Software job ads rebound unevenly

The software total-postings index rose from 62.3 in March 2025 to 77.1 in September 2026. Against the trough, that is a large percentage increase; against the February 2022 peak of 229.2, it recovers 8.9% of the peak-to-trough index-point loss. The peak does not establish what a normal posting level should be. The 23.8% rebound does not mean nearly a quarter of the missing jobs returned, because this is an advertising index rather than a count of jobs or hires. The low was selected retrospectively from this history, not established as a permanent turning point.

Timing also frustrates a simple AI account. By November 2022, software total postings had already fallen 35.8% from their February monthly peak. They then fell another 47.6% through September 2026. The correction was already substantial by the November comparison month. The later interval could contain an AI contribution alongside other changes. Starting a chart in November would hide the earlier recruitment contraction.

The latest direction depends on the measure. Total software postings were 19.1% higher in September than a year earlier. New postings, defined by Indeed as adverts on the site for seven days or fewer, averaged 6.5% lower in July through September than in April through June. Both can occur without telling us how many vacancies were filled. Changes in advert duration, the mix of postings and platform practices remain possible explanations for the split, rather than findings established by it.

The total/new distinction also changes the comparison with the whole platform. Since November 2022, software total postings fell 47.6% against 30.1% for all sectors. For new adverts, the corresponding declines were 43.6% and 40.1%. Software is weaker in both comparisons, but the extra decline is much smaller for new postings. An account of an isolated software collapse misses the wider recruitment cooling.

Across all 47 current Indeed job categories, 46 had lower total-postings indices in September 2026 than in November 2022. The same count holds for new postings. There is considerable variation within that decline: only 19 categories' total indices were below their February 2020 baseline, compared with 24 for new adverts. Losing part of the pandemic expansion is common; falling below the pre-pandemic level is less universal.

Software's total index was the sixth-lowest of the 47 against the native 2020 baseline, while its new-postings index was eleventh-lowest. Data & Analytics and IT Infrastructure, Operations & Support were among the categories with total indices below software. Changing the comparison to November 2022 puts software fifth-lowest for total postings and thirteenth-lowest for new ones.

These are Indeed's normalized-title categories, not official occupational groups, and a category index does not reveal the number of workers or adverts behind it. The national index is a separately published series; adding category indices would not reconstruct it.

Computer systems design and related services payroll fell from 2,482,900 jobs in November 2022 to a preliminary 2,355,400 in September 2026, a decline of 127,500 jobs, or 5.1%. The latest level remains 7.3% above the industry's 2019 monthly average. Total nonfarm payroll grew 3.1% over the same November-to-September comparison.

The payroll population differs from the job board's software category. BLS groups jobs by the employer's industry, including nontechnical employees within computer systems design and excluding developers employed by banks, retailers and other industries. Payroll counts jobs rather than unique people. The modest industry decline and deep advertising decline therefore cannot yield an advert-to-job conversion ratio.

The broader BLS Information industry provides an actual hiring measure, although it excludes computer systems design and is different from Indeed's title categories. Its monthly hires rate fell from 3.5% in November 2022 to a preliminary 1.6% in August 2026. Averaging the latest twelve months gives 2.64%, still below the 2019 monthly-rate average of 3.31%. National hiring was also subdued: August's total-nonfarm hires rate was 3.3%. Retaining an existing job creates no opening for an applicant, so weak hiring can restrict both entrants and experienced engineers seeking a move.

Information's August layoffs and discharges rate was 1.0%, below its 2019 monthly average of 1.35%. But the June-through-August average was 1.47%, and the latest twelve-month average was 1.60%, both above that baseline. These are arithmetic averages of published monthly rates, not BLS's official weighted annual rates. National layoffs remained below their own 2019 average across all three windows; the Information result cannot be generalized to the entire economy.

Layoffs are also only part of the outflow. The August BLS separation tables report 44,000 Information hires and 68,000 total separations, comprising 30,000 quits, 28,000 layoffs or discharges and 10,000 other separations. Subtracting layoffs alone would omit 40,000 exits. These rounded survey flows are not an exact reconciliation with payroll, but they prevent treating few dismissals as evidence that a workforce cannot shrink.

In the New York Fed's latest annual outcomes by major, covering 2024, unemployment was 7.0% for computer science and 7.8% for computer engineering, against 4.2% overall. Among employed graduates, 19.1% of computer science graduates and 15.8% of computer engineering graduates worked in occupations classified as not requiring a college degree, compared with 39.4% overall.

These rates cover ages 22 to 27 with a bachelor's degree or higher, excluding current students. The occupation-based measure is called underemployment; it does not show whether the work matches the person's major. The same table reports early-career median wages of $87,000 for computer science and $90,000 for computer engineering, versus $58,000 overall, but those medians cover full-time workers with a bachelor's degree only. They cannot describe the earnings of the unemployed or guarantee what a new applicant will earn.

The more recent monthly graduate series ends in June 2026 and covers graduates across majors, not a refreshed computing-major cohort. Recent-graduate unemployment was 5.7%, compared with 2.9% for all college graduates, a broader group that includes the recent graduates. The producer uses seasonally adjusted three-month moving averages and defines recent graduates by age and education, not an observed graduation date.

Recent-graduate underemployment was 42.0% in June. That was 3.0 percentage points above November 2022, but only 0.7 points above June 2019.

The education pipeline was expanding while recruitment contracted. Aggregating NCES's original completions files, narrow computer-science bachelor's awards rose from 30,145 in 2018-19 to a provisional 51,304 in 2024-25, an increase of 70.2%. The broader computer and information sciences family rose from 88,638 to 138,367, up 56.1%, while all bachelor's awards increased just 0.3%. Narrow computer science is included in that broader family; the counts must not be added.

These are seven actual academic-year observations, with the latest covering July 2024 through June 2025. They count awards, including those to nonresident recipients, rather than unique graduates or first-time US jobseekers. Narrow program codes also differ from the New York Fed's major categories. Dividing degrees by Indeed adverts would manufacture a supply/demand ratio from incompatible populations.

More potential applicants could face the same number of positions even if employers' willingness to hire held steady. Here, degree growth coexists with weaker advertised demand. Growth in computing awards was already underway before late 2022, which also complicates assigning the whole adjustment to the AI timeline.

Within Stanford and ADP Research's software-developer payroll sample, employment among ages 22 to 25 went from an index of 100 in November 2022 to 80.52 in August 2026. That is a 19.48% decline from the group's own baseline, calculated from the authors' public index. Ages 26 to 30 declined 6.1%, while all older bands grew. Growth was not monotonically greater with age: ages 41 to 49 rose 19.5%, compared with 7.9% for ages 50 and older.

The source follows occupation-matched workers in firms with five years of ADP payroll history, excluding firm entry, exit and provider switches. These selected firms do not represent the national software workforce. Workers also age into and out of the bands, current age-cell counts and confidence intervals are unavailable, and age does not establish seniority. The pre-November trajectory is retained in the chart rather than making every group appear to start from the same employment history.

The Stanford paper's current 19% headline is a different statistic. It compares employment for ages 22 to 25 in highly AI-exposed occupations with where it would have stood had it kept pace with less-exposed peers through June 2026. It is a broader occupational comparison and a relative shortfall, not the software chart's own-baseline decline. The researchers report that adjustment appears to operate primarily through weaker hiring of young workers, with worse patterns where AI usage is more oriented toward automating tasks.

The August revision also makes attribution less secure. Education controls shrink the gaps, some differential trends precede widespread generative AI use, and adjusted estimates are sensitive to pipeline and specification changes. The overall national benchmark comparison shows smaller gaps than the ADP analysis sample, while white-collar sector comparisons agree more closely. A credible task-substitution mechanism does not determine how much of the observed age pattern it caused.

In a hypothetical engineering team, an experienced engineer with a coding agent takes over documented implementation work that would previously have supported a beginner's position. Management cancels that hire. The product continues to need its established engineers, while the plan loses a place where someone could acquire production judgment. Cheaper implementation could also allow more projects, so this is a possible staffing decision rather than a forecast of the eventual number of jobs.

The September Census working paper Graduating into Disruption follows actual bachelor's graduates from participating institutions, covering about 29% of US bachelor's degrees from 2016 through 2024, with employment records through 2025's third quarter. Its authors report a regression-adjusted five-percentage-point decline in initial employment for the most AI-exposed decile of majors relative to the bottom six deciles, alongside a 13% decline in initial full-quarter earnings.

Initial employment here means earnings covered by unemployment insurance in the quarter after graduation. A separate ACS check suggests changes in self-employment and graduate schooling could account for up to half of that covered-employment gap, although those choices can themselves respond to weak demand. It is not a decomposition of the same linked graduates; the populations differ. The earnings outcome is conditional on a full quarter in a dominant job, two quarters after graduation. The employment result is not a CPS unemployment rate.

In the study's partner sample, growth in exposed-major graduate supply was absorbed through 2022. Growth then slowed, yet employed-graduate counts fell. That challenges a supply-only account within that cohort, whose coverage and terminal year differ from the national IPEDS totals. Participating institutions and majors remain selective populations, and concurrent field-specific shocks limit the AI attribution.

In the second quarter of 2026, Yale Budget Lab's published unemployment difference for more-exposed occupations was about +0.125 percentage points, with the authors' 95% interval running from -1.098 to +1.348 points. That interval includes zero and considerable positive or negative differences. It cannot support a claim that the effect is precisely zero.

Unemployment intervals were wholly above zero in the third quarter of 2023 and second quarter of 2025. These are quarter-specific intervals rather than a simultaneous band for the whole path, so selecting those quarters would not settle attribution either. The latest quarterly endpoint is June 2026's quarter, despite the September tracker release incorporating August data in its separate monthly measures.

The method constructs a weighted comparison resembling exposed occupations' earlier trajectory. Exposed occupations receive equal weights, so this is not the effect on the average exposed worker. Gains and losses across occupations can cancel, and unemployment classified by a previous occupation misses people who have never held one. An all-age occupational comparison can consequently coexist with a difficult first transition into computing work. Its interval describes uncertainty within the model, not protection against an inaccurate exposure measure or a competing shock affecting the same occupations.

In The Broken Ladder, Peter John Lambert and Yannick Schindler compare AI and remote-work exposure in a joint specification. For their state-by-detailed-occupation monthly design, the generative-AI exposure coefficient on the junior share of new hires changes from -1.34 percentage points per exposure standard deviation when estimated alone to -0.41 when work-from-home exposure is included. The work-from-home coefficient changes from -1.69 to -1.42. The firm-level design shows a similar pattern.

These June 2026 author estimates compare 2017 through 2019 with 2023 through 2025, pooling selected résumé-derived hiring records across the US, UK, Canada and Australia. Junior status comes from role categories, not ages. The table gives the authors' coefficients and standard errors. Those errors are not confidence intervals, and coefficient changes do not identify the share of the decline caused by either exposure.

Stanford's ADP reanalysis moves in the opposite direction. Its 2025 AI-exposure coefficient changes from -1.06 when estimated alone to -1.52 with remote-work exposure, with standard errors of 0.30 and 0.38. These are percentage points in an age-based junior share of new hires per exposure standard deviation, relative to 2019. The age proxy and annual comparison differ from LSE's role-based seniority and pooled-period design; they are not interchangeable estimates.

The Census graduate study also reports that remote-work controls did not erase its main pattern. Its major-level correlation between AI and remote-work exposure was 0.88, making the two difficult to separate even with controls. Robustness to a control and uncertainty about identification belong together here, rather than treating either as a verdict on the competing explanation.

The proposed organizational mechanism is the cost of supervising and developing beginners. A firm expecting slower learning or heavier demands on experienced colleagues may prefer someone already trained. That makes mentoring capacity relevant to recruitment without proving that remote work caused the decline or that office mandates would reverse it.

In the New York Fed's August 2026 regional surveys, 15% of AI-using service respondents said they hired fewer staff because of AI over the preceding six months, compared with 4% reporting layoffs. But 13% reported hiring more staff, and 34% retrained staff. Manufacturers reported 6% hiring fewer and 22% retraining, with published zeros for layoffs and additional hiring.

These are shares of AI-using respondents, not shares of workers. The source does not report which responses overlapped, and subtracting the hire-more share from the hire-fewer share would not produce net job losses. A reported zero is rounded survey data. Manufacturing covers New York State; services also include northern New Jersey and Fairfield County, Connecticut. The answers do not identify junior or engineering roles. Hiring fewer is a respondent's comparison with hiring without AI, not necessarily with last year's hires. Self-attribution and regional coverage limit causal and national claims.

In the Census Business Trends and Outlook Survey's October 8 release, 23.8% of employer businesses in scope reported current AI use, against 50.2% in Information and 42.7% in professional, scientific and technical services. These refer to September 7 through 20, 2026, and include the survey's US and Puerto Rico coverage. They are firm shares in broad sectors, not adoption rates for software workers. The definition includes machine learning and voice recognition as well as generative AI. Census restarted the AI series after broadening its question in late 2025, so an apparent jump across that change cannot be read as adoption growth alone.

A team can observe increased software output and still have more than one reasonable staffing response. It might complete the existing backlog with fewer planned hires, fund work that was previously uneconomical or redirect engineers toward maintenance and review. The recruitment rebound and the regional survey's opposing hiring responses are compatible with that variety.

For a proposed internal review, compare completed hires with headcount over the same months, splitting the hires by the experience the role required. Retain cancelled requisitions and record whether management removed the work, assigned it to existing employees or changed the requirements. Compare live adverts with newly approved roles and actual fills, so a lingering listing cannot stand in for recruitment. Review a quarter or a year alongside the latest month, following the window sensitivity visible in the layoffs data.

An AI replacement claim needs the additional link between tool use, transferred work and the cancelled position. In the hypothetical team, that would mean showing the implementation work moved to the experienced engineer, accounting for the review it required and checking whether the extra capacity funded new projects. Keep the supervision plan visible as well: a role requiring immediate independence offers a different entry path from one with time allocated for learning.

The posting rebound does not reveal whether newly advertised roles include time for beginners to learn. A team whose remaining roles all require production experience relies on other employers continuing to fund the positions in which beginners gain it. If it removes its own training position to capture an output gain, that reliance belongs in the staffing decision.