Are entry-level jobs disappearing?

What does the unemployment comparison show?

Some exposed fields show weaker entry hiring. The survey measures how unemployment moved across exposure groups. The groups can also differ in industry mix and earlier hiring trends.

The Current Population Survey (CPS) is the household survey behind U.S. unemployment statistics. I compare civilians ages 22–27 and 35–54 who have a bachelor’s degree or higher and are out of school.¹³

An unemployed person is jobless, available for work and has actively searched during the prior four weeks, or is on temporary layoff awaiting recall, under the survey’s definitions. The unemployment rate divides unemployed people by the labor force: employed plus unemployed. People outside the labor force are not in that denominator. Occupation is the current or previous reported occupation; a new job seeker without one cannot be assigned an exposure score.

I freeze the occupation rankings before examining the later outcomes. Low and high groups come from employment-weighted thirds of 2022 occupations, with tied scores kept together. Three published measures test different ideas:

Sources: author task scores, Anthropic February 2025 release, BLS categories.

The scores classify 2022 occupations retrospectively. I subtract the low-exposure group’s change in unemployment from the high-exposure group’s change. A positive difference means unemployment rose more in the high-exposure group.

Under the task-capability measure, unemployment rises from 2.26% to 4.46% in the high group and 5.27% to 6.64% in the low group. The difference in those changes is +0.84 percentage points. The other measures give −1.29 and −0.01 percentage points. All refer to January–October 2022 versus January–October 2026.¹³

How wide are the unemployment intervals?

The middle column gives the measured change. The outside columns give the endpoints of its approximate 95% interval.

Units: percentage points. All estimates concern the same 2022–2026 windows and young-graduate population.

These intervals are wide enough to include meaningful deterioration as well as relative improvement. Repeated observations of the same household are clustered when calculating uncertainty; public files lack what is needed to reproduce the official complex survey variance, so the intervals are approximate. Changing the exposure measure changes the sign, so no single column is the answer.

Have young workers lost employment or pay?

Employment and earnings changes lean negative, and their intervals include zero.

I measure employment in each exposure group per 100 graduates in the whole age group. The measure captures nonemployment and movement between occupations. The high-exposure group’s change relative to the low-exposure group ranged from −2.25 to −0.50 employed people per 100 graduates.¹³

The real earnings comparison ranges from −$2.27 to −$1.64 per hour, expressed in October 2026 purchasing power. Among eligible wage-and-salary workers with valid reports, it uses a weighted mean of weekly earnings divided by usual main-job hours, after applying a common $2,884.61 nominal weekly ceiling to handle changing public-file earnings treatment. It is a capped hourly-equivalent proxy. Changes in who remains employed, reported hours, earnings allocation and the ceiling can change the result.¹⁴

The underemployment proxy uses BLS entry-education requirements to flag jobs that typically require less than a bachelor’s degree. Its relative changes are −1.02, −2.15 and −1.61 percentage points across the three exposure measures. Each interval includes zero. I classify occupations with consistent mapped requirements. The New York Fed uses a different O*NET measure.¹⁵

Does the result change with the age range?

All the unemployment intervals still include zero when the age range or comparison window changes.

2022–2026 changes. Except the last row, each uses January–October.

Subtracting the corresponding change for older graduates produces age-interaction estimates of +0.50, −1.61 and −0.39 percentage points. Those intervals also cross zero. Restricting the sample to citizens at birth does not identify the effect of visa policy; it only tests whether that sample restriction changes this result.¹³

When did software job postings begin to fall?

Software postings had already fallen by ChatGPT’s public launch, so the downturn began earlier.

Indeed’s seasonally adjusted software-development postings index fell 39.1% from its February 28, 2022 peak to November 30, 2022, the day ChatGPT launched. On November 6, 2026, it was 18.6% below its February 2020 benchmark and 23.6% above its year-earlier level.¹⁶

Indeed counts advertisements across experience levels. In September 2026, BLS counted 1.310 million professional and business services openings. Across the whole economy, nonfarm payrolls rose by a preliminary 165,000 in October, and the unemployment rate was 4.1%.^{17, 18}

What did Stanford and the Cleveland Fed find?

The studies use different workers and measures. Exposure scores help identify where to look for changes.

Stanford’s revised August 12, 2026 payroll study reports a 19% descriptive relative employment shortfall for workers ages 22–25, comparing its two highest exposure groups with the three lower groups through June 2026. It calls its findings “early, descriptive indicators.” Its payroll sample covers full-time workers, without restricting them to degree holders; this report’s CPS sample also includes part-time workers. Employer coverage, education and earlier trends matter.¹⁹

In the study’s public Figure 3 data through September 2026, the youngest most-exposed quintile reached 87.66 and the least-exposed quintile reached 108.40, with November 2022 equal to 100. That is a 19.13% relative gap. For ages 35–40, the corresponding figures are 109.00 and 114.30, a 4.64% gap. This calculation compares the outer quintiles in September. The revised paper compares two higher groups with three lower groups through June.¹⁹

The Cleveland Fed’s November 9, 2026 commentary also finds worse unemployment outcomes in more-exposed occupations, with a stronger young-worker gradient and weaker job finding. It uses ages 23–54, without restricting the sample to college graduates: unemployment compares May 2025–April 2026 with January 2015–December 2019; job-finding flows compare January 2024–April 2026 with that earlier baseline. Its scores and windows differ from this report’s. The appendix recomputes its public binned chart.²⁰

The comparisons depend on exposure scores, the prior occupations job seekers can report and the hiring trends already underway.

Is computer science still worth it?

Computer science combines higher earnings among full-time graduates with higher unemployment. The New York Fed’s 2024 major table puts computer science unemployment at 7.0%, versus 4.2% overall. Among full-time early-career workers with a bachelor’s degree only, median annual earnings were $87,000 for computer science and $58,000 overall. New York Fed major table.

The broader graduate market is also difficult. The Fed’s November 2 release puts recent-graduate unemployment at 5.2% in September 2026—5.227% before rounding—and underemployment at 42.5%, or 42.455% before rounding. Here, underemployment means an employed graduate works in an occupation that typically does not require a bachelor’s degree. It is a share of employed graduates, so it cannot be added to the unemployment rate. Sources: unemployment CSV, underemployment CSV, definitions and release date.

The corresponding unemployment average for January 1990–September 2026 is 4.546%, rounded to 4.5%. That dated benchmark includes the Fed’s estimated October 2025 observation; our raw CPS series leaves that uncollected month missing. The monthly series covers non-enrolled people ages 22–27 with a bachelor’s degree or higher and already includes seasonal adjustment and a three-month moving average.

Engineering outcomes vary substantially by field. Computer engineering’s unemployment estimate is higher than computer science’s in the latest table, while civil and electrical engineering are lower.

Sources: original NYFed files for 2021, 2022, 2023, and 2024. Rates are rounded to one decimal.

The same table puts employed CS graduates’ underemployment below the overall rate. That combination—higher unemployment but lower occupational underemployment among those employed—shows why one statistic cannot summarize a major’s prospects. The wage columns cover full-time workers ages 22–27 with a bachelor’s degree only; the unemployment columns also include advanced-degree holders. Source: NYFed definitions.

The table supplies no sample counts or standard errors. Its Overall row includes the named majors; it is not a calculated “all other majors” group. Without population weights, averaging or subtracting the major rates would produce the wrong comparison.

What do studies of recent graduates find?

One study measures unemployment among young degree holders. Another follows graduating cohorts into covered jobs and compares early earnings.

Robert Fairlie and Jane Wu study non-enrolled people ages 22–25 whose highest degree is a bachelor’s. Their unadjusted June–August unemployment rate averaged 7.3% in 2026. It was 7.1% in 2022, 6.3% in 2023, 7.8% in 2024 and 7.2% in 2025. The authors found no statistically significant summer increase in 2026. September 2026 paper, pages 16–19 and Tables 1–3.

Their expanded measure, which includes people outside the labor force who want a job, rises from 9.4% in summer 2025 to 10.4% in summer 2026. This report’s main occupational comparison uses the composite labor-force weight, a broader 22–27, bachelor’s-or-higher population, and January–October windows, so it asks a different question. Source: Fairlie–Wu definitions and summer results, printed pages 8–9 and 16.

Cody Orr, Lee Tucker, and Lawrence Warren’s September 2026 Census working paper finds a deterioration in initial employment and pay for highly exposed majors. Its adjusted estimates compare the most-exposed tenth of majors with the bottom six tenths: employment in the quarter after graduation is five percentage points lower, and dominant-job full-quarter earnings two quarters after graduation are about 13% lower. The dominant job has the highest earnings among qualifying full-quarter jobs, identified using earnings in that job in adjacent quarters; it is not necessarily a graduate’s first job. The analysis links actual degree completions with employer records and covers roughly 29% of U.S. bachelor’s degrees awarded in 2016–2024. Source: CES 26-56, abstract and printed pages 3–5 and 10–11.

That design follows graduates even when they enter a different occupation or do not appear in covered employment, which an unemployment comparison based on a current or previous occupation can miss, especially for first-job seekers with no occupation to report. Conversely, absence from an employer record is not the survey definition of unemployment. The Census study’s detailed records are restricted: its published tables are auditable, but its adjusted model cannot be rerun from public files. Its findings inform how to read our survey results but are not the same estimate.

What do UT Austin first-job outcomes show?

UT’s public reports document successful placements for particular colleges, but their different cohorts and denominators do not form a comparable university-wide employment series. The most complete report I found is McCombs' BBA Class of 2025 report.

It covers 1,264 graduates from summer 2024, fall 2024, and spring 2025: 925 seeking employment, 299 not seeking, and 40 with no information. Its 792 accepted offers imply 85.6% of job seekers, rounded to the report’s 86% headline, within six months of graduation. The denominator matters: 86% is not the share of the entire graduating class. Among 621 salary reports, mean annual base salary was $89,377 and median salary $85,000. Source: full McCombs report, introduction and printed pages 2–4.

Cockrell’s salary page reports $88,141 for bachelor’s graduates, last updated May 22, 2025, without a cohort or respondent count. The CNS computer-science page combines 2014–2023 survey respondents, making its 81% employed figure unsuitable as a current AI-period outcome. Sources: Cockrell, CNS computer science.

The iSchool’s 2025 headline percentages do not reconcile with its detailed outcome counts. These gaps prevent a defensible comparison across colleges or an inference about AI’s effect on UT graduates. Source: iSchool Class of 2025 outcomes.

How many jobs have been lost to AI?

Employers named AI as the reason for 121,401 announced cuts in January–October 2026. That was 19.7% of the 617,189 total announced cuts Challenger tracked.²

Challenger counted 15,341 AI-attributed announcements in March, or 25.3% of that month’s 60,620 announced cuts. October had 1,265, or 2.9% of 43,994. A large announcement can drive the monthly share.^{1, 2}

Challenger counts positions employers announce they plan to eliminate, including some early retirements and voluntary departures. The series gives the employer’s stated reason for a planned cut. It lacks worker age, degree and career stage.

What do layoff notices show?

State notices record planned separations at workplaces and sometimes list job titles.

The study reviewed 69 selected notices from Texas, Washington and California. Texas’s accessible official export lists four Meta Austin notices dated November 16, 2022, with separations planned for January 13, 2023. The counts of 154, 7, 11 and 50 sum to 222.³

Washington’s Microsoft letter dated September 22, 2026 schedules 277 positions for elimination on November 21, 2026, after this report’s cutoff. It cites reorganization and restructuring and says some positions may move to other Microsoft locations.⁴ Reading the main letters for 14 notices found no explicit statement that AI caused those separations. That finding applies only to the reviewed letters. Texas coverage ends with notices dated June 23, 2026 in the accessible export; later Texas filings remain a source gap. Notices are not summed into a national total because their coverage differs and amendments can overlap.

What do companies disclose about AI and staffing?

Some filings explicitly connect workforce changes to AI strategy. They rarely quantify how many jobs AI alone displaced.

Accenture’s fiscal 2025 filing described “exiting people in a compressed timeline where reskilling is not a viable path” alongside efficiencies “including through AI.”⁵ Its fiscal 2026 filing reports $923 million of costs for the completed program: $628 million of employee severance and $295 million primarily for asset impairments. The program began in fiscal 2025's fourth quarter and ended in fiscal 2026's first quarter. Accenture also reported approximately 814,000 people on August 31, 2026, against 779,000 one year earlier.^{5, 6} Hiring, acquisitions and departures can occur together.

The study’s headcount comparison covers 262 annual observations for 50 selected SEC filers from 2021. Each company reports its own employee or full-time-equivalent measure, usually worldwide. Acquisitions and disposals change the group counted. Dell’s January 2021 figure, for example, included VMware employees.

Could something other than AI explain weaker hiring?

The decline began before several later events, which helps test competing explanations.

Interest rates and the hiring boom. The Federal Reserve raised its target range to 0.25%–0.50% on March 16, 2022, and to 4.25%–4.50% on December 14, 2022.⁷ Higher financing costs offer an ordinary reason to reduce expansion. Meta’s November 9, 2022 explanation also pointed to pandemic growth expectations that failed to persist, weaker revenue and a hiring freeze.⁸ Estimating AI’s effect would require accounting for the earlier reversal in hiring.

Research tax treatment. Starting with tax years after December 31, 2021, section 174 required domestic research costs, including software development, to be deducted over five years, with 15 years for foreign research. Legislation enacted July 4, 2025 restored immediate domestic deductions for tax years beginning after December 31, 2024.⁹ Continued weakness after the policy changed would weaken a tax-only explanation.

Offshoring and visas. BEA’s latest outward multinational employment data cover 2023, and its next release is scheduled for November 20, 2026.¹⁰ In Global Nurse Force, the district court’s September 30, 2026 order vacated and enjoined the challenged agency policies implementing the H-1B payment, including the 2026 extension.¹¹

Energy. EIA dates the sharp 2026 oil-market disruption to military action on February 28, 2026 and interrupted Strait of Hormuz traffic.¹² It can be a candidate explanation for later weakness; it cannot explain why software hiring first turned down in 2022. None of these records supplies the missing occupation-by-age causal comparison.

Which claims survive the evidence check?

The strongest evidence describes a difficult and uneven entry market. The claim that AI caused a quantified nationwide loss of entry-level jobs goes beyond these public records.

What would change the answer?

A stable exposure-related deterioration across public datasets, with credible earlier comparisons and narrower intervals, would strengthen the case for an AI effect. Evidence that the weakness follows non-AI demand, financing or hiring cycles after comparable controls would weaken it.

A stronger test would follow graduation cohorts into employment, include people with no prior occupation, measure hires and departures, and date an employer’s AI adoption before any workforce change. Pay comparisons would benefit from uncapped earnings with consistent hours and worker composition.

Use the occupation lookup to compare employment, unemployment and pay for ages 22–27 and 35–54 by occupation, age or measure. The intervals show the uncertainty around each estimate.

Frequently asked questions

Is AI taking entry-level jobs?

Some exposed fields show weaker entry employment. In the CPS comparison of January–October 2022 and 2026, relative unemployment changes range from −1.29 to +0.84 percentage points across three exposure measures. All three approximate 95% intervals include zero.

What is the unemployment rate for recent college graduates?

The New York Fed’s latest published rate is 5.2% for September 2026, released November 2. It covers non-enrolled people ages 22–27 with a bachelor’s degree or higher and is already seasonally adjusted and smoothed over three months.

How do computer science graduates fare in the job market?

The New York Fed’s 2024 major table puts computer science unemployment at 7.0%, versus 4.2% overall. Full-time early-career computer science workers with a bachelor’s degree only earned a median $87,000 a year, compared with $58,000 overall.

How many layoffs were blamed on AI in 2026?

Challenger reports 121,401 announced cuts with AI as the stated reason in January–October 2026. It counts employer announcements and does not track which workers were ultimately laid off.

Did software postings fall before ChatGPT?

Indeed’s software postings index fell 39.1% from its February 28, 2022 peak to November 30, 2022, the public launch date.

Data, methods and refresh dates

The CPS workbook uses native Census occupation codes common to 2022 and 2026. The occupation lookup CSV uses broader occupational categories and covers both age groups from 2015 to 2026.

The lookup pools January–October each year. In 2025 it uses January–September because the CPS skipped October. Estimates below the display thresholds are blank. The workbook also tests the comparison using January–September in both endpoint years.

The next routine check is May 12, 2027, followed by November 12, 2027. New BLS and New York Fed releases or a material study revision can change the figures sooner. The BEA multinational release scheduled for November 20, 2026 will update the offshoring figures.

Who keeps the value of legal AI? looks at how task-level time savings affect staffing and revenue. ElevenLabs: What a finished minute is worth compares time savings and cash savings.

Appendix: Cleveland Fed public chart replication

The released Figure 1 has twenty binned points under each exposure measure. Recomputing ordinary least squares—the line that minimizes squared vertical distances—gives fitted highest-minus-lowest gaps of 1.285, 1.339 and 0.906 percentage points for human-rated task exposure, model-rated exposure and observed usage. These figures describe the Cleveland Fed sample. Valid confidence bands would require additional survey inputs.

Primary sources for the original model

1. Challenger March 2026 report. PDF page 7, Table 4. ↩
2. Challenger October 2026 report. Table 4 on PDF page 10 gives reasons for cuts. Table 1 on page 6 gives totals. See also the release. ↩
3. Texas WARN notice export, records 1543–1546 for Meta Austin, dated November 16, 2022. The live export can change. ↩
4. Washington ESD Microsoft notice, September 22, 2026, PDF page 1. ↩
5. Accenture FY2025 Form 10-K. See Item 1, Our People, and Note 1, Business Optimization. ↩
6. Accenture FY2026 Form 10-K. Item 1 gives workforce totals. Note 16, footnote 3, gives costs spanning Q4 FY2025 and Q1 FY2026. ↩
7. March 16, 2022 FOMC statement and December 14, 2022 statement. See the decision paragraphs. Implementation began the following day. ↩
8. Meta’s November 9, 2022 statement, sections on the causes and the hiring freeze. ↩
9. Historical section 174, subsections (a)(2)(B) and (c)(3). Current section 174A, subsections (a) and (d)(3) and section 70302 effective-date notes. ↩
10. BEA 2023 release and multinational release schedule. The figures cover U.S. parent companies and their foreign affiliates. ↩
11. Global Nurse Force v. Trump, district-court order, case 4:25-cv-08454-HSG, ECF 130, September 30, 2026, pages 2 and 33–34. The order addresses Proclamations 10973 and 11069. ↩
12. EIA April 7, 2026 analysis, introduction and crude oil prices section. ↩
13. U.S. Census Bureau CPS Basic monthly public-use files, January 2015–October 2026. The CPS workbook recalculates the comparisons in CPS Calc, Cov Calc and Contrasts. Coverage by group and period. ↩
14. Census earnings rounding and topcoding and BLS public-use guidance. The unadjusted CPI-U series gives the October 2026 price index of 337.511. ↩
15. CPS occupation mapping uses BLS employment projections and Census crosswalks. Occupations with mixed education requirements remain unclassified. ↩
16. Indeed software-development postings via FRED, November 10, 2026 vintage. The latest observation is November 6. The full series is hosted at the source, where seasonally adjusted figures can be revised. ↩
17. BLS September 2026 JOLTS release, released November 3, 2026, professional and business services job-opening level; series JTS540099JOL. Workbook converts published thousands to jobs. ↩
18. BLS October 2026 Employment Situation, November 6, 2026. Table B-1 gives the preliminary nonfarm payroll change. Table A-1 gives the unemployment rate. ↩
19. Stanford Digital Economy Lab’s August 12, 2026 revised study covers payroll through June. The public Figure 3 data in the October 16 vintage cover through September. Exact object URL in the source register. ↩
20. Cleveland Fed, Job Finding and Unemployment in the Age of AI, November 9, 2026. See Figures 1–4 and the workbook linked there. ↩