with the broadest controls, exposure-linked declines only become significant in 2024.

FT synthesis of the PwC 2026 Global AI Jobs Barometer: entry-level roles are being 'seniorised' rather than eliminated. Newcomers used to learn by doing simple repetitive work (the apprenticeship rung) — AI took exactly that. What remains for entry-level workers now leans on judgment, idea generation, and interpersonal skills — historically senior. Job openings for seniorised entry-level roles have grown 35% since 2019 while other entry-level roles shrank 10%; entry-level roles most exposed to AI are 7x more likely to require traditionally senior-level skills.

A landmark statement signed by 16 Nobel Laureates and 200+ economists and AI researchers, organized by Brynjolfsson, Agrawal, Korinek, and Cunningham. Warns AI could drive an economic transformation larger than the Industrial Revolution but on a vastly shorter timeline, with large-scale job displacement as a primary risk. Calls for deeper research, policy frameworks, and institutions to ensure AI complements human capabilities. Korinek: 'Steam, electricity, and computers each gave societies decades to adapt; AI may give us only a few years.'

Casselman synthesizes the AI measurement problem: different data sources give contradictory answers on basic questions — is AI causing job losses or gains, which workers are most exposed, is the productivity boom real? Highlights: a new Yale Budget Lab monthly 'occupational churn' measure designed as an early-warning system that tracks how the composition of an industry's jobs shifts before total headcount does; new Ramp/Revelio Labs research showing the companies using AI most intensely are ADDING jobs faster than laggards (opposite direction from displacement narrative, but Ramp's clients skew tech-savvy so representativeness is uncertain); a Nathan Goldschlag/Economic Innovation Group report on the measurement challenge with concrete policy recommendations; and a bipartisan Senate bill (Sen. Mark Kelly, D-AZ) that would expand federal AI labor data collection and mandate an annual federal report. Notes the underlying federal statistical system is deteriorating from falling response rates and funding cuts (former BLS Commissioner McEntarfer says $10M/yr would help). Frames current confusion as J-curve territory — most firms still on the downward experimentation phase before productivity gains materialize. Companion to Casselman's June 10 'Hidden Workers' piece; the current article is a stronger external validation of jobsdata.ai's own operating premise (show the source mix, don't collapse to a single number).

Chandar (coauthor of the 'Canaries' paper) explains why he was one of 5 of 16 economists on a WSJ panel to predict AI would cause net job loss — alongside Acemoglu, Henderson, Restrepo, and Wolfers. All 16 agreed AI would boost productivity; 8 predicted no change in net jobs and 2 predicted net growth. The 5 net-loss economists also unanimously said AI would replace rather than complement workers and would reduce demand for white-collar jobs. Chandar's key clarification: his prediction is not a 'jobs bloodbath' story. He expects AI to make the economy rich enough — via capital income from AI assets or transfers — that the income effect dominates the substitution effect, lowering long-run labor force participation. Historically, prime-age LFP has been flat since the 1960s and labor-supply adjustments have come through hours, not participation. He frames Kinder's 'messy middle' as the short-to-medium-run risk lens. A useful disambiguation of what mainstream labor economists actually mean by 'AI net job loss.'

A bipartisan consortium called RAISE US launches with a 'people strategy' for the AI era, led by former Commerce Secretary Gina Raimondo (CEO) and former Indiana Gov. Eric Holcomb. Founding employers include Amazon, Microsoft, Bank of America, and Eli Lilly; OpenAI and Anthropic are involved; MIT economist David Autor sits on the advisory board. The group has raised more than $500M — roughly half its multiyear goal — and will initially work with state governments in Arkansas, Maryland, Utah, and Connecticut. Beyond conventional retraining, the mandate explicitly includes revisiting decades-old policies like unemployment insurance (e.g., letting displaced workers keep benefits while starting AI-enabled businesses) and developing corporate incentives for employers to retain and reskill workers whose jobs AI disrupts. Holcomb: 'The scale of change and reshaping of jobs and careers is different — and more so than it has ever been, I think, in human history.' Notable for jobsdata.ai readers because it is the first large bipartisan, multi-state institutional response treating AI labor disruption as a policy problem rather than a technology problem — and because the founding employers are the same firms whose AI deployments are driving displacement signals tracked on this site.

The Economist argues the center of gravity in AI economics is shifting out of universities and into AI labs and government statistical offices. Academic economists have been comparatively slow: the share of NBER papers on AI is rising but not fast, and even in 2024 COVID-related papers outnumbered AI ones. A core reason is skepticism — a survey by Basil Halperin and colleagues found that even under a 'rapid' AI scenario by 2030 (AI matching or surpassing the brainiest humans), the median academic economist expects US GDP growth of just 3.5% in 2050, versus 5.3% for AI researchers, and only 11% of leading economists (University of Chicago IGM survey) agree AI will substantially raise unemployment in advanced countries over the next decade. Daron Acemoglu's early-2024 model (1,000+ citations) implies modest aggregate productivity gains, a result Tyler Cowen attributes to its assumption that AI won't do new things. Meanwhile labs are hiring economists: Anthropic appointed Anton Korinek, OpenAI hired Ronnie Chatterji as chief economist, and Google DeepMind hired Alex Imas as 'director of AGI economics.' By 2019 more than two-thirds of AI researchers worked in industry (up from under half in 2001, per Ufuk Akcigit), and junior lab economist roles can pay $300,000+. The piece flags conflicts of interest and the risk that AI economics drifts from open science toward proprietary innovation.

Casselman reframes the AI-displacement debate away from software engineers toward the larger, quieter population economists worry about most: customer service representatives, bookkeepers, payroll clerks, and HR specialists — tens of millions of jobs, disproportionately held by women, many without college degrees. Molly Kinder: 'I worry that AI will be to high-school-educated women what deindustrialization was to high-school-educated men.' Cites Northwestern's Yin & Ogut reweighting showing usage-based exposure measures understate impacts on workers without degrees, older workers, and people of color; GovAI's exposure-times-adaptive-capacity framework identifying back-office workers as both highly exposed and least able to adapt; and Muro/Heck's 'gateway jobs' research on AI carving out the career ladder's middle rungs. Balances with Forsythe's caution that prior automation waves created jobs, and notes there is still little firm evidence AI has hurt the labor market as a whole.

Stanford's Digital Economy Lab launched a monthly-updated public dashboard suite tracking AI's real economic footprint — the live, continuously refreshed successor to the Canaries research. Three components: an Employment & AI Exposure dashboard built with ADP Research on payroll records covering millions of workers at thousands of private companies; the Canaries dashboard tracking early-career employment in AI-exposed occupations, now showing a 16% relative employment decline for workers aged 22-25 in the most-exposed roles, concentrated where AI automates rather than augments; and a Takeoff Tracker scanning 12 macroeconomic indicators (productivity, capital share, energy use) that currently read mostly neutral — no macro takeoff visible yet. A standing, citable answer to 'what does the data show right now?'

A four-expert panel — Daron Acemoglu, Dean Ball, Ethan Mollick, and Clara Shih — on how workers should prepare for a hybrid AI-human workforce. Mollick cites a Procter & Gamble experiment with 776 employees in which individuals using AI performed as well as two-person teams without it, and warns that the apprenticeship model for training junior workers has 'all collapsed.' Acemoglu challenges the agent-supervisor vision of work ('How many Marcus Chens can the American economy employ?') and argues AI investment is misdirected toward automation rather than augmenting trades facing shortages — a novice electrician with the right AI tool could be 10x as productive. Shih reports a 'tale of two cities' in entry-level hiring: candidates fluent in AI agents land jobs while others see those roles disappear. Consensus advice: curious generalists, AI-augmented skilled trades, and owning projects end to end.

A well-reported case study of Schneider Electric (160K employees) choosing augmentation over replacement. In Q4 2025, AI answered 75% of 150K customer service queries correctly — but agents still review and deliver every response, preserving headcount while cutting response times. On the factory floor in Le Vaudreuil, AI optimized silver-tip washing cycles (73% waste reduction) and quality inspection without eliminating operators. Erik Brynjolfsson frames the thesis: businesses can reap bigger gains by making workers productive than by cutting them. The counterpoint emerges from within: Schneider's own AI-assisted workforce developed a plug-and-play contactor that no longer requires an electrician to wire. Anton Korinek (now at Anthropic) voices the deeper concern: the direction of AI development is increasingly hard to steer, and the augmentation window may narrow. A useful companion to the Goldman Sachs op-ed from the same week — two different frames for the same optimist position, one from a CEO and one from the factory floor.

Goldman Sachs CEO argues the 'job apocalypse' narrative overshoots. Acknowledges Goldman economists estimate AI may automate 25% of current work hours over the next decade, and cites a Stanford finding that entry-level employment in the most AI-exposed occupations has already declined 16% relative to least-exposed roles. But argues this won't translate to net job elimination: complexity expands to fill freed capacity (his own analyst example — stock charting went from 6 hours to seconds, yet Goldman employs more people than ever), cultural preferences preserve human roles (ATMs didn't reduce bank employment), and gross US job churn of 25–35M annually dwarfs net creation of a few million. Notes Goldman's own data center demand has created 200K+ construction jobs since 2022. Calls for joint public-private reskilling investment. A Tier 3 opinion piece, but notable as a major CEO staking out the optimist position with specific internal data.

Stripe's internal data on two AI-economy signals. First, new business formation: total US business applications are accelerating but 'high-propensity' (likely-to-hire) applications aren't keeping pace — Stripe Atlas data confirms solo founders are the overwhelming driver of the acceleration, especially in Q1 2026, consistent with AI lowering barriers to solopreneurship. Second, the SaaSpocalypse: when software stocks shed ~$1T in early 2026 on AI disruption fears, Stripe's pay-in volumes for the 100 largest non-AI SaaS companies showed only a brief dip and swift recovery — the sell-off was expectations, not current activity. Also frames the Solow paradox for AI: the near-term absence of aggregate productivity acceleration shouldn't be read as evidence that none is coming, citing electrification's 30-year lag before productivity doubled.

NY Fed researchers combine Anthropic's AI exposure measure with Lightcast vacancy data (through January 2026) and BLS OEWS employment data to test whether AI is already affecting hiring. Key finding: while AI-exposed occupations show relative vacancy declines, the divergence began before ChatGPT's release in late 2022 and shows no additional break in trajectory afterward. There is no divergence between junior and senior positions within highly exposed occupations — undermining the thesis that AI is specifically hollowing out entry-level roles. The authors conclude that 'while AI may be contributing to recent labor market developments, it is not the main driver of the slowdown in hiring.' NY Fed business surveys indicate firms intend to incorporate AI mainly via retraining, with limited effects on hiring. An important empirical check on the displacement narrative from a Tier 1 government source.

The first papal encyclical to treat artificial intelligence as a central topic. Chapter Four directly addresses AI and labor: warns of a 'significant and rapid contraction in available jobs' creating chain reactions for families and local economies (¶151), notes wealthy societies 'automate rapidly and chaotically, reducing the need for a workforce' while poorer regions are trapped in hybrid economies of underpaid labor and partial technology (¶153), and flags wage polarization — 'outsized remuneration for a highly specialized minority alongside declining wages for a large portion of the workforce' (¶151). The encyclical also warns AI can 'paradoxically de-skill workers, subject them to automated surveillance and relegate them to rigid and repetitive tasks' (¶150). Frames the policy response around subsidiarity, integral human development, and the principle that 'the pursuit of greater profits cannot justify choices that systematically sacrifice jobs' (¶152). Builds on Leo XIII's Rerum Novarum (1891) tradition of Catholic Social Doctrine applied to labor.

A 'two realities' essay on the widening gap between AI's frontier capability and its enterprise implementation. On the frontier: an OpenAI model made progress on the Erdős unit-distance problem — given n points in a plane, how many pairs can be exactly one unit apart — a question open since 1946, finding an infinite family of constructions via algebraic number theory through cross-domain reasoning rather than brute-force search, a Lean formalization, or a recovered proof. Outside mathematicians reviewed it and Fields Medalist Timothy Gowers called it 'a milestone in AI mathematics.' On the floor: Starbucks rolled out an AI inventory-counting tool across more than 11,000 North American company-owned stores in 2025, then scrapped it nine months later over persistent inaccuracies (including confusing similar milk types), reverting to manual counting and daily replenishment. Ojha's central claim is that capability does not automatically become productivity — it must pass through incentives, politics, procurement, org charts, and the competence of the people implementing it. She warns of 'metric theater': companies routing existing workflows through AI features, bundling AI into existing contracts, and subsidizing adoption, then reporting usage numbers that, like adjusted EBITDA, require reading the footnotes. The firms that benefit most won't have the loudest mandates or highest internal usage numbers; they'll redesign workflows, train people, set real standards, and measure actual output.

Strada Institute (with Artemis Strategy Group) surveyed 1,498 US executives and senior talent leaders March 3-22, 2026, weighted to represent employers with 5+ staff that hire at the entry level. The headline: AI is a net positive for entry-level hiring so far. In 2025, 46% of employers that have at least explored AI report it increased entry-level hiring vs 13% reporting a decrease (nearly 4-to-1); for 2026, 2.7x as many expect AI to raise hiring as to cut it (46% positive vs 17% negative). Greater AI use is the most frequently cited single positive driver (27% of firms naming a significant positive factor). But the bar is rising: 42% say AI increased analytical/judgment tasks for entry-level staff while 41% say routine admin tasks shrank, and the minority reducing headcount concentrate cuts in administrative/data-entry (46%), customer support (44%), and data analytics (41%) roles. 92% of employers are engaging with AI in some way (22% strategically integrated); only 8% have no plans. Notably, employers rank AI literacy the least important entry-level skill — behind critical thinking, communication, and collaboration — and value relevant work experience over a 4.0 GPA with no work history.

The first global task-level automation atlas: 18,797 O*NET tasks scored across 124 countries, producing 2.33M task-country labels. Core insight: automation risk is not fixed at the task level — the same task carries different exposure depending on local wages, technology adoption rates, workforce skills, and production environment. Covers nations representing 99%+ of global GDP and population. Provides the cross-country comparative baseline that US-centric indices (Eloundou, Felten, Tomei) cannot offer, and directly challenges the assumption that AI displaces uniformly across geographies.

Ng argues the 'AI jobpocalypse' narrative is overblown, citing U.S. unemployment at 4.3% and strong software engineer hiring despite coding agents. He identifies three drivers of the narrative: frontier AI labs incentivized to overstate capability, SaaS companies anchoring pricing to employee salaries rather than software benchmarks, and businesses attributing pandemic-era overhiring corrections to AI. Predicts an 'AI jobapalooza' — net job creation through AI engineering roles and transformed non-AI work. Complements the Ezra Klein (NYT) and Yale Budget Lab pieces from the same week with an industry-insider perspective on why incumbents amplify displacement fears.

Brookings FCAI briefing synthesizing the productivity–diffusion–distribution trilemma. Frontier capability is racing ahead (Stanford AI Index: training compute doubling every 5 months; private industry produced ~90% of notable 2024 models), but real-world productivity is lagging. The headline finding: an NBER Feb 2026 survey of ~6,000 executives (US/UK/DE/AU) reports ~70% of firms 'actively use AI,' yet executives spend only ~1.5 hrs/wk on it and ~90% of firms report no impact on employment or productivity over the past three years. Micro-evidence remains bimodal: a customer-support GenAI study showed +15% productivity (concentrated in novices), while METR's randomized trial found experienced open-source developers using early-2025 AI tools were 19% SLOWER on their own repos. Adoption stats: US BTOS Feb 2026 shows 17.5% of US businesses used AI in at least one function in the last two weeks; Eurostat 2025 shows 19.95% of EU firms with 10+ employees. Distributional risk concentrates at entry-level (Stanford Digital Economy Lab); ~88% of language-tagged models on Hugging Face are English-only, widening Global North/South divides. Task-based macro estimates put AI's TFP contribution at <0.66% over 10 years.

The strongest null-result paper to date. Using synthetic differences-in-differences to compare AI-exposed (top tercile) vs. a synthetic-control group built from unexposed occupations, the authors find no statistically significant AI effect on employment shares or real hourly wages through 2026Q1. Unemployment rose ~0.5pp in the latest quarter for the AI-exposed group (more for 16–34 year olds) but remains statistically insignificant. Honest about the limits: LLMs are still improving, exposure metrics may misclassify, CPS is underpowered for the 22–27 cohort. Required reading for anyone calibrating confidence about what the data already shows.

Blair and Guo introduce the concept of 'algorithmic credentialism' — AI-powered hiring screens trained on historical data encode bachelor's-degree requirements as skill proxies, potentially violating civil rights law under disparate-impact doctrine. The framework matters as AI screening proliferates: it implies entry-level access could narrow not because AI replaces workers but because algorithmic filters silently re-impose credential bias that human screens were being pushed to drop.

Not an AI paper — but a critical baseline for thinking about customer-service automation. Dube finds California's AB 1228 (April 2024) raised fast-food wages ~7% with a tight employment own-wage elasticity bracket (−0.29 to +0.26, median −0.02 across 32 specs) despite the floor reaching ~77% of state median hourly wages. The result: even an aggressive wage floor in a sector facing kiosk and ordering automation produced near-zero employment effects through 2025Q3. Useful prior for separating regulation-driven from AI-driven wage and employment changes.

Wolfe and Dangol analyze public communications from five leading AI companies and argue that the demand for high-skill data annotation has created an 'expert gig economy' — commodifying professional expertise into scalable, lower-paid task work. The piece is qualitative, but it names a mechanism that quantitative studies miss: AI labs increasingly source expert judgment via gig platforms, which compresses the professional wage premium even when overall white-collar employment looks stable.

a16z's bullish counter to the displacement narrative: historical precedent (agriculture, electrification, spreadsheets) shows productivity gains expand labor markets rather than shrink them. Current data backs this — 90%+ of firms report no AI employment impact, software dev jobs are rising, and augmentation out-mentions substitution 8:1 on earnings calls. The aggregate picture is neutral-to-positive, with task reallocation underway but no economy-wide job destruction.

Noah Smith traces the AI industry's pivot from 'AI will make humans obsolete' to 'AI will augment workers and expand jobs' — driven by deteriorating public opinion, political risk, and competitive positioning (OpenAI vs Anthropic). The new pitch rests on task creation and Jevons Paradox (induced demand), plus the 'relational sector' thesis (humans paid just to be human). Smith endorses it as better PR and possibly self-fulfilling for AI research direction, though notes the pivot is partly strategic.

Klein argues mass unemployment is unlikely because the macrodata isn't matching the anecdata — unemployment was 4.3% in March 2026 vs 4.4% in March 2020, average hourly earnings are stable, and demand for software engineers is booming despite Claude Code. Drawing on Alex Imas's 'what becomes scarce' framework, he predicts labor will shift toward the 'relational sector' as wealthier consumers pay premiums for human-made goods and services. The harder scenario, he warns, is partial displacement of ~8M workers — the U.S. responds poorly to localized shocks (cf. the China shock's 2M jobs).

A reported essay from inside the SF AI bubble: the 'San Francisco consensus' is that the median worker is screwed, and even doomers, accelerationists, and labs differ mostly on what to do about it. Amodei's '50% of entry-level white-collar by 2030' anchors the discourse; Block CEO Jack Dorsey laid off ~half his staff in March citing coding agents; OpenAI's GDPVal benchmark went from sub-human to >80% win rate vs human professionals in months; Anthropic enterprise-agent revenue jumped from $9B to $30B annualized. OpenAI's new white paper floats a 32-hour week, public wealth fund, and capital-gains hikes; Anthropic has set up an institute but not yet endorsed specific policy. David Shor finds 72% of voters fear AI will drive down wages — the rare populist message that polls across the political spectrum.

Pethokoukis interviews Wharton's Daniel Rock on AI and work. Key framing: exposure ≠ automation. Rock covers the productivity J-curve (slow early gains as firms reorganize workflows), adoption bottlenecks, and a measured growth outlook that pushes back on Silicon Valley claims of imminent white-collar doom.

LMICs employ <10% of workers in skilled knowledge work vs. 41% in high-income countries, limiting the 'grafting' strategy rich countries use for AI adoption. But cheap intelligence could leapfrog: full automation may benefit LMICs differently than augmentation strategies. Key open question: will AI augment scarce knowledge workers or automate knowledge work entirely? LMICs face less political resistance and fewer legacy institutions — a different optimization landscape.

NBER WP 34986 (forthcoming Annual Review of Economics): 18% of US workers hold jobs introduced since 1970. New work commands a wage premium 4× larger for tech-linked roles. Advanced-degree workers are 2.9pp more likely to land new work. Public policy can drive new work creation. New work is the central mechanism counteracting automation-driven displacement.

Nature surveys the existential-risk debate: only 3% of ~4,000 AI researchers name extinction as their top worry (UCL preprint), yet 53% give it ≥10% probability — up from 47% in 2023 (AI Impacts). Dario Amodei puts P(doom) at 25%. Critics including Gary Marcus and Casey Mock argue doom narratives distract from documented current harms and give firms a regulatory shield.

A supply-side rebuttal to Amodei's claim that AI will eliminate half of entry-level white-collar jobs in 1-5 years. Labour markets price jobs, not tasks: when components of a bundle are expensive to separate from the rest, AI helps with parts while humans keep the work. Frey/Osborne in 2013 put 94% automation probability on accountants; a decade later BLS counts 1.6M accountants and auditors at $81,680 median pay and projects +5% growth through 2034, while bookkeeping clerks (a 'weak bundle') are projected -6%. Travel agent employment is 60% below its dot-com peak, yet surviving agents' weekly earnings rose from 87% to 99% of the private-sector average (2000-2025). Also: organizations need residual decision rights — a human who can be sued, fired, and held accountable — that AI agents don't yet have.

Survey of 80,508 Claude.ai users connects qualitative worker sentiment to Anthropic's Economic Index usage data. One fifth voiced concern about AI-driven displacement, and worry tracks exposure: every 10pp of observed exposure adds 1.3pp of perceived threat, and top-quartile exposure workers mention it three times as often as the bottom quartile. Early-career respondents are much more concerned than seniors, and only 60% of early-career users said they personally benefited from AI versus 80% of senior professionals. Mean productivity rating was 5.1/7 ('substantially more productive'); 48% cite scope (new tasks), 40% speed. Management (mostly entrepreneurs) and computer/math show the biggest gains; lawyers and scientists the mildest. Speedup and threat form a U-shape: the workers AI slowed and the workers it sped up most are both more anxious.

Analysis of 4.5M+ federal civil cases and 46M PACER docket entries shows self-represented (pro se) filings broke a 20-year steady state of ~11% to reach 16.8% in FY2025, with pro se case counts nearly doubling from a pre-AI average of 23,210 to 41,490. The rise is concentrated in 'simple' case types (civil rights, consumer credit, foreclosure) and essentially absent in patent or securities fraud. An AI-text detector applied to 1,600 random complaints finds AI-generated text rising from 1.0% (2023) to 18.0% (early 2026). Case durations and disposition mix are unchanged, but docket entries per court from pro se cases are up 158% vs pre-AI — the supply of judicial capacity is fixed while demand has surged.

Comprehensive UK survey of 412 occupations (24.8M workers): no detectable displacement signal three years post-ChatGPT, though occupation-level divergence is sharp — IT analysts up 38%, call centre workers down 19% since 2021. Software sector is the strongest early signal: employment fell 4.5% in H2 2025 as AI coding tools arrived, productivity growth accelerated from 0.8%/yr to 3.8%/yr post-ChatGPT, and expert exposure models systematically overstate actual adoption.

US labor productivity surged from 1.3% annually (2013–2019) to 2.2% (2019–2025), concentrated in three tech-exposed sectors posting 3.2–3.9% growth while the rest of the economy grew at 0.1%, driven by AI deployment and digital transformation that is hollowing out entry-level white-collar jobs even as output accelerates.

AI exposure has already reduced wages 4.5% at substitutable firms and threatens 6.1M U.S. clerical workers, demanding a four-pillar policy response: strengthening social safety nets and worker protections, regulating AI surveillance and algorithmic wage-setting, investing in competitive public AI infrastructure, and banning AI-enabled price discrimination while shifting data-center electricity costs to firms.

OpenAI's framework categorizes 147.9M US jobs into four AI-transition archetypes, finding 18% face near-term automation risk, yet a 66.2 percentage-point gap exists between theoretical AI exposure (90%) and realized exposure (23.8%) in high-risk jobs. Counterintuitively, unemployment rose more in the "less immediate change" archetype than in high-automation-risk jobs since 2024Q1, suggesting AI exposure alone doesn't drive displacement.

No substantial labor market impact from AI detected as of March 2026, with all key metrics remaining flat or on pre-existing trends; only notable finding is widening dissimilarity between older and younger college graduates' AI exposure.

Only 39% of employed Americans use AI at work, with stark disparities by education (58.7% college vs. 22.9% non-college) and income ($15.9% under $50K vs. 66.3% over $200K); among users, 66% report productivity gains, yet a critical training gap exists—38% of workers value AI training but only 15.9% receive it from employers, with untrained workers willing to pay 11.4% of salary for access.

AI will shift jobs toward relational sectors (care, hospitality, craft, education) as consumer preferences favor human-made goods, exemplified by human art commanding a 44% exclusivity premium over AI-generated art and Starbucks reversing store automation.

GDP growth is increasingly decoupled from job creation as AI-driven productivity gains expand output without hiring, a pattern potentially benign if productivity gains translate to higher wages rather than concentrated wealth.

Employment among young software developers has plummeted nearly 20% since 2024 as AI disruption moves from prediction to reality, while generative AI adoption has reached 53% globally in just three years. US AI researcher inflow has collapsed 89% since 2017, and the country's competitive lead over China has nearly disappeared.

OpenAI proposes a comprehensive policy framework for the AI transition including a public wealth fund, 32-hour workweeks, portable benefits, and tax shifts from payroll to capital—warning that productivity gains won't automatically benefit workers without deliberate intervention. The blueprint also calls for AI safety auditing, international coordination, and worker voice in deployment decisions.

AI applications concentrate deeply in narrow task categories, with 92% of AI apps targeting just 6.8% of work activities—primarily information creation and software tasks—while physical and interactive work remain largely unaffected despite 6x growth in AI tools from 2022-2024.

Economists have reversed their dismissal of AI's labor threat, with BCG estimating over 50% of US jobs will be reshaped within 2-3 years as advanced reasoning models begin displacing workers across entry-level and professional roles.

A survey of 159 experts and 401 public respondents projects AI will boost GDP by 0.5 percentage points above baseline by 2030, but a 14% probability rapid-deployment scenario shows GDP reaching ~4% alongside 10 million job losses and wealth concentration to the top 10% by 2050. Expert disagreement on AI's economic impact stems primarily from differing beliefs about economic effects rather than AI capability timelines, with 71.8% of economists favoring job retraining over job guarantees.

A rapid AI scenario generates only +1.5 percentage points additional annual GDP growth versus a slow scenario (+45% cumulative over 25 years), with economists significantly more pessimistic than AI experts, citing social backlash and historical diffusion constraints as key limiting factors.

15.6 million non-degree workers face high AI exposure, with 3.5 million lacking adaptive capacity to transition to better jobs. Nearly half of career pathways from entry-level Gateway occupations to higher-wage positions are highly AI-exposed, concentrated in Sun Belt and Northeast metros.

AI's comparative advantage constraint means cheaper compute won't eliminate human jobs—opportunity costs keep human labor valuable even if AI outperforms humans at every task. However, wages will decline at full automation, with major risks from energy scarcity, wealth concentration, and transition disruptions.

The American AI Jobs Risk Index identifies 9.3M US jobs (6% of workforce) as vulnerable to AI displacement, with information, finance, and professional services sectors facing the steepest risks and innovation hubs like San Jose experiencing 9.9% job loss potential. Writers, programmers, and web designers are most at risk, with $757B in annual income threatened across the economy.

Nearly half of jobs now use Claude for at least a quarter of their tasks, with high-tenure users achieving 10% higher success rates, while broader adoption has shifted task composition toward lower-wage work ($49.3 to $47.9/hr) and concentrated early success among high-skill users.

When tasks are quality complements, automating one task increases returns to automating others, creating bundles where partial automation can raise worker incomes by forcing focus on remaining tasks. Standard displacement measures fail because they ignore task complementarities and bottleneck structures.

AI displacement risk depends on job structure: high-dimensional jobs (consulting, medicine) see wage gains from partial automation due to productivity focus effects, while low-dimensional jobs (trucking, warehousing) face genuine displacement because firms have stronger incentives to fully automate when few complementary tasks remain.

Despite 250 years of predictions that technological advancement would cause mass unemployment, those forecasts proved largely wrong—though the study acknowledges real costs to displaced workers. Annual working hours fell from 2,950 in 1870 to 1,500 in 1998, demonstrating that technology's long-run economic benefits outweigh short-term disruption.

January-February 2026 CPS data shows no measurable AI impact on labor markets, with all exposure metrics, occupational shifts, and employment changes remaining within historical ranges.

Victorian industrial novels offer insight into technological disruption: handloom weavers experienced 50% real wage declines (1806-1820), and the resulting labor unrest and social upheaval parallels dynamics we may face during AI's transition, though potentially at a faster scale.

AI will automate cognitive labor within months, with a Stanford CTO reporting 4-week engineering tasks completed in 45 minutes, followed by physical labor automation in 5-10 years as accelerated R&D advances robotics. The resulting displacement may cause deeper identity crises for knowledge workers than economic harm, creating a bifurcated economy of zero-cost cognitive goods and scarce physical/experiential goods.

Anthropic surveyed 81,000 Claude users about their AI usage, aspirations, and concerns in the largest qualitative study of its kind, completed in one week.

CFOs surveyed expect minimal aggregate job loss (<0.4%) but report productivity gains three times higher than their workforce changes imply, with finance roles seeing >2% productivity growth while routine clerical roles decline 0.76 percentage points annually. 85.4% of firms plan to invest in AI in 2026.

Swedish employer data show that employment of 22-25-year-olds in high-AI occupations declined 5.5% relative to low-AI occupations by 2025H1, while workers 50+ saw 1.3% gains, with effects twice as large for young women. This replicates prior US findings of AI disproportionately affecting young workers' hiring prospects.

6.1 million clerical and administrative workers—86% women—face the highest AI threat due to low adaptability, while white-collar jobs are positioned to be disrupted first despite no measurable job displacement occurring yet.

Explores how AI automation extends beyond job displacement to challenge deeper questions about work's role in identity, purpose, and social meaning.

Dutch youth employment in GenAI-vulnerable occupations declined 13% from Q4 2022 to Q3 2025, while employment in other sectors grew 3%, with job postings in vulnerable occupations dropping 25%.

Individual AI productivity gains of 10x are not translating to firm-level value, mirroring how electrified textile mills saw no output gains for 30 years until organizational redesign. The research proposes an 'Institutional Intelligence' framework with 7 pillars to bridge this gap.

ATMs didn't reduce bank teller jobs due to complementarity, but iPhones caused a 51% collapse in teller employment (332K to 164K, 2010-2022). Paradigm replacement displaces jobs while task automation within existing systems does not.

Evidence on AI labor impact is inconclusive; disruption pace matches prior tech transitions. Under 1/5 of firms using AI per Census BTOS.

New 'observed exposure' metric combining LLM capability with real usage. No systematic unemployment rise, but young worker hiring slowing in exposed occupations.

Eight-month study of 200 employees found 83% said AI increased their workload through greater pace, scope, and hours -- leading to burnout and cognitive fatigue.

First canonical aggregator of research on how AI is impacting jobs at the occupational level, with implications and actions for job seekers.

Millions of women in clerical and customer service roles face AI exposure, echoing the pattern of manufacturing's toll on men during deindustrialization.

Employment declined 1% since late 2022 in the 10% of sectors most exposed to AI. Wages in AI-exposed sectors rising faster, suggesting augmentation and displacement happening simultaneously.

No meaningful relationship between productivity and AI adoption economy-wide, but companies that quantified task-level AI productivity saw a median 30% gain.

Overall employment growing robustly, but 22-25 year olds' employment stagnant since late 2022. AI-exposed jobs driving the decline for young workers.

Two-thirds of 3,500 euro-area firms use AI. Companies with frequent AI use are 4% more likely to hire additional staff. Only 15% cite reducing labour costs as a factor.

Little evidence of AI disruption in labor data. Forward-looking labor market components have improved. AI data center construction driving a pick-up in construction hiring.

Survey of ~6,000 executives: 90%+ report no employment impact from AI over past 3 years. Firms expect only -0.7% employment reduction over next 3 years (US: -1.2%).

No evidence that AI reduces employment in the short run. AI augments worker output -- enabling faster task completion and better decisions -- without displacing labour.

Comprehensive review: 35.9% of US workers used generative AI by December 2025. Adoption accelerating but unevenly distributed across industries and skill levels.

Of 37.1 million workers in the top quartile of AI exposure, 26.5 million are in occupations with above-median adaptive capacity. Not all exposed workers are equally vulnerable.

~10% of US work affected short-term. AI will boost productivity and GDP by 1.5% by 2035. Jobs AI can completely replace saw 0.75% employment fall 2021-2024.

AI can already replace 11.7% of the US workforce ($1.2T in wages). Tech layoffs represent just 2.2% of total wage exposure -- the visible tip of the iceberg.

AI will increase TFP by only 0.53-0.66% over 10 years. Much less transformative than the hype suggests, because AI can only automate a fraction of tasks in most occupations.

37.4% of US workers used AI on the job in the past week. Adoption highest among college-educated (52%) and workers earning $100K+ (58%). Survey of 25,000 workers.

Formal economic models show that advanced AI could theoretically cause negative economic growth if labor income collapses and capital owners stop consuming, though this requires extreme conditions. Most economists consider this unlikely but expect demand-side constraints to moderate AI's growth impact.