The Early Impacts of AI on Employment among Recent College Graduates

Probably not a substantial claim: may not pass S3 (load-bearing).

CPS microdata have been used extensively in previous studies of unemployment.

“The CPS microdata have been used extensively in previous studies of unemployment”

• cites Hoynes, Miller and Schaller 2012 sourceNot checked — no full text of the source
• cites Couch, Fairlie and Xu 2020 sourceSupported

Classified as probably not a substantial claim.

Remote work exposure is measured using the teleworkable classification of Dingel and Neiman (2020).

“using the teleworkable classification of Dingel and Neiman (2020)”

• cites Dingel and Neiman 2020 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Deterioration in AI-exposed jobs predates the launch of ChatGPT.

“Frank et al. (2026) emphasize deterioration in AI-exposed jobs that predates the launch of ChatGPT”

• cites Frank et al. 2026 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Previous studies document early labor market shifts associated with AI using payroll records, job postings, and unemployment insurance claims.

“Previous studies document early labor market shifts associated with AI using payroll records, job postings, and unemployment insurance claims (Brynjolfsson et al., 2026; Tucker, 2026; Liu and Webber, 2026; Hyman et al., 2026).”

• cites Brynjolfsson et al., 2026Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Tucker, 2026Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Liu and Webber, 2026 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The teleworkability measure of Dingel and Neiman classifies occupations using O*NET task and work-context descriptors and is built on 2010 O*NET-SOC codes.

“Dingel and Neiman classify occupations using O*NET task and work-context descriptors.Their measure is built on 2010 O*NET-SOC codes, which we bridge to 2018 Census occupation categories using the Census occupation code crosswalk.”

• cites Dingel and Neiman, 2020 sourcePartly supported

  The Classification of occupations section says the measure uses O*NET Work Context and Generalized Work Activities questionnaires to classify occupations, supporting the descriptor-based part of the claim. Our copy refers to SOC codes but does not identify them as 2010 O*NET-SOC codes. The bridge to 2018 Census categories is described as the citing authors’ own step, not as something attributed to Dingel and Neiman. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

Among recent college graduates, 45 percent are in occupations with a teleworkable value of 1 and 34 percent with a value of 0.

“Among recent college graduates, 45 percent are in occupations with a teleworkable value of 1, and 34 percent are in occupations with a teleworkable value of 0.”

May combine two claims.

Point estimates on interactions with both theoretical and observed occupational AI-exposure measures are positive but not statistically significant.

“Point estimates on interactions with both theoretical and observed occupational AI-exposure measures are positive but not statistically significant.”

May combine two claims.

As of March 2026, computer and mathematical occupations were assessed as 94 percent theoretically feasible for AI but only 33 percent covered in observed usage.

“As of March 2026, computer and mathematical occupations were assessed as 94 percent theoretically feasible for AI but only 33 percent were actually covered in observed usage”

May combine two claims.

None of the Table 2 summer 2026 interaction estimates is statistically significant, and the point estimates are small.

“None of the coefficient estimates is statistically significant, and the point estimates are small in magnitude.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The expanded unemployment measure follows many of the same movements as the standard measure, though the difference between the two rates varies over time.

“The expanded measure follows many of the same movements as the standard measure of unemployment, although the difference between the two rates also varies over time.”

May combine two claims.

The unemployment rate for recent college graduates increased from 5.4% in May 2026 to 7.8% in June, then declined to 7.4% in July and 6.8% in August.

“The rate increased from 5.4 percent in May 2026 to 7.8 percent in June, before declining slightly to 7.4 percent in July and 6.8 percent in August.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Unemployment rates and other economic outcomes returned to pre-pandemic levels near the beginning of 2022.

“Unemployment rates and other economic outcomes returned to pre-pandemic levels near the beginning of 2022 (Fairlie 2026).”

• cites Fairlie 2026Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Frontier models gained about 30 points on Humanity's Last Exam in a year.

“frontier models gained ~30 points on Humanity's Last Exam in a year”

May combine two claims.

All four Table 3 summer 2026 interaction point estimates are positive but none is statistically significant.

“All four point estimates are positive but none is statistically significant.”

May combine two claims.

The expanded unemployment rate increased from 7.9% in May 2026 to 11.1% in June, then declined to 10.7% in July and 9.4% in August.

“The expanded unemployment rate increased from 7.9 percent in May 2026 to 11.1 percent in June and then declined slightly to 10.7 percent in July and 9.4 percent in August.”

May combine two claims.

Across eight specifications using observed AI exposure and theoretical AI exposure, the point estimates are consistently positive but none are statistically significant.

“Across the eight specifications, the point estimates are consistently positive but none are statistically significant.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The interaction between the young graduate indicator and the summer indicator is large and precisely estimated (0.014 to 0.030), confirming that the summer increase in unemployment is much larger for recent college graduates than for older college graduates in general.

“The interaction between the young graduate indicator and the summer indicator, in contrast, is large and precisely estimated (0.014 to 0.030), confirming that the summer increase in unemployment is much larger for recent college graduates than for older college graduates in general.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In Table 3, the interaction between the college graduate indicator and summer 2026 is 0.002 (0.009) with a linear trend and 0.006 (0.010) in the event-study specification for the standard measure, and 0.013 (0.011) and 0.012 (0.012) for the expanded measure.

“The interaction between the college graduate indicator and summer 2026 is 0.002 (0.009) in the specification with a linear trend and 0.006 (0.010) in the event-study specification for the standard measure, and 0.013 (0.011) and 0.012 (0.012) for the expanded measure.”

May combine two claims.

Neither the standard nor the expanded summer 2026 coefficient is statistically significant.

“Neither coefficient estimate is statistically significant.”

May combine two claims.

The estimated coefficient on the summer 2026 dummy is -0.008 (SE 0.007) for the standard unemployment measure and 0.001 (0.009) for the expanded measure.

“The estimated coefficient on the summer 2026 dummy is -0.008 with a standard error of 0.007 for the standard unemployment measure (Specification 1) and 0.001 with a standard error of 0.009 for the expanded measure that adds those not officially in the labor force but report wanting a job (Specification 3).”

May combine two claims.

Unemployment in summer 2026 remains within the range observed since 2022, and the summer increase follows a pattern also evident in earlier years.

“Examining these levels, unemployment in summer 2026 remains within the range observed since 2022, and the summer increase follows a pattern that is also evident in earlier years.”

May combine two claims.

The summer 2026 unemployment rate is higher than in 2022, 2023 and 2025, but below 2024.

“The unemployment rate in summer 2026 is therefore higher than in 2022, 2023 and 2025, but below 2024.”

May combine two claims.

Unemployment among recent college graduates averaged 5.3% in winter 2026, compared with 5.7% in winter 2025.

“Unemployment among recent college graduates averaged 5.3 percent in winter 2026, compared with 5.7 percent in winter 2025.”

May combine two claims.

Averaging over June, July and August, unemployment was 7.3% in 2026, compared with 7.1% in 2022, 6.3% in 2023, 7.8% in 2024 and 7.2% in 2025.

“Averaging over June, July and August, unemployment was 7.3 percent in 2026, compared with 7.1 percent in 2022, 6.3 percent in 2023, 7.8 percent in 2024 and 7.2 percent in 2025.”

May combine two claims.

As in Table 2, the interactions for earlier summers are close to zero, and the summer increase in unemployment is larger for college graduates than for young non-college workers.

“As in Table 2, the interactions for earlier summers are close to zero, and the summer increase in unemployment is larger for college graduates than for young non-college workers.”

May combine two claims.

In the event-study specification, the interactions for summers 2023 to 2025 are close to zero, so no summer since 2022 stands out.

“The interactions for summers 2023 to 2025 in the event-study specification are also close to zero, so no summer since 2022 stands out.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In Table 2, the interaction between the young college graduate indicator and summer 2026 is -0.007 (0.008) with a linear trend and -0.005 (0.009) in the event-study specification for the standard measure, and 0.002 (0.009) and 0.004 (0.010) for the expanded measure.

“The interaction between the young college graduate indicator and summer 2026 is -0.007 (0.008) in the specification with a linear trend and -0.005 (0.009) in the event-study specification, and 0.002 (0.009) and 0.004 (0.010) respectively for the expanded unemployment measure.”

May combine two claims.

The estimated interaction between summer 2026 and the theoretical AI exposure measure of Eloundou et al. is small and not statistically significant in either the standard or the expanded specification.

“| Summer 2026 × AI exposure | 0.009 (0.008) | 0.012 (0.009) | 0.008 (0.008) | 0.012 (0.009) |”

• Also stated in: § Figures and · plain

  “AI exposure is defined using the measure fromEloundou et al. (2024)which is GPT-4 LLM-coded.”

• cites Eloundou et al. (2024) sourceNot checked — no full text of the source

May combine two claims.

The estimated interaction between young college graduate status and summer 2026 is small and not statistically significant in either the standard or the expanded specification, under both the linear-trend and event-study specifications.

“| Young × Summer 2026 | -0.007 (0.008) | -0.005 (0.009) | 0.002 (0.009) | 0.004 (0.010) |”

May combine two claims.

The standard unemployment rate was also slightly higher for young college graduates than for young non-graduates in summers 2022 and 2024, but the 2026 gap is larger.

“The standard rate was also slightly higher among young college graduates in summers 2022 and 2024, but the 2026 gap is larger.”

May combine two claims.

The point estimate of the interaction with occupational teleworkability is positive but difficult to separate cleanly from AI exposure.

“The point estimate of the interaction with occupational teleworkability is positive but difficult to separate cleanly from AI exposure.”

May combine two claims.

The estimated interaction between summer 2026 and the teleworkability of an occupation is positive and statistically significant for college graduates aged 22-25, in both the standard and expanded specifications.

“| Summer 2026 × Teleworkable | 0.028* (0.015) | 0.040** (0.017) | 0.028* (0.016) | 0.039** (0.018) |”

May combine two claims.

On the expanded measure, the summer 2026 rate rose from 9.4% to 10.4% for college graduates aged 22-25 while falling from 11.6% to 10.7% for non-graduates of the same age, narrowing the college advantage from 2.2 to 0.3 percentage points, the smallest of the five summers.

“The rate among young college graduates increased from 9.4 percent in summer 2025 to 10.4 percent in summer 2026, while the rate among young noncollege graduates declined from 11.6 to 10.7 percent.The gap favoring college graduates therefore narrowed from 2.2 percentage points to only 0.3 percentage points, the smallest of the five summers.”

May combine two claims.

Among college graduates, the summer 2026 unemployment rate was 2.8% for ages 30-49 (up from 2.7% in summer 2025) and 7.3% for ages 22-25 (up from 7.2%), leaving the gap between young and older graduates unchanged at 4.5 percentage points.

“Figure3ashows that unemployment among older graduates is lower and exhibits less pronounced seasonal variation.Their unemployment rate increased from 2.7 percent in summer 2025 to 2.8 percent in summer 2026, while the rate among young graduates increased from 7.2 to 7.3 percent.The young-older gap therefore was stable at 4.5 percentage points.”

May combine two claims.

Relative to older college graduates ages 30-49, the summer 2026 estimates are statistically insignificant, and relative to young workers without a college degree they are also statistically insignificant.

“Relative to older college graduates ages 30-49, the summer 2026 estimates are statistically insignificant, and relative to young workers without a college degree estimates are also statistically insignificant.”

May combine two claims.

Neither the official nor the expanded measure exhibits a statistically significant departure from earlier summers.

“Neither the official nor the expanded measure, however, exhibits a statistically significant departure from earlier summers.”

May combine two claims.

The summer 2026 coefficient itself, which applies to graduates in occupations that cannot be performed from home, is negative in every column, indicating that the differential captures both higher unemployment in teleworkable occupations and lower unemployment in non-teleworkable occupations.

“The summer 2026 coefficient itself, which applies to graduates in occupations that cannot be performed from home, is negative in every column, pointing to the differential capturing not only higher unemployment in teleworkable occupations but also lower unemployment in non-teleworkable occupations.”

May combine two claims.

For the teleworkability measure, there are statistically significant positive point estimates using both the standard and expanded measure, in both the base difference-in-differences and event-study specifications.

“We find statistically significant, positive point estimates using both the standard and expanded measure, in both the base difference-in-differences and event-study specifications.”

May combine two claims.

Both the standard and the expanded unemployment measures for recent college graduates vary substantially within each year, following a clear seasonal pattern in which unemployment rises sharply in the summer months.

“Both measures show substantial variation over each year, with unemployment following a clear seasonal pattern across all years, in which unemployment among recent college graduates rises sharply during the summer months and then”

May combine two claims.

On the expanded measure that adds the sidelined unemployed, the summer 2026 rate rose from 9.4% to 10.4% for college graduates aged 22-25 and from 4.1% to 4.2% for those aged 30-49, widening the young-older gap from 5.3 to 6.2 percentage points.

“Under the expanded definition, the rate among young college graduates increased from 9.4 to 10.4 percent, compared with an increase from 4.1 to 4.2 percent among older college graduates.The expanded gap widened from 5.3 percentage points in summer 2025 to 6.2 percentage points in summer 2026.”

May combine two claims.

The coefficient estimates of interactions between summer 2026 and both measures of occupational AI exposure are positive but not statistically significant.

“The coefficient estimates of interactions between summer 2026 and both measures of occupational AI exposure are positive, but not statistically significant.”

May combine two claims.

The estimated interaction between college graduate status and summer 2026 is small and not statistically significant in either the standard or the expanded specification, under both the linear-trend and event-study specifications.

“| College × Summer 2026 | 0.002 (0.009) | 0.006 (0.010) | 0.013 (0.011) | 0.012 (0.012) |”

Probably not a substantial claim: may not pass S4 (based).

Many recent college graduates who are unemployed or out of the labor force but want a job do not report an occupation because of their lack of previous experience in the workforce.

“Also, many recent college graduates who are unemployed or out of the labor force but want a job do not report an occupation because of their lack of previous experience in the workforce.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Because of the federal government shutdown in fall 2025, the CPS did not collect October 2025 data, and the data were not collected retroactively.

“Due to the federal government shutdown in fall 2025, the CPS data are missing the October 2025 data collection.”

• Also stated in: § Figures and · plain

  “As noted by the BLS, “CPS data were not collected for October 2025 due to the lapse in appropriations and were not collected retroactively.”

• cites BLSNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Averaged over June, July and August, the underemployment proxy measure was 9.1% in 2026 versus 8.2% in 2022, and the occupational mismatch proxy measure was 17.1% versus 15.6% in 2022.

“Averaging over June, July and August, the underemployment proxy measure was 9.1 percent in 2026 compared with 8.2 percent in 2022, and the occupational mismatch proxy measure was 17.1 percent compared with 15.6 percent in 2022.”

May combine two claims.

The estimated interaction between summer 2026 and the Anthropic-based AI exposure measure is small and not statistically significant in either the standard or the expanded specification.

“| Summer 2026 × AI exposure | 0.006 (0.007) | 0.008 (0.008) | 0.006 (0.007) | 0.009 (0.008) |”

• Also stated in: § Figures and · plain

  “AI exposure is defined using the standardized AI-exposure measure fromMassenkoff et al. (2026)which is the Anthropic Economic Index based on Claude usage patterns.”

• cites Massenkoff et al. (2026)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Among recent college graduates, 16% of the unemployed and 36% of the unemployed plus sidelined unemployed do not report an occupation.

“As noted above, among recent college graduates, 16 percent of the unemployed and 36 percent of the unemployed plus sidelined unemployed (i.e.”

May combine two claims.

The aggregate comparisons do not detect an increase in unemployment among recent college graduates in summer 2026, but they could mask increases in occupations most exposed to AI that are offset by decreases elsewhere.

“While the aggregate comparisons do not detect an increase in unemployment among recent college graduates in summer 2026, they could mask increases in the occupations most exposed to AI that are offset by decreases elsewhere.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Young non-college graduates might also be affected by AI, but they are likely less affected than recent college graduates because AI likely replaces standardized technical tasks more than consumer-facing or physical work.

“Although young individuals who did not graduate from college might also be affected by changes in AI use in the workplace they are likely to be less affected than recent college graduates given that AI likely replaces standardized technical tasks more so than consumer-facing or physical work.”

Probably not a substantial claim: may not pass S3 (load-bearing).

January 2022 is a useful starting point because 2022 has been described as a turning point in the pandemic and its labor market effects.

“January 2022 is a useful starting point because 2022 has been described as a turning point in the pandemic and its labor market effects (Fairlie 2026).”

• cites Fairlie 2026Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Brynjolfsson, Chandar, and Chen use ADP records to examine changes in young workers' employment across occupations with different AI exposure.

“Brynjolfsson, Chandar, and Chen (2026), for example, use ADP records to examine changes in young workers' employment across occupations with different AI exposure.”

• cites Brynjolfsson, Chandar, and Chen 2026Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

For June 2026, the reference week was Sunday, June 7 through Saturday, June 13, 2026.

“For June 2026, the reference week was Sunday, June 7 through Saturday, June 13, 2026.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The CPS is usually released within a month of data collection.

“The CPS is usually released within a month of when the data are collected.”

May combine two claims.

Automating entry-level tasks might reduce the immediate value of junior workers' output and change firms' incentives to hire and train them.

“Automating these entry-level tasks might therefore reduce the immediate value of junior workers' output while changing firms' incentives to hire and train junior workers.”

Classified as probably not a substantial claim. May combine two claims.

The comparison-group regressions control for the prior trend in overall unemployment and the trend in the young-older gap in unemployment.

“We also control for the prior trend in overall unemployment and the trend in the young-older gap in unemployment which might be important for identifying the effects of AI.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The CPS is the only source of monthly estimates of employment, unemployment, hours worked, college enrollment, and occupations.

“The CPS is the only source of monthly estimates of employment, unemployment, hours worked, college enrollment, and occupations.”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing). May combine two claims.

The growth of remote work impedes the informal training and mentoring junior workers require, making firms more reluctant to hire inexperienced workers.

“Studying an earlier period, they estimate that the growth of remote work impedes the informal training and mentoring that junior workers require, which makes firms more reluctant to hire inexperienced workers.”

• cites Emanuel, Harrington and Pallais, 2026 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Theoretical AI exposure is measured using the Eloundou et al. (2024) classification.

“For theoretical exposure, we use the measure developed by Eloundou et al. (2024)”

• cites Eloundou et al. 2024 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The main specification includes a linear time trend to capture longer-term pre-trends in unemployment.

“The main specification includes a linear time trend to capture longer-term pre-trends in unemployment.”

May combine two claims.

The event-study approach has the disadvantage that it uses only 2022 to characterize the non-AI labor market, and 2022 is four years before 2026.

“The approach has the disadvantage that it only uses information from 2022 to characterize the non-AI labor market, and 2022 is a full four years before 2026.”

Probably not a substantial claim: may not pass S2 (checkable).

Taken together, the adoption evidence makes 2026 particularly informative to examine.

“Taken together, the adoption evidence makes 2026 particularly informative to examine.”

Classified as probably not a substantial claim. May combine two claims.

The CPS is representative of the entire U.S. population and contains observations for more than 130,000 people.

“These surveys, conducted monthly by the U.S. Bureau of the Census and the U.S. Bureau of Labor Statistics, are representative of the entire U.S. population and contain observations for more than 130,000 people.”

Classified as probably not a substantial claim.

Observed AI exposure is measured using the Anthropic Economic Index by Massenkoff et al. (2026).

“For observed exposure, we use the Anthropic Economic Index by Massenkoff et al. (2026)”

• cites Massenkoff et al. 2026Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

The official unemployment measure generally requires individuals without work to have actively searched in the previous four weeks.

“The official unemployment measure generally requires individuals without work to have actively searched in the previous four weeks.”

May combine two claims.

AI may augment workers and be particularly beneficial to new entrants.

“AI may augment workers and be particularly beneficial to new entrants”

Classified as probably not a substantial claim.

The CPS captures current work activity, including recent graduates who have a job but have not started yet.

“The CPS captures the current work activity of the individual including recent graduates who have a job but have not started yet.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Payroll records can capture employment changes at firms using specific payroll software.

“Payroll records can capture employment changes at firms utilizing specific payroll software”

Classified as probably not a substantial claim.

An association between occupational AI exposure and early-career outcomes could reflect more than one change within firms.

“An association between occupational AI exposure and early-career outcomes could therefore reflect more than one change within firms.”

Classified as probably not a substantial claim. May combine two claims.

The regressions control for individual characteristics (age, gender, race) and geographic characteristics (region and central city/suburb/rural location).

“The regressions control for individual characteristics (age, gender, race), and geographic characteristics (region of the country and central city/suburb/rural location).”

Probably not a substantial claim: may not pass S3 (load-bearing).

The CPS survey reference period is generally the calendar week containing the 12th day of the month.

“The CPS survey reference period is generally the calendar week that contains the 12 th day of the month.”

Classified as probably not a substantial claim. May combine two claims.

The results provide a timely assessment of early-career labor market conditions without establishing that AI caused the observed changes.

“Taken together, the results provide a timely assessment of early-career labor market conditions without establishing that AI caused the observed changes.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Unemployment insurance claims apply to workers with qualifying prior employment histories.

“unemployment insurance claims apply to workers with qualifying prior employment histories”

May combine two claims.

The CPS covers the population regardless of employment status, allowing observation of the extensive margins of labor force entry and exit and nonparticipation.

“Our analysis complements this emerging literature by leveraging individual-level data from the CPS which covers the population regardless of employment status, making it possible to observe the extensive margins of labor force entry and exit as well as nonparticipation.”

Classified as probably not a substantial claim.

The analysis controls for prior trends in unemployment because they may be important for identifying AI disruptions.

“The inclusion of a control for prior trends in unemployment might be important in identifying potential disruptions created by AI.”

May combine two claims.

Technical progress can either increase or reduce the advantage of experience.

“Thus, technical progress can either increase or reduce the advantage of experience.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The choice of the base year is difficult, especially because of the pandemic's disruption to labor markets.

“We show below that the choice of the base year is difficult especially because of the pandemic's disruption to labor markets.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Job postings measure advertised labor demand rather than realized hiring.

“job postings measure advertised labor demand rather than realized hiring”

Classified as probably not a substantial claim.

The years 2020 and 2021 are excluded from the seasonal unemployment profile because of the economic disruption and high unemployment caused by COVID.

“The years 2020 and 2021 are excluded because of the economic disruption caused by COVID and resulting high unemployment rates during these years.”

Probably not a substantial claim: may not pass S2 (checkable).

These findings do not rule out larger effects in the future.

“These findings do not rule out larger effects in the future.”

May combine two claims.

The regression estimates are consistent with the raw rate comparison: the summer 2026 coefficient is small and not statistically distinguishable from zero in all specifications, and the event-study estimates do not place the largest increases in summer 2026.

“The regression estimates are consistent with this comparison of raw rates, as the coefficient on summer 2026 is small and not statistically distinguishable from zero in all specifications, and the event-study estimates do not place the largest increases in summer 2026.”

Probably not a substantial claim: may not pass S2 (checkable).

These patterns suggest some caution in interpreting estimates of the effects of AI exposure and remote work availability.

“These patterns suggest some caution in interpreting estimates of the effects of AI-exposure and remote work availability.”

Classified as probably not a substantial claim.

The results provide a descriptive picture of the labor market for recent college graduates in the first summer in which AI use in the workplace was widespread enough to expect observable effects on employment.

“Taken together, the results provide a descriptive picture of the labor market for recent college graduates in the first summer in which AI use in the workplace was widespread enough to expect observable effects on employment.”

• AssumptionRests on an assumption: Summer 2026 is the first summer in which AI use in the workplace was widespread enough to expect observable effects on employment.

May combine two claims.

Recent increases in the intensity of workplace AI use make this an informative period to examine, but neither the timing of adoption nor the available comparison groups provides a clean basis for causal interpretation.

“Recent increases in the intensity of workplace AI use make this an informative period to examine, but neither the timing of adoption nor the available comparison groups provides a clean basis for causal interpretation.”

• AssumptionRests on an assumption: There have been recent accelerations in firms' AI usage intensity.

Probably not a substantial claim: may not pass S3 (load-bearing).

These data sources provide valuable evidence on different aspects of the labor market such as employment at certain firms, advertised labor demand, and job loss among workers with prior employment.

“These data sources provide valuable evidence on different aspects of the labor market such as employment at certain firms, advertised labor demand, and job loss among workers with prior employment.”

Classified as probably not a substantial claim.

The CPS is uniquely poised to examine this question, providing individual-level information on current school enrollment, own educational attainment, and detailed labor force status, beyond any particular employer or platform.

“The CPS is uniquely poised to let us examine this question, providing us with individual-level information on current school enrollment, own educational attainment, and detailed labor force status, beyond any particular employer or platform.”

Classified as probably not a substantial claim. May combine two claims.

College graduates and non-graduates aged 22-25 are in the same age range and face the same broad macroeconomic conditions, though changes in labor demand may affect them differently.

“This comparison is worth examining further because the two groups are in the same age range and face the same broad macroeconomic conditions, although changes in labor demand may affect them differently.”

May combine two claims.

Although neither comparison group is a clean control, the fact that recent graduates did not lose significant ground against two very different groups reinforces the conclusion from the level estimates.

“While neither group is a clean control, the fact that recent graduates did not lose significant ground against two very different groups nevertheless reinforces the conclusion from the level estimates.”

Classified as probably not a substantial claim. May combine two claims.

The main analysis includes graduates who have not established an occupational history, even though these individuals cannot always be included in the occupational-exposure analysis.

“Importantly, our main analysis includes graduates who have not established an occupational history, even though these individuals cannot always be included in the occupational-exposure analysis.”