Do U.S. Job Postings Show AI Lowering Demand?

What is the strongest evidence for a real decline?

Yes, U.S. postings show weaker relative demand for some occupations whose tasks are more automatable. The clearest focused evidence comes from a Dallas Fed comparison within industries and among firms that remained in the data. This is a shift in advertised demand, not proof that AI caused every decline, that total hiring fell by the same amount, or that a particular worker will lose a job. “Relative” means one group’s postings weakened compared with a less-exposed group; it does not mean every exposed occupation had fewer openings absolutely. In “Job postings show early signs of AI automation impact,” Dallas Fed researchers link O*NET occupation tasks to a measure based on observed AI-platform use, then examine Lightcast postings after ChatGPT’s late-2022 release. Comparing more- and less-exposed occupations within the same industry, they estimate postings for the more-exposed positions were about 5 percent lower relative to the comparison group by the end of 2023 and about 8 percent lower by the first quarter of 2025, for a 10-percentage-point difference in automatable task share. In a panel of Texas firms present throughout the period, more-exposed firms’ postings for these occupations fell by roughly 5–6 percent by mid-2024 and 8–9 percent by early 2026. The study is observational, uses online ads that underrepresent some work, and cannot isolate concurrent economic changes. Its estimated total-posting reductions of 1.8 percent in Texas in 2024 and 2.6 percent in 2025 are regional estimates, not U.S. totals. The broader picture is not uniform contraction. PwC’s 2026 U.S. “AI Jobs Barometer” analyzes Lightcast postings by occupation exposure quartile. From 2012 to 2025, the lowest-exposure quartile reached about 4.7 times its 2012 posting level, compared with about 1.9 times for the highest-exposure quartile. Yet all four quartiles added postings from 2024 to 2025, and the highest-exposure quartile still had the most absolutely, about 13.7 million. This is a long-run descriptive gradient, not an estimate of AI’s causal effect: exposure groups can differ for many reasons, and the chart does not identify why their paths diverged. A Federal Reserve Board note, “AI Adoption and Firms’ Job-Posting Behavior,” further complicates a claim of broad hiring collapse. Combining Lightcast ads with industry survey measures and a firm measure based on AI or machine-learning skills in earlier ads, it finds no negative relationship with subsequent total posting behavior in its models. The authors describe estimated effects as near-zero or slightly positive, while cautioning against a causal interpretation. Total firm volume can remain steady while the occupational mix changes, so this result does not directly refute Dallas’s narrower finding. Together, the evidence supports a real but uneven and method-sensitive hiring signal for particular task bundles. It cannot show that every exposed role is shrinking. The rest of the analysis asks why credible studies differ and how skill requirements can rise alongside weaker demand for some roles. Postings are recruiting signals, not hires, employment, layoffs, or a personal displacement forecast.

Sources: Job postings show early signs of AI automation impact; 2026 AI Jobs Barometer: United States; AI Adoption and Firms’ Job-Posting Behavior

What does the broad national trend actually say?

A broad U.S. posting trend points to slower long-run growth in occupations placed in higher AI-exposure groups, but it does not establish that AI caused demand to fall. PwC’s 2026 U.S. analysis of Lightcast online ads compares exposure quartiles from a 2012 baseline: by 2025, postings in the lowest-exposure quartile stood at about 4.7 times their 2012 level, while postings in the highest stood at about 1.9 times. That is a difference in cumulative growth, not evidence that high-exposure postings fell or that AI caused employers to reduce hiring. The choice of baseline matters. An earlier PwC comparison from 2019 to 2024 made a shorter-window contrast; the updated 2026 report presents the quartile series beginning in 2012 and extending through 2025. Those windows answer different descriptive questions. A 2012 baseline spans shifts in industries, occupational composition, online recruiting, and business cycles; it cannot isolate the effect of generative AI introduced much later. A post-2019 comparison also includes other demand and recruiting changes. The chart flags a pattern to investigate, but does not identify its cause. The absolute counts also change the headline. In PwC’s comparison of 2024 with 2025, postings increased in every exposure quartile: the lowest group rose from about 4.1 million to 4.6 million, the third from 8.0 million to 9.2 million, the second from 11.6 million to 13.4 million, and the highest from 12.7 million to 13.7 million. The most exposed quartile still had the largest total in 2025. Thus, “grew more slowly since 2012” and “posted more openings in 2025 than in 2024” can both be true. Relative growth compares each group with its own earlier level; absolute volume counts the ads at a given time. Neither measure, on its own, says how many positions were filled or how many workers were employed. This broad comparison is consistent with the Dallas Fed’s more focused finding that postings for relatively more automatable occupations weakened within industries over its later study period. The analyses do different work: PwC describes long-run differences between occupation groups, while Dallas compares occupations within industry and reports a relative decline. Together, they make hiring pressure plausible for some work, but the broad chart does not supply the causal test Dallas attempts. Nor does the Dallas result mean every high-exposure occupation is declining. Exposure quartiles are group labels, not descriptions of every job’s daily task mix. One occupation can include roles with different tools, customers, discretion, review duties, and accountability. A posting count for that occupation also does not reveal which tasks changed or whether an employer adopted AI. PwC’s chart therefore cannot tell a particular worker whether their employer is hiring less, whether their location has fewer openings, or whether a specific deliverable is being automated. It can show where a more careful question is warranted: which roles within a broad group are growing, shrinking, or changing their advertised work? For a worker, the practical reading is narrow. Treat a long-run exposure gradient as a reason to inspect recent postings for the actual role and region, then compare recurring duties and required skills over time. Do not translate the 4.7-to-1.9 ratio into an individual risk score or assume that the highest-exposure group is in absolute decline. Broad charts can point toward a labor-market pattern; current local evidence and task-level changes are needed before choosing an upgrade, adjacent move, or larger career change.

Sources: 2026 AI Jobs Barometer: United States; Job postings show early signs of AI automation impact

How strong is the within-industry Dallas Fed signal?

The Dallas Fed analysis offers stronger evidence of an exposure-linked hiring pattern than a chart that simply compares all postings before and after ChatGPT’s release. It asks whether occupations with more automatable tasks lost ground relative to less-exposed occupations in the same industry, then checks whether the pattern also appears among firms that remained in the data throughout the period. The result is a relative pullback for certain work, but this observational posting study cannot prove that AI alone caused an employer to cancel a vacancy or predict an individual worker’s outcome. Comparing occupations within industries matters because hiring can move differently across sectors . If technology firms slow recruitment while health care firms expand it, an economy-wide comparison could make an occupation look weaker simply because it is concentrated in the contracting sector. The Dallas Fed model compares more- and less-exposed occupations inside the same industry, reducing that source of distortion. It reports that postings for more-exposed positions were about 5 percent lower relative to less-exposed positions by the end of 2023, and about 8 percent lower by the first quarter of 2025. These estimates represent the difference associated with a 10-percentage-point gap in the share of tasks classified as automatable; they do not mean every exposed occupation’s postings fell 8 percent in absolute terms. In “Job postings show early signs of AI automation impact,” the Dallas Fed links occupation tasks from O*NET with a measure built from observed use of Anthropic’s Claude platform. The index is intended to capture tasks generative AI is already being used to perform, rather than only tasks a future system might handle. But it is not an audit of which Texas employers adopted AI, how extensively they deployed it, or whether a tool replaced or assisted workers. An occupation score is a lens for comparing task bundles, not a direct measure of each firm’s implementation. The continuing-firm comparison addresses another possibility: perhaps the occupational mix changed because new firms entered with different staffing patterns, or firms hiring for exposed work disappeared. Researchers follow a balanced panel of Texas firms present in Lightcast throughout the period and classify them using occupations posted before ChatGPT’s release. These continuing firms showed similar declines: postings for the more-exposed group were approximately 5 to 6 percent lower by mid-2024 and 8 to 9 percent lower by early 2026. The paper says these estimates are not statistically distinguishable from the occupation-industry results. That consistency makes a simple “new firms explain it” account less persuasive. It does not show what happened inside each firm or rule out changing budgets or recruiting practices. The Dallas Fed estimates that AI automation exposure reduced total Lightcast postings in Texas by approximately 1.8 percent in 2024 and 2.6 percent in 2025. These modeled Texas aggregates draw on exposure scores, estimated AI use, and the state’s industry composition; they are not U.S. totals. Occupation-industry estimates using U.S. postings were quantitatively and qualitatively similar, supporting the relative pattern beyond Texas without making the Texas aggregate a national estimate. Remaining explanations include post-pandemic hiring normalization, weaker demand for particular products, investment and organizational redesign occurring alongside AI, and changes in how online ads are classified. The within-industry and continuing-firm checks narrow some alternatives but do not isolate every concurrent change or create a randomized comparison. Lightcast records online ads and underrepresents some occupations, including farming and construction. Fewer advertised openings also do not tell us how many realized hires disappeared: jobs may be filled internally, through other channels, or not filled. The defensible conclusion is that the Dallas Fed finds a substantial, method-sensitive signal: within comparable industries, more-exposed task bundles lost relative posting share, and continuing Texas firms showed a similar pattern. This makes an AI contribution to weaker demand plausible for some work, with consequences for people entering those roles or trying to move into them. The data cannot assign the cause of one vacancy or determine an individual outcome. Stronger causal evidence would connect measured adoption to actual hires or employment , alongside recruiting signals.

Sources: Job postings show early signs of AI automation impact

Why do credible U.S. studies reach different answers?

These studies reach different answers because “AI exposure” and “lower demand” do not name one shared measure. A posting decline among occupations whose tasks appear automatable can occur alongside stable or rising total postings at firms using AI, if hiring shifts toward other occupations. The results are therefore in tension, but they are not direct opposites. In “Do Job Postings Show Early Labor Market Effects of AI?”, New York Fed researchers combine an occupational exposure measure with national Lightcast listings and ask whether postings for more-exposed occupations fell disproportionately after ChatGPT’s late-2022 release. Their summary says overall hiring slowed, but the postings offer little indication of a distinct AI-driven decline in labor demand. It supports a cautious null finding for the comparison described, not a claim that no exposed occupation or worker faced weaker prospects. “AI Adoption and Firms’ Job-Posting Behavior” asks a different question. The Board of Governors analysis links Lightcast postings to Census Business Trends and Outlook Survey reports of business AI use, then examines subsequent total postings at industry and firm levels. For September 2023 through November 2025, it finds no evidence that higher measured adoption reduced total posting behavior. Its firm-level adoption proxy is especially narrow: whether a prior posting requested AI or machine-learning expertise. Thus, this result weighs against broad net hiring collapse in the observed firms and industries; it does not rule out a change in which jobs they advertise. The Dallas Fed’s “Job postings show early signs of AI automation impact” measures the outcome at a more specific level. It maps O*NET tasks to an Anthropic task-automation measure, then compares more- and less-exposed occupations within the same industry. The analysis reports relative posting gaps of about 5 percent by late 2023 and 8 percent by the first quarter of 2025 for a 10-percentage-point exposure difference; continuing Texas firms also shifted their posting mix away from more-exposed occupations. The researchers report nationwide occupation-industry robustness checks, while their aggregate Texas estimate remains regional. This design can detect compositional pressure that a firm’s total posting count obscures. It remains observational, not proof that AI caused every difference. The measures must stay separate. Task exposure estimates whether work could be automated under a defined index; reported business use measures adoption more directly; an AI skill in an advertisement captures a stated hiring requirement. None is interchangeable with the others. Nor does a posting count establish filled jobs, employment, layoffs, or a worker’s redundancy risk. Different comparison groups, periods, and outcome levels can produce different findings without one study invalidating another. Taken together, the null findings challenge the claim that U.S. employers have broadly cut total hiring because of AI. They do not erase evidence of weaker relative demand for selected task bundles. The defensible reading is uneven, level-specific pressure: occupational mix may shift even when employer totals do not. Broader claims require evidence that links measured adoption to repeated changes in both job composition and realized employment.

Sources: Do Job Postings Show Early Labor Market Effects of AI?; AI Adoption and Firms’ Job-Posting Behavior; Job postings show early signs of AI automation impact

Why does the answer change with the time window?

The answer changes because “since when?” changes the comparison. Indeed Hiring Lab’s U.S. posting analysis finds that occupations with greater potential exposure to AI generally saw larger declines from May 2022, the labor market’s posting peak, through May 2026. But between May 2025 and May 2026, the relationship flipped: more-exposed occupations generally had larger posting rebounds. These are two descriptions of different windows, not proof that one cancels the other or that AI caused either pattern. The longer window is not a clean before-and-after test of ChatGPT. Indeed reports that vacancies in exposed occupations began declining before ChatGPT’s late-2022 release, and it labels the exposure relationship uncontrolled. That timing matters: a trend already underway cannot be assigned wholly to a tool introduced later. Nor does a statistically significant association establish its cause. Other hiring and sector changes may contribute alongside AI-related task changes. The analysis identifies a pattern, not a measured effect of adoption on each occupation. The newer window supplies a counter-signal, though not a causal estimate. In Indeed’s comparison, the exposure-posting relationship turns positive over the year ending May 2026. Software development illustrates both the rebound and its limits: U.S. postings rose almost 15% from late February 2025 through June 2026, while total postings fell 7% over that same period. Yet software postings remained about 27.5% below February 2020 levels. A rise from a depressed base need not restore the earlier level or mean the occupation has recovered. The composition of the software rebound also matters. Indeed Hiring Lab reports that senior positions accounted for 71% of the increase in software-development postings from May 2025 to May 2026, while jobs mentioning AI in their title accounted for 37%; those groups overlap. That describes advertised openings, not proof that employers prefer every experienced applicant or junior work is disappearing. It does caution against treating a rising occupation-wide line as an equal opportunity signal for every career stage. Other forces were changing too, so the timing alongside agentic coding tools cannot show they created the increase. The Dallas Fed’s study ends at a different point, first-quarter 2025 for its occupation comparison. It reports that, within industries, postings for more-exposed positions were about 8% lower relative to less-exposed positions by then, for a 10-percentage-point difference in automatable task share. That finding and Indeed’s later rebound can both be true: one estimates a relative decline through early 2025; the other describes a reversal during a subsequent year. Later movement does not erase the earlier pattern, and the earlier pattern does not establish what happened after its observation window. For a worker, the useful signal is therefore not one national cumulative curve. Check recent postings in the location and role family you could realistically pursue, across repeated weeks or months, and notice whether the number of openings and their seniority, duties, and stated AI requirements move together. A rebound is not proof of broad recovery or a reliable opening for junior applicants. Treat it as a reason to investigate current local demand while keeping any next step proportionate to your income, time, and other constraints.

Sources: AI and Job Postings: From Destruction to Creation; Job postings show early signs of AI automation impact

Can skill requirements rise while exposed-role demand weakens?

Yes. Posting counts and requested skills measure different employer choices. A company could advertise fewer openings for tasks it can automate while asking applicants in remaining or redesigned roles to use AI-related tools. Manufacturing offers a useful cross-check: task-based measures often classify it as less exposed because production work includes physical tasks. The Federal Reserve Board’s “AI on the Factory Floor: Evidence from Manufacturing Job Postings” analyzes Lightcast online ads and reports that AI-related requirements rose rapidly in late 2025. By July 2026, they appeared in 11 percent of manufacturing postings, against 8 percent across all sectors. Broad AI and machine-learning requirements also rose in production-occupation ads, though they remained lower than in manufacturing overall; generative-AI skills were essentially absent from those production ads through the first half of 2026. The figures describe advertised skill mix, not whether manufacturers added jobs or changed current workers’ tasks.

The distinction matters because the note measures what employers request when recruiting, not what hired workers do. It treats Lightcast ads as a timely hiring-demand indicator, while noting that online postings and employer-reported vacancies are different measures. Ads can miss internal promotions and hiring through other channels; a requested skill also does not verify that a firm has implemented AI. The analysis can show changing advertised requirements, but not adoption prevalence, worker-level task change, or net employment growth.

The International Labour Organization review “AI and jobs. A review of theory, estimates, and evidence” explains why these signals can diverge. Synthesizing experiments and digital-trace studies, it reports substitution between people and machines in some writing and translation work, alongside rising demand for AI skills in some data. It cautions that exposure measures alone do not show adoption or whether use substitutes for workers, augments them, or changes expertise. It supports a set of possible mechanisms, not a verdict for every U.S. employer: some task demand may shrink, other workers may be assisted, and tasks may be recombined into roles with different requirements.

Workers should track two signals separately: whether openings are becoming more or less numerous, and whether their listed duties or skills are changing. Compare recent local ads, then ask whether recurring requirements connect to work you already do. If AI literacy would help you verify outputs in a real workflow, a small project or targeted course may be proportionate. The evidence does not justify a generic expensive credential by itself, or show that learning one tool secures a job. Choose a next step that fits the work you want, feedback, time, and budget; treat repeated local requirements as a reason to investigate, not a promise of demand.

Sources: AI on the Factory Floor: Evidence from Manufacturing Job Postings; AI and jobs. A review of theory, estimates, and evidence

Which alternative explanations and measurement limits still matter?

Postings are useful early indicators, but cannot by themselves show that AI caused a hiring change, establish realized hires or employment, or reveal layoffs and displacement. A shared decline may reflect broad vacancy cooling; a gap is more suggestive but still depends on the comparison. “Canadian employment trends in the era of generative artificial intelligence: Early evidence” examined Canadian occupational exposure groups. Higher-exposure, lower-complementarity vacancies declined at a rate similar to lower-exposure vacancies from late 2022 through the third quarter of 2025. This is a scope check, not a U.S. estimate: it illustrates how shared declines complicate an AI-specific explanation. The authors also identify post-pandemic adjustment and other economic changes as alternatives.

An online advertisement records a stated opening, not whether it was filled or whether work shifted to existing employees. Some jobs are less consistently advertised online; the Dallas Fed notes limited Lightcast coverage for farming, construction, building maintenance, and personal services. Employer use may change without appearing in ads, while platform coverage and coding complicate comparisons. Internal promotion, referrals, and other unadvertised recruitment are missing from posting counts. A slowdown, changed investment plans, or reorganization can reduce ads as AI tools become available. A before-and-after pattern cannot separate these influences by itself.

The International Labour Organization’s “AI and jobs. A review of theory, estimates, and evidence” shows why outcomes differ: task capability or productivity evidence does not establish employer adoption, and neither alone shows whether firms substitute for workers, augment their work, or change total labor demand. The Federal Reserve Board’s “AI Adoption and Firms’ Job-Posting Behavior” makes a related boundary explicit: total firm and industry postings can miss pressure concentrated in particular occupations. Stable totals can coexist with a changed role mix.

A stronger AI-linked demand case requires repeated patterns across independent posting sources, evidence of adoption at affected employers, and follow-through in occupation-level hires and total employment over time. Persistent changes in hires and employment alongside measured adoption would raise confidence. Stable total hiring alongside changes in occupations and advertised tasks would fit redesign or reallocation better than net contraction. Until outcomes converge, the verdict remains revisable: postings identify a signal to investigate, not its full cause or consequence.

Sources: Canadian employment trends in the era of generative artificial intelligence: Early evidence; AI and jobs. A review of theory, estimates, and evidence; Job postings show early signs of AI automation impact; AI Adoption and Firms’ Job-Posting Behavior

What should a worker do with this signal now?

Use national evidence as a prompt for a task and local-market check, not an abrupt career pivot. First, map one recurring deliverable: which steps are routine and which require human judgment or accountability? Second, identify where AI might change the workflow. Verify outputs against work standards and note where human review remains. Third, compare local postings for possible roles. Look for recurring shifts in duties, seniority, and requested AI or review skills, not one-off ads. The Dallas Fed analysis finds relative posting declines in more exposed occupations, while Indeed's 2022–26 series shows a recent reversal. Neither establishes what will happen in your workplace. The Federal Reserve's manufacturing analysis finds rising advertised AI-skill requirements, which do not prove adoption or job growth. Compare staying and redesigning your work, an adjacent move using existing experience, and larger retraining against income, location, time, cost, health, and family needs. Try a reversible workflow or learning change before an expensive credential. If task changes and local hiring signals recur, choose a feasible next step; otherwise, keep monitoring. The free task checker organizes pressure signals. The paid career roadmap compares scenarios and constraints, but neither predicts personal job loss or guarantees employment or income.

Sources: Job postings show early signs of AI automation impact; AI and Job Postings: From Destruction to Creation; AI on the Factory Floor: Evidence from Manufacturing Job Postings