Why Hasn’t AI Increased Unemployment?  🔗 **Source:** [Why Hasn’t AI Increased Unemployment?](https://twitter.com/PeterMcCrory/status/2079979321607745905/?rw_tt_thread=True) > [!note]+ Document Notes > --- > tags: [ai-labor-market, labor-augmentation, occupational-change, productivity, human-expertise] > author: "Peter McCrory" > source: "https://x.com/p_mccrory/status/2014281227663524038" > reference: "01ky8rcpvrc9kds4q1n04emz23" > date created: 2026-07-24 > --- > # 🏷️ Title: Why hasn’t AI increased unemployment? > ## 🧭 Table of Contents > Summary; main content and notes; references; paraphrase; newsletter; action items; hypotheses; methodology; questions; future research; implications; learning, accessibility, DEIB, logic, DARVO, conclusion; refactor appendix; template library. > ## 🔗 References > - Source: https://x.com/p_mccrory/status/2014281227663524038 > - Reader document: 01ky8rcpvrc9kds4q1n04emz23 > ## 📝 Summary > The essay argues that recent US labor data does not yet show a material AI-driven rise in unemployment. AI remains skill-biased and labor-augmenting because jobs contain mixed tasks, model capability is uneven, and expert oversight still determines outcomes. The evidence supports caution about present claims, not certainty about the future. > ## 📖 Main Content > The author cites stable unemployment, roughly balanced openings, growing but incomplete firm adoption, modest productivity changes, limited occupational churn, and no occupation whose full task bundle is handled by the cited model. Software examples show experts planning, reviewing, and recovering from failures. Weak links constrain end-to-end gains even when code output rises. > ### 🗒️ Notes > - Claim: Task automation does not imply full job automation because jobs contain linked responsibilities. > - Mechanism: Uneven model capability raises the value of expert oversight and recovery. > - Verification focus: Weak links cap system-level productivity even when one production metric jumps. > - Scope: This analysis covers the captured source body. It does not treat linked claims as independently verified unless the source supplies direct evidence. > ## 📚 References > - Primary source: https://x.com/p_mccrory/status/2014281227663524038 > - Follow linked primary documents before using numerical, legal, medical, or policy claims in publication. > ## 🗣️ Paraphrase > AI can perform parts of jobs while leaving planning, context, coordination, and recovery to people. Jobs then change through task rebundling. Aggregate data has not yet separated a large AI unemployment effect from macro conditions, though weaker hiring among some young exposed workers deserves study. > ## 📰 Newsletter > **Why this matters for WordPress:** The essay argues that recent US labor data does not yet show a material AI-driven rise in unemployment. AI remains skill-biased and labor-augmenting because jobs contain mixed tasks, model capability is uneven, and expert oversight still determines outcomes. The evidence supports caution about present claims, not certainty about the future. For teams building open web software, the useful move is to convert the source’s central claim into a small test, publish limits, and keep the people affected by the decision in the review loop. [Read the source](https://x.com/p_mccrory/status/2014281227663524038). > ## ✅ Action Items > - DevRel: turn claims into testable demos; OSPO: track governance and dependency effects; Community: invite affected groups into review; WordPress — Core: assess platform impact; Core Performance: measure cost and latency; Core AI: define model-specific tests; Core Program Team: coordinate milestones; Mobile: test small-screen use; Playground: build a reproducible demo; Community: gather contributor feedback; Training: create a lesson; Marketing: state limits; Support: collect failure patterns; Design: review workflows; Documentation: record behavior and caveats; Photos: not applicable unless visual assets are needed; TV: consider a short explainer; Polyglots: plan translation context; Accessibility: test WCAG 2.2 AA; Meta: review site and API effects; CLI: expose repeatable checks where relevant; Hosting: test deployment constraints; Tide: add automated quality checks where applicable; Openverse: review open-asset implications; Themes: test presentation effects; Plugins: publish compatibility guidance. The WPOCC: consider community-governance effects. FAIR: note implications for federated open-source collaboration. > ## 🧪 Hypotheses > 1. If task automation does not imply full job automation because jobs contain linked responsibilities., teams will observe a measurable change in decision quality or failure rate. > 2. If teams test the mechanism in a bounded pilot, results will differ by task complexity and user expertise. > 3. Publishing limits beside outcomes will reduce overgeneralization in later decisions. > ## 📊 Methodology > Use a preregistered comparison with a baseline, clear unit of analysis, task-level outcomes, failure categories, cost and time measures, and interviews with affected users. Preserve raw inputs and report null results. For a short announcement, first obtain primary documentation and reproduce the stated behavior. > ## ❓ Questions > - Which claim would change the decision if it proved false? > - What is the correct unit: task, job, team, organization, or ecosystem? > - Which groups bear costs that the source does not measure? > - What evidence would distinguish adoption from dependable outcomes? > ## 🔮 Future Research > Replicate across organizations and user groups; compare short- and long-term effects; measure distributional impact; test alternative explanations; and track whether early gains survive operational use. > ## 💡 Implications > Practice: use bounded trials. Policy: require traceable evidence for high-impact claims. Security: inspect failure paths and dependencies. Education: teach evaluation, not tool loyalty. Funding: tie continued spending to measured outcomes and public limits. > ## 🧠 Learning Styles Assessment > - Visual: add a claim-to-evidence diagram. > - Aural: discuss the argument and counterargument aloud. > - Verbal: rewrite the thesis in plain language. > - Kinesthetic: run the proposed pilot. > - Logical: map premises, mechanism, evidence, and conclusion. > - Social: compare interpretations in a small group. > - Solitary: keep a prediction and error log. > - Naturalistic: examine resource and environmental effects. > - Reading level: approximately grade 10–12; technical or institutional terms need definitions. > ## 📊 Accessibility Assessment > - Perceivable: provide alt text, captions, sufficient contrast, and text equivalents (WCAG 1). > - Operable: support keyboard access, visible focus, and adequate target sizes (WCAG 2). > - Understandable: use clear headings, definitions, and consistent navigation (WCAG 3). > - Compatible: use semantic markup, valid names, roles, and values (WCAG 4). > - Cognitive: shorten dense passages and expose the argument structure. > - Motion and media: avoid forced motion; provide controls and transcripts. > - Testing: combine automated checks with keyboard, screen-reader, zoom, reflow, and cognitive review against WCAG 2.2 AA. > ## 🌍 DEIB Assessment > - Cultural breadth: identify the source’s geographic and institutional scope. > - Inclusive language: separate groups from stereotypes and avoid universal claims. > - Neurodiversity support: provide summaries, headings, examples, and predictable structure. > - Representation: seek evidence from people affected by the proposed change. > - Psychological safety: invite dissent without requiring public disclosure of harm. > ## ✅ Logical Validity > The essay uses several independent indicators and states uncertainty. Limits include US scope, short observation windows, internal-company evidence, exposure measures that may not match real adoption, and confounders such as interest rates and remote-work shifts. > - Strong reasoning: the source offers a clear thesis and at least one mechanism. > - Fallacy check: watch for hasty generalization, selection bias, false cause, motive attribution, and analogy used as proof. > - Fix: add comparison data, definitions, uncertainty ranges, and counterevidence. > - Tone: explanatory and persuasive. > ## 🛡️ DARVO Analysis > No complete DARVO sequence is established by the captured source. Where conflict appears, separate denial, criticism, and claims of harm rather than assigning a pattern from rhetoric al  I thought I’d share a few high-level reflections and a framework that helps me make sense of why we (so far) don’t see significant impact of AI on the US labor market. I focus on the US because (a) AI adoption is high here, (b) it’s the world’s largest economy, (c) [compared to Europe](https://www.ecb.europa.eu/pub/pdf/sintra/ecb.forumcentbankpub2025_Schoefer_paper.en.pdf), our labor market institutions historically facilitate faster adoption of labor-displacing technologies, and (d) we have an unusually rich array of labor market statistics for understanding current conditions. I shared these thoughts recently within Anthropic. First, AI adoption is high in the US. So it’s reasonable to look for its effects on productivity, or on unemployment. Labor productivity growth has been strong in recent years, and there’s suggestive evidence that AI has partly contributed. But the US labor market is currently stable and close to maximum employment. In my view, AI has caused no material increase in the unemployment rate to date. Even if we focus on workers with high exposure to current patterns of AI automation, we don’t see unexpected increases in unemployment in recent years (though as I’ll discuss below, these workers are more concerned about job loss). Why don’t we see any impact of AI adoption on unemployment? AI (so far) has the hallmarks of a “skill-biased,” labor-augmenting technology. Even as AI automates some aspects of work, complementary human expertise amplifies what AI (or humans) can achieve alone. AI broadens the scope of what people can accomplish, which increases the returns to working with AI. Now, the future is still quite uncertain. Model capabilities are advancing rapidly, and AI systems may soon be able to [autonomously develop](https://www.anthropic.com/institute/recursive-self-improvement) their own successors. More generally intelligent AI systems could lead to labor displacement that hasn’t yet materialized. Whether that might happen (and how quickly) is an open question, one my team is working to understand. In many ways, this short essay is my attempt to synthesize Anthropic's economic research over the past 18 months to understand how people use AI and what that implies for work, the labor market, and the broader economy right now. This work has emphasized that the future is still uncertain and the implications are likely to be uneven. So far, we've seen muted unemployment effects; this essay presents my framework for understanding why (and what might change in the future). **Important background fact: The US labor market is currently stable.** * The unemployment rate in June was 4.2%, which the Fed [views](https://www.federalreserve.gov/monetarypolicy/files/fomcprojtabl20260617.pdf) as the level consistent with full employment and stable prices. * The ratio of job openings to unemployed workers recently rose to just over 1 in April, which some economists argue [implies](https://www.brookings.edu/wp-content/uploads/2024/09/17160-BPEA-BPEA-FA24_WEB_MichaillatSaez.pdf) that demand for and supply of labor are roughly and efficiently balanced. More sophisticated indicators, like the [NY Fed’s Heise, Pearce, Weber (HPW) Labor Market Tightness Index](https://www.newyorkfed.org/research/labor-market-tightness/), also point to a balanced labor market. * The prime-age employment-to-population ratio remains close to multi-decade highs, which reflects broad-based labor market strength that emerged during the post-pandemic expansion. * Weekly initial claims for unemployment insurance have been stably low over the past four years. And rates of layoffs and discharges remain at or slightly below pre-pandemic levels. **So should we even expect an impact from AI on the labor market yet? I think the answer is yes: The AI sector is large enough that we can look for discernible macroeconomic effects.** * ~20% of firms [use](https://www.census.gov/programs-surveys/btos.html) AI in at least one business function. In the information sector, which is 5.5% of GDP, the share is 40%. * Quality-adjusted AI output grew over 2,000% per year in both 2024 and 2025, according to [estimates](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6813982) by Patrick McKelvey and my colleague @akorinek. Even from a small initial base, this suggests that we should see signs of AI’s impact in the aggregate. * **I believe we’re starting to see AI’s impact *in aggregate productivity statistics*:** * Labor productivity growth has been relatively strong in recent years: the ratio of output per hour of work increased 2.0% per year from 2022Q1 to 2026Q1, as compared to 1.6% in the four years prior to the pandemic. * Sectors of the US economy with higher AI adoption have [tended](https://www.stlouisfed.org/on-the-economy/2026/mar/mind-gap-ai-adoption-europe-us) to have faster labor productivity growth over the past four years. * Total factor productivity growth—how economists like to measure productivity—has been [more modest in recent years](https://www.frbsf.org/research-and-insights/data-and-indicators/total-factor-productivity-tfp/) (especially after accounting for factor utilization). This weakens the evidence, but mixed signals on productivity could themselves indicate that an acceleration is [underway](https://www.frbsf.org/research-and-insights/publications/economic-letter/2026/05/have-we-entered-era-of-high-productivity-growth/). * Firms make hiring decisions based on beliefs about the future value of employing a new worker, relative to the costs. * Firms take into account how productive workers will be in the future, wages over the course of [employment](https://doi.org/10.1016/j.jmoneco.2014.07.007), and the expected duration of the match. * For this reason, firms’ beliefs about the future should shape labor market outcomes *today*, in the form of job separations, hiring, and occupational switching. But across quits, layoffs, and job finding rates, we see a [pattern of stability](https://github.com/qlmonthly/updates/blob/main/QLMonthlyUpdateMay2026.pdf). There’s likewise [no evidence](https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market) of unusually fast occupational churn due to AI. * Software Engineering job posts have [broadly rebounded](https://www.citadelsecurities.com/news-and-insights/global-macro-strategy/the-economics-of-intelligence/) since May 2025, compared to all other job posts, indicating rising demand for workers in an occupation where AI capabilities are advancing the most. **Is there any evidence that job displacement is happening, even if it’s not (yet) macroeconomically consequential? The evidence is mixed. But overall, I’m unconvinced.** * As Maxim Massenkoff and I documented in [our labor impact report,](https://www.anthropic.com/research/labor-market-impacts) we haven’t seen worsening unemployment rates for workers in roles with a large share of tasks that Claude is being used to automate, relative to workers in other roles. Updating this analysis with more recent data from the BLS doesn’t change this result. * We do find *some* suggestive evidence that hiring rates for young workers in highly AI-exposed roles have weakened over the past year or so. That’s consistent with the evidence in *[Canaries in the Coal Mine](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/)*, a paper by researchers at the Stanford Digital Economy Lab. * **But this evidence for young worker displacement should be interpreted with caution.** It’s hard to discern causal effects, because AI emerged in an unusually volatile macroeconomic environment: unwinding of pandemic-era dislocations, rapid tightening of monetary policy, commodity price volatility following Russia’s invasion of Ukraine, and sustained global policy uncertainty (e.g. from trade wars). Because hiring is a form of investment, broad economic uncertainty can itself weigh on hiring. * Another way to put it: from 2022 to now, the US experienced the largest non-recessionary labor market slowdown on record (the “[immaculate disinflation](https://www.nytimes.com/interactive/2023/11/17/business/what-is-immaculate-disinflation.html)”). This coincided with a “low hire, low fire” labor market. This kind of labor market hits early-career entrants hardest. Right now, young workers may be struggling to find jobs for macroeconomic reasons other than AI. * Separately, some recent work [argues](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6787638) that the rise in remote work after the pandemic has had an impact on young workers in roles that are highly exposed to AI. It’s challenging to isolate the impact of AI from other shocks buffeting the economy. * While we don’t see unemployment effects yet, we do find that workers in roles with tasks that Claude is used to automate do [express greater concern](https://www.anthropic.com/research/81k-economics) about losing their jobs than those in less exposed roles. If you believe that the US labor market is currently healthy, that AI could in principle be generating macroeconomically discernible effects, and that this hasn’t yet produced displacement for highly AI-exposed roles, then the next question is obvious: **Why hasn’t AI caused a meaningful increase in unemployment?** **So far, AI is both *skill-biased* and *labor-augmenting*.** It complements domain expertise. It relies on humans in the loop to direct and evaluate the most complex work. And it rewards AI proficiency. Model capabilities are improving fast, but remain [stubbornly jagged](https://helentoner.substack.com/p/taking-jaggedness-seriously). To fill in the pockets of the jagged frontier, expert oversight is needed to steer incredibly capable AI systems, and to recover when they falter. Of course, some jobs are more exposed to outright displacement by automation. For instance, technical writers, data entry workers, customer support representatives, and computer programmers are jobs where AI can reliably handle the core set of tasks and responsibilities. Even though we haven’t seen any increase in unemployment for workers in these sorts of roles, occupations with higher observed exposure are [projected by the BLS](https://www.anthropic.com/research/labor-market-impacts) to grow less through 2034. But *so far*, the broader picture is one of labor *augmentation*. The effects in the labor market are set to be uneven as a result, even as capabilities advance rapidly. This does not mean that all skills that currently command a premium in the labor market will do so in the future. Some types of expertise may become less valuable (e.g. pure coding implementation) even as others become more valuable (e.g. managerial skills of delegation and evaluation). Why do we think AI is a skill-biased, labor-augmenting technology? I’ll make a few observations based on work published by the [Anthropic Economic Research](https://www.anthropic.com/research/team/economic-research) team, and evidence from the [Anthropic Economic Index](https://www.anthropic.com/economic-index): * Despite the incredible advance of AI and rapid adoption throughout the economy, there’s no job in the O*NET taxonomy (a Department of Labor catalog of occupations and their typical tasks) for which* all\* associated tasks are systematically handled by Claude. If jobs are fixed bundles of tasks—they aren’t, but more on that in a moment—then the essential, non-automated aspects of work both constrain the overall productivity lift and amplify the returns to labor. Tasks that Claude can’t handle may depend on interpersonal coordination, in-person interactions, or engagement with the physical world that, so far, only humans can do (see “[The task is not the job](https://www.siliconcontinent.com/p/why-desk-jobs-survive-and-amodei)” by @lugaricano). * In our January report, *[Economic Primitives](https://www.anthropic.com/research/anthropic-economic-index-january-2026-report)*, we find that sophisticated user inputs and complex Claude outputs are highly correlated. In other words, when Claude builds a complex economic model, in practice it does so under the guidance of someone providing complementary, expert direction. * While Claude tends to succeed on average, the model struggles most on the more complex work. This is broadly in line with the [METR task-horizon results](https://metr.org/time-horizons/), in which higher reliability comes at the cost of shorter task horizons. We find suggestive evidence that humans-in-the-loop increase the complexity of work that Claude handles. * In our March report, *[Learning Curves](https://www.anthropic.com/research/economic-index-march-2026-report)*, we found that even after just six months of use, people are more likely to interact with Claude as a thought partner, and have more successful interactions with Claude. If AI was good enough on its own, we wouldn’t expect to see this effect. * Widespread task automation can still augment labor. Why? Because jobs are *not* fixed bundles of tasks. New technologies have historically led to large changes within existing jobs, even as some jobs go away. And they’ve produced entirely new types of work that [combine](https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3) new technical capabilities with complementary human expertise. We see signs of this effect in our research: A commonly cited source of *perceived* productivity among [81k Claude users was scope](https://www.anthropic.com/research/81k-economics)—being able to do more, more proficiently. Such empowerment from AI may redraw the boundaries of our roles, and produce new bundlings of tasks within jobs—automating some, reinforcing the importance of others—while on net increasing the marginal product of labor. * In our June Economic Index report, *[Cadences](https://www.anthropic.com/research/economic-index-june-2026-report)*, over a third of all respondents from the [Anthropic Economic Index Survey](https://www.anthropic.com/research/economic-index-survey-announcement) expect AI to be able to do most or all of the tasks they do in their jobs in the next twelve months. A similar share of respondents expect job responsibilities to change for themselves and their colleagues in the coming year. But expectations of job loss were notably lower. Indeed, those who use Claude in more automated ways tend to be more optimistic about AI’s impact on their pay, job security, and ability to find a job. * As AI capabilities improve, we’ll see increasingly capable agents that can autonomously handle complex, long-horizon, valuable tasks. Will AI still augment labor? To get a handle on this question, we [recently analyzed patterns of Claude Code usage](https://www.anthropic.com/research/claude-code-expertise) to see if agentic coding is altering the returns to expertise. Claude Code has been used on more and more valuable tasks over the seven months we tracked, but we’ve seen persistent returns to human expertise: that is, people make planning decisions, and delegate implementation to Claude. People with more domain expertise succeed in their tasks more often, and recover more consistently when Claude makes an error. The return to straightforward coding ability may have fallen, but agentic coding has so far increased the value of other, complementary skills. The future is hard to parse. It may very well be the case that the skill-biased, labor-augmenting aspect of AI will go away as models continue to improve. The jagged frontier may become smoother. And at a certain point, the returns to human expertise may diminish. **A big reason there's so much uncertainty about the future is that AI may automate innovation itself.** Endowing machines with general cognitive capabilities is a direct catalyst for further innovation, in ways that past general purpose technologies weren’t. The internal combustion engine couldn’t invent new modes of transportation. In otherwise standard economic models, automating innovation—which we might see first in recursive self-improvement—can produce economic singularities: infinite growth in finite time. Will such singularities occur? Not if there are essential tasks that are never automated, whether for technical reasons or societal constraints. Those “weak links” are the limits on growth: “Economic growth may be constrained not by what we do well but rather by what is essential and yet hard to improve,” as [Aghion, Jones, and Jones (2017)](https://www.nber.org/system/files/working_papers/w23928/w23928.pdf) argue. Such weak links can keep the labor share of income elevated in the long-run, even under [very rapid, widespread, but incomplete automation](https://web.stanford.edu/~chadj/JonesTonetti_Automation.pdf). The “weak links” logic holds not just for producing goods, but also for automating innovation. Whether AI dramatically accelerates productivity growth requires not just improving capabilities but [broad-based automation](https://www.kellogg.northwestern.edu/faculty/jones-ben/htm/Artificial_Intelligence_in_Research_and_Development.pdf) of tasks involved in the R&D process. Recent [evidence](https://www.nber.org/papers/w35275) suggests that weak links are currently a constraint on automating software production: a 10-20x increase in lines of code generated after the introduction of coding agents yielded only a 30% increase in software releases, and no total increase in app usage. Scaling laws are hard to argue with. The models are going to get better. Much better. I expect this will drive faster productivity growth, and maybe even more clear signs of RSI. But I don’t expect unemployment to be noticeably higher a year from now—at least not because of AI., Why Hasn’t AI Increased Unemployment?, Peter McCrory, 01ky8rcpvrc9kds4q1n04emz23.