Labor Market Consequences of Generative AI: Early Evidence from Norway

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

Acemoglu, Autor, and Johnson (2026) identify early-career workers in affected occupations as the most vulnerable group, based on the notion that AI cannot yet substitute for experience accumulated by older workers.

“Acemoglu, Autor, and Johnson (2026) identify early-career workers in affected occupations as the most vulnerable group, based on the notion that artificial intelligence cannot yet substitute for experience accumulated by older workers.”

• Also stated in: § 2.2 Considerations specific to Generative AI · hedged

  “Nevertheless, Acemoglu et al. (2026) caution that while AI might not yet be able to substitute for experienced workers, a substantial age-gradient in capabilities exists, which puts entry-level workers in affected occupations at risk of competing with GenAI systems in terms of basic, codified knowledge.”

• cites Acemoglu et al., 2026 sourceDisputed

  Section 4.2.5 (p. 18) says expertise-leveling technologies can increase the value of less-expert workers while reducing the scarcity value of workers who previously performed those tasks. The summary table (p. 13) states, “Entrants benefit. Incumbent expertise potentially devalued.” That is contrary to the claimed rationale that accumulated experience protects older workers and makes early-career workers the most vulnerable. (LLM, conf=0.78)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

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

Acemoglu and Restrepo (2019) formalize the tension between displacement and reinstatement forces in a general-equilibrium model, showing that automation strictly reduces labor demand unless offset by the creation of new tasks in which labor has a comparative advantage.

“Acemoglu and Restrepo (2019) formalize this tension between displacement and reinstatement forces in a generalequilibrium model, showing that automation strictly reduces labor demand unless offset by the creation of new tasks in which labor has a comparative advantage.”

• cites Acemoglu and Restrepo, 2019 sourceDisputed

  Split into 2 atoms — 0 confirmed, 1 partially supported, 1 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

May combine two claims.

Norway is a coordinated labor market whose institutions differ sharply from those of the United States and the United Kingdom and, more subtly, from those of other Nordic countries.

“Second, Norway is a coordinated labor market whose institutions differ sharply from those of the United States and the United Kingdom and, more subtly, from those of other Nordic countries (Bhuller, Moene, Mogstad, & Vestad, 2022;Iacono, 2018;Mogstad, Salvanes, & Torsvik, 2025).”

• cites Bhuller et al., 2022 sourcePartly supported

  The Union Coordination section (page not provided) describes Norway as highly coordinated and contrasts Scandinavian bargaining systems with the United States’ low coordination. The Union Density and Bargaining Coverage section (page not provided) also identifies the United States and United Kingdom as having the lowest union influence. The passages do not establish the claimed, subtler institutional differences between Norway and other Nordic countries; they largely describe those countries together. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites Iacono, 2018 sourceNot checked — no full text of the source
• cites Mogstad et al., 2025Not 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).

Humlum and Vestergaard (2025) estimate null effects of observed AI-chatbot adoption on earnings and hours in a survey-linked subsample of Danish administrative records.

“Furthermore, while not testing the age-gradient directly and focusing on the intensive rather than the extensive margin, Humlum and Vestergaard (2025) estimate null effects of observed AI-chatbot adoption on earnings and hours in a survey-linked subsample of Danish administrative records.”

• Also stated in: § 2.3 Empirical evidence · hedged

  “While not testing the age-gradient hypothesis directly, Humlum and Vestergaard (2025) link adoption surveys to Danish administrative labor records and estimate null effects of AI-chatbot adoption on earnings and hours at both worker and workplace levels, although adoption is associated with occupational switching and task restructuring.”

• cites Humlum and Vestergaard, 2025 sourceSupported

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

Early studies on online platforms for freelancers and vacancy postings find substantial negative effects of GenAI on employment and remuneration.

“Early studies on online platforms for freelancers and vacancy postings find substantial negative effects of GenAI on employment and remuneration (Hui, Reshef, & Zhou, 2024;Teeselink, 2025;Teutloff et al., 2025).”

• cites Hui, Reshef, & Zhou, 2024 sourceNot checked — no full text of the source
• cites Teeselink, 2025 sourceNot checked — no full text of the source
• cites Teutloff et al., 2025 sourceSupported

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

Handa et al. (2025) map Claude.ai conversations onto O*NET tasks and occupations and document that AI use concentrates in software engineering, writing and analytical tasks, peaks in mid-to high-wage occupations rather than at either extreme of the wage distribution, and splits roughly evenly between automation and augmentation.

“Moving from predicted exposure to measured use, Handa et al. (2025) map Claude.aiconversations onto O*NET tasks and occupations and document that AI use concentrates in software engineering, writing and analytical tasks, peaks in mid-to high-wage occupations rather than at either extreme of the wage distribution, and splits roughly evenly between automation and augmentation.”

• cites Handa et al., 2025 sourceSupported

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

Pizzinelli et al. (2023) extend this measure to allow for potential complementarities with labor which could dampen displacement effects.

“Pizzinelli, Panton, Mendes Tavares, Cazzaniga, and Li (2023) extend this measure to allow for potential complementarities with labor which could dampen displacement effects.”

• cites Pizzinelli et al., 2023 sourceSupported

May combine two claims.

At the incumbent level, pooled DiD estimates are subject to noisy pretrends, and Rambachan and Roth (2023) identified sets include zero for wage, employment and mobility outcomes.

“We find that at the incumbent level, pooled DiD estimates are subject to noisy pretrends, and Rambachan and Roth (2023) identified sets include zero for wage, employ-ment and mobility outcomes.”

• cites Rambachan and Roth, 2023 sourceSupported

May combine two claims.

In the absence of firm-level adoption data for GenAI, and following the related literature, we treat the public release of ChatGPT in November 2022 as an availability shock for Generative AI tools.

“In the absence of firm-level adoption data for GenAI, and following the related literature (Brynjolfsson, Chandar, & Chen, 2025;Humlum & Vestergaard, 2025;Kauhanen & Rouvinen, 2026;Lodefalk et al., 2026), we treat the public release of ChatGPT in November 2022 as an availability shock for Generative AI tools.”

• cites Brynjolfsson, Chandar, & Chen, 2025Not matched

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

• cites Humlum & Vestergaard, 2025 sourceSupported
• cites Kauhanen & Rouvinen, 2026 sourceNot checked — no full text of the source

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

Acemoglu and Restrepo (2022) attribute 50 to 70 percent of observed changes in the US wage structure between 1980 and 2016 to the task displacement of worker groups specialized in routine work in industries experiencing rapid automation.

“Acemoglu and Restrepo (2022) attribute 50 to 70 percent of observed changes in the US wage structure between 1980 and 2016 to the task displacement of worker groups specialized in routine work in industries experiencing rapid automation.”

• cites Acemoglu and Restrepo, 2022 sourceSupported

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

Frank et al. (2026) caution that in U.S. unemployment-insurance records labor-market deterioration in AI-exposed occupations already began in early 2022.

“Lastly, while Frank, Sabet, Simon, Bana, and Yu (2026) caution that in U.S. unemployment-insurance records labor-market deterioration in AI-exposed occupations already began in early 2022”

• cites Frank et al., 2026 sourceSupported

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

Fanfani (2023) finds that contractual wage growth raises actual pay levels but has significantly negative employment effects, consistent with canonical centralized-bargaining models.

“Fanfani (2023) uses Italian administrative data matched to industry-wide collective agreements, and finds that contractual wage growth raises actual pay levels but has significantly negative employment effects, consistent with canonical centralized-bargaining models.”

• cites Fanfani, 2023 sourceSupported

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

Koch, Manuylov, and Smolka (2021) show that on firm level, robot adoption leads to output gains and net job creation within adopters, but substantial job losses at nonadopters.

“Koch, Manuylov, and Smolka (2021) show that on firm level, robot adoption leads to output gains and net job creation within adopters, but substantial job losses at nonadopters.”

• cites Koch et al., 2021 sourceSupported

May combine two claims.

The authors read the worker-level result as a null effect on incumbent labor-market outcomes, consistent with their institutional priors and with Humlum and Vestergaard's (2025) null worker-level findings.

“We interpret this as a null effect on incumbent labor-market outcomes, in line with our institutional priors and the null worker-level effects of Humlum and Vestergaard (2025) in Danish administrative data.”

• cites Humlum and Vestergaard (2025) sourceSupported

May combine two claims.

Three of the four pooled worker-level coefficients are significant at the 1% level, but event-study pre-trends are pronounced and HonestDiD bounds include zero for all four outcomes.

“Three of the four pooled coefficients are significant at the 1% level, but event-study pre-trends are pronounced and the HonestDiD bounds of Rambachan and Roth (2023), which adjust the post-treatment effect for the magnitude of the observed preperiod drift, include zero for all four outcomes.”

• cites Rambachan and Roth (2023) sourceSupported

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

An occupation's remote-work exposure correlates strongly with its GenAI exposure.

“Lambert and Schindler (2026) show that occupational exposure to working from home is strongly correlated with GenAI exposure”

• cites Lambert and Schindler (2026) sourceSupported

May combine two claims.

For all four worker-level outcomes, the HonestDiD identified set includes zero at M = 0 and the bounds widen monotonically as M increases.

“For all four outcomes the identified set already includes zero at M = 0 (no allowed violation), and the bounds widen monotonically with M .”

• cites Rambachan & Roth, 2023 sourcePartly supported

  Section 6.3 says that larger M allows larger deviations from linearity and that the reported confidence intervals widen as M increases. It does not report four worker-level outcomes or show that their identified sets contain zero at M = 0. In its own example, it says the robust confidence set for women contains only positive values when M < 0.01; it also explains that M = 0 still permits linear violations of parallel trends. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

• AssumptionThe authors state this as an assumption: The HonestDiD analysis assumes that violations of parallel trends are bounded by linear pre-trend extrapolations parameterized by M.

May combine two claims.

Event-study plots show strong pre-trends and the Rambachan and Roth (2023) bounds indicate null effects for most specifications.

“However, relevant event-study plots confirm (strong) pre-trends and Rambachan and Roth (2023) bounds squarely indicate null effects for most specifications.”

• cites Rambachan and Roth (2023) sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

The four occupational AI exposure indices are measured on different scales, ranging from a standardized score with mean near zero and standard deviation near one to a right-skewed usage-weighted share.

“The four indices live on different scales -Felten et al. (2021) delivers a standardized score with mean roughly zero and standard deviation roughly one; Pizzinelli et al. (2023) delivers a positive, bounded score on a wider range;”

• cites Felten et al. (2021) sourceNot checked — no full text of the source
• cites Pizzinelli et al. (2023) sourcePartly supported

  Section 2.3 (p. 15) says Pizzinelli et al.’s complementarity measure θ is bounded between 0 and 1, and Figure B.7 (p. 58) says AIOE and C-AIOE are normalized to that range for its plots. These passages support the existence of bounded scores, but do not establish that Pizzinelli et al.’s occupational AI exposure index is inherently a positive score on a wider range. They also do not substantiate the paraphrase’s right-skewed, usage-weighted-share description. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites Eloundou et al. (2024) sourceNot checked — no full text of the source

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

Cui et al. (2025), Gambacorta, Qiu, Shan, and Rees (2024), and Peng, Kalliamvakou, Cihon, and Demirer (2023) document strong productivity increases for software development tasks, with larger gains among less experienced developers.

“Cui et al. (2025); Gambacorta, Qiu, Shan, and Rees (2024); Peng, Kalliamvakou, Cihon, and Demirer (2023) document strong productivity increases for software development tasks, again with larger gains among less experienced developers.”

• cites Cui et al. (2025) sourceNot checked — no full text of the source
• cites Gambacorta, Qiu, Shan, and Rees (2024)Not matched

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

• cites Peng, Kalliamvakou, Cihon, and Demirer (2023) sourceSupported

May combine two claims.

Estimated coefficients for young workers are negative but small and statistically insignificant.

“While estimated coefficients for young workers are negative, in line with the existing literature, they are small and statistically insignificant.”

• Also stated in: § 1 Introduction · plain

  “Across all three designs, point estimates for the 22-25 age band are negative but never individually significant.”

May combine two claims.

The study finds no robust evidence of employment displacement among young workers in highly AI-exposed occupations, nor any robust response across other age cohorts or on incumbent labor-market outcomes.

“we find no robust evidence of employment displacement among young workers in highly AI-exposed occupations, nor any robust response across other age cohorts or on incumbent labor-market outcomes.”

• Also stated in: § 1 Introduction · plain

  “Overall, we observe no significant deterioration of the Norwegian labor market in response to GenAI, neither for early-career workers nor for incumbents.”

May combine two claims.

The four indices share most of their structure: at the occupational level the rank correlation between Felten et al. (2021), Eloundou et al. (2024) and Massenkoff and McCrory (2026) ranges from 0.79 to 0.89, while Pizzinelli et al. (2023) correlates more weakly (0.44-0.52) with the others.

“Across the 396-399 occupations for which exposure can be assigned (depending on the index), the four indices share most of their structure: at the occupational level the rank correlation between Felten et al. (2021), Eloundou et al. (2024) and Massenkoff and McCrory (2026) ranges from 0.79 to 0.89, while Pizzinelli et al. (2023) correlates more weakly (0.44-0.52) with the others.”

May combine two claims.

Norwegian inter-occupation mobility rates over our sample window are low and very stable over time, bounding the magnitude of any such bias.

“Furthermore, we observe that Norwegian inter-occupation mobility rates over our sample window are low and very stable over time,foot_5 bounding the magnitude of any such bias.”

May combine two claims.

On the shorter window, the 50+ GPT coefficient remains positive and significant on a dirty pre-trend.

“the 50+ GPT coefficient remains positive and significant on a dirty pre-trend.”

May combine two claims.

The 22-25 headcount cell under Eloundou is again negative (-0.024) but not statistically distinguishable from zero.

“The 22-25 headcount cell is again negative (-0.024) but again not statistically distinguishable from zero.”

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

May combine two claims.

The 22-25 SDID headcount ATT is again negative but insignificant, and a backdating placebo gives a larger absolute value.

“The 22-25 headcount ATT is again negative but not significant, while a backdating placebo yields a larger absolute estimate.”

May combine two claims.

For ages 22-25, the top exposure quartile shows a sharp employment drop starting in late 2023, whereas the other three quartiles stay roughly flat.

“in the youngest age band (22-25), employment in the highest-exposure quartile (Q4) declines markedly from late 2023 onward while the three lower quartiles remain broadly stable”

May combine two claims.

Loosening the filter to include firms with partial cross-quartile presence greatly enlarges the panel, but those extra firms usually have almost no within-firm Q4-versus-Q1-Q3 turnover and add little identifying variation.

“relaxing our filter to admit firms with partial cross-quartile presence (closer to their approach) inflates the panel substantially, but the additional firms typically have little to no within-firm Q4-vs-Q1-Q3 turnover and therefore contribute little identifying variation”

May combine two claims.

The other age bands use larger firm samples but produce either dirty pre-trends, null estimates, or both; the all-ages and 50+ panels in particular show steady secular growth in high-AI employment that the longer pre-period exposes.

“The remaining age bands are estimated on larger firm samples but either deliver dirty pre-trends, null point estimates, or both: the all-ages and 50+ panels in particular show steady secular increases in high-AI employment that our longer pre-period reveals.”

May combine two claims.

The 31-34 headcount and wage cells that were significant at the 5% level under Massenkoff become +0.020 (p = 0.29) and +0.012 (p = 0.21) under Eloundou, and the 50+ wage cell becomes +0.008 (p = 0.24).

“the 31-34 headcount and wage cells that were significant at the 5% level under Massenkoff and McCrory (2026) (Table 2) become +0.020(p = 0.29) and +0.012 (p = 0.21) respectively, and the 50+ wage cell becomes +0.008 (p = 0.24).”

• cites Massenkoff and McCrory (2026)Not matched

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

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

May combine two claims.

The SDID 22-25 headcount ATT is -0.013, negative like the within-firm -0.07 but about ten times smaller, with a 95% confidence interval that includes zero ([-0.083, +0.058]).

“the SDID 22-25 headcount ATT of -0.013 shares the negative sign with the within-firm -0.07 from Section 5.3 but lies an order of magnitude smaller in absolute value, with a 95% confidence interval that comfortably includes zero ([-0.083,+0.058])”

May combine two claims.

The 22-25 event-study path has a clean pre-trend and turns down faster near the end of the sample.

“the 22-25 event-study trajectory shows a clean pre-trend and accelerates downwards towards the end of the sample period”

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

In Table 12, results are again unstable across specifications.

“Here again results are unstable across specifications.”

May combine two claims.

The 2019H2 backdating yields a 22-25 headcount ATT of -0.030, larger in absolute value than the real-treatment estimate and itself indistinguishable from zero.

“The 2019H2 backdating produces a 22-25 headcount ATT of -0.030 (larger in magnitude than the realtreatment ATT) which is itself statistically indistinguishable from zero”

May combine two claims.

Under the Eloundou index, the 22-25 ChatGPT coefficient is again negative, smaller in magnitude (-0.066), and not statistically significant.

“The 22-25 Chat-GPT coefficient is again negative but smaller in magnitude (-0.066) and not statistically significant.”

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

May combine two claims.

Q3 workers are three to four years older than Q4 and earn substantially more.

“yet Q3 is older still by three to four years and earns substantially more”

May combine two claims.

Under the shorter window, the 22-25 ChatGPT coefficient stays negative but insignificant; under the Eloundou index it is also negative and similar in size (-0.066) but still insignificant.

“The 22-25 ChatGPT coefficient under the shorter window has the same negative sign but remains insignificant, under the Eloundou index the same coefficient is also negative and of similar magnitude (-0.066), but remains insignificant.”

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

The mean age of Q4 workers is 41.5, making them the second-oldest of the four exposure bins.

“Q4 workers have a mean age of 41.5 (the second-oldest among the four bins)”

May combine two claims.

In the 26-30 and 41-49 groups, Q3 and Q4 rise together until the end of 2022, after which Q4 turns down while Q3 continues to grow.

“For the 26-30 and 41-49 bands, Q3 and Q4 share a common upward trend through end 2022 but Q4 reverses post-treatment while Q3 keeps growing”

May combine two claims.

For 22-25-year-olds, the post-ChatGPT within-firm estimate is negative but insignificant, based on a small panel of just 45 firms.

“For workers aged 22-25, the post-ChatGPT estimate is negative but not statistically significant, and identified on a small panel of only 45 firms that survive our filters and balance requirements.”

May combine two claims.

On the shorter window, the 22-25 GPT coefficient remains negative, is smaller in magnitude (-0.012), and remains insignificant.

“the 22-25 GPT coefficient remains negative, but of smaller magnitude (-0.012) and still insignificant”

May combine two claims.

Only the 22-25 and 41-49 age bands show a negative ChatGPT coefficient, and it is significant only in the latter.

“The ChatGPT coefficient βGPT is negative only for the 22-25 and the 41-49 age bands, and only significant for the latter.”

May combine two claims.

For the 22-25 cohort, the shift-share estimate is again negative with a clean pre-trend but insignificant, based on a larger panel of 82 firms.

“For the 22-25 cohort the estimate is again negative on a clean pretrend but not statistically significant, although identified on a larger panel of 82 firms.”

May combine two claims.

The 50+ panel shows a strong pre-trend, with F(14) = 8.68.

“The strong pre-trend (F (14) = 8.68)”

May combine two claims.

For the 50+ panel, the ChatGPT coefficient is positive and statistically significant at +0.069.

“the 50+ panel returns a positive and statistically significant ChatGPT coefficient (+0.069 * * * )”

May combine two claims.

On the headcount outcome, every age band other than 22-25 places the most synthetic-control weight on Q3, then Q2, then Q1, with the pattern strongest for 31-34, 35-40 and 50+, where Q3 weights reach 0.43-0.45 against Q1 weights of 0.24-0.26.

“First, on the headcount outcome, every age band other than 22-25 puts the most weight on Q3, then Q2, then Q1; the pattern is most pronounced for 31-34, 35-40 and 50+, where Q3 weights reach 0.43-0.45against Q1 weights of 0.24-0.26.”

May combine two claims.

The 22-25 panel under Eloundou mirrors the Massenkoff result: pre-treatment coefficients are flat (consistent with the clean joint pre-trend test) and post-treatment coefficients drift modestly negative but are not statistically distinguishable from zero.

“The 22-25 panel mirrors the Massenkoff result: pre-treatment coefficients are flat (consistent with the clean joint pre-trend test) and post-treatment coefficients drift modestly negative but are not statistically distinguishable from zero.”

• cites Eloundou et al. (2024) sourceNot checked — no full text of the source
• cites Massenkoff and McCrory (2026)Not matched

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

May combine two claims.

The wage panel is visibly inconsistent with parallel trends: the pre-2022 coefficients walk monotonically from -0.008 (2015) to -0.002 (2021)-every one of them significantly different from zero at the 1% level-and the post-2022 estimates of +0.001 (2023) and +0.001 (2024) sit on this seven-year secular trend rather than breaking from it at ChatGPT's release.

“The wage panel is visibly inconsistent with parallel trends: the pre-2022 coefficients walk monotonically from -0.008 (2015) to -0.002 (2021)-every one of them significantly different from zero at the 1% level-and the post-2022 estimates of +0.001 (2023) and +0.001 (2024) sit on this seven-year secular trend rather than breaking from it at ChatGPT's release.”

• AssumptionRests on an assumption: The ChatGPT release in November 2022 constitutes a discrete treatment shock for AI-exposed occupations.

May combine two claims.

The AI exposure quartiles differ along observable dimensions: Q3 is older (mean age 44.5) and higher-paid (NOK 69,853 monthly cash wage) on average; Q1 has the highest immigrant share (29%) and the highest male share (64%); Q4 has the youngest mean age aside from...

“The quartiles differ along observable dimensions: Q3 is older (mean age 44.5) and higher-paid (NOK 69,853 monthly cash wage) on average; Q1 has the highest immigrant share (29%) and the highest male share (64%); Q4 has the youngest mean age aside from”

May combine two claims.

In the 31-34 and 35-40 age groups, employment in high-exposure occupations grew faster than in low-exposure occupations, particularly in Q3.

“These positives align with the descriptive pattern in Figure 1, where in these age groups employment in high-exposure occupations grew faster than in low-exposure occupations in particular in Q3.”

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

Restricting the panel to private-sector firms and full-time main jobs leaves 112.6 million worker-months.

“The resulting 112.6 million worker-months are then further restricted, depending on”

May combine two claims.

The 22-25 panel is remarkably flat through 2022 and drops below zero post-treatment.

“The 22-25 panel is by contrast remarkably flat through 2022 and drops below zero post-treatment”

May combine two claims.

The young-worker gap between Q4 and Q1 widens markedly post-ChatGPT: by early 2025, employment in high-AI occupations is roughly twenty percentage points lower (relative to October 2022) than employment in low-AI occupations among young workers, while the analogous gap among older workers is of opposite sign.

“The young-worker gap between Q4 and Q1 widens markedly post-ChatGPT: by early 2025, employment in high-AI occupations is roughly twenty percentage points lower (relative to October 2022) than employment in low-AI occupations among young workers, while the analogous gap among older workers is of opposite sign.”

• AssumptionRests on an assumption: The ChatGPT release in November 2022 constitutes a discrete treatment shock for AI-exposed occupations.

May combine two claims.

The 22-25 panels show a treated trajectory that tracks the synthetic control closely throughout the pre-period and exhibits no visible break at 2022H2 on either outcome.

“The 22-25 panels show a treated trajectory that tracks the synthetic control closely throughout the pre-period and exhibits no visible break at 2022H2 on either outcome.”

• AssumptionRests on an assumption: The synthetic control method provides a credible counterfactual for the treated age bands.

May combine two claims.

High-AI occupations grew faster than low-AI occupations throughout the pre-period and continued at roughly the same rate after 2022.

“Table 16 reports the sum of these weights have been growing faster than low-AI occupations throughout the pre-period, and continued to do so after 2022 at roughly the same rate.”

May combine two claims.

The synthetic-control matches are visibly close to the treated series in every panel through 2022H2; post-treatment divergences are within the bootstrap noise floor, with the 31-34 panels showing the most visible (though still modest) divergence consistent with their nominal significance in Table 2.

“The synthetic-control matches are visibly close to the treated series in every panel through 2022H2; post-treatment divergences are within the bootstrap noise floor, with the 31-34 panels showing the most visible (though still modest) divergence consistent with their nominal significance in Table2.”

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

The 26-30, 31-34 and 35-40 panels resemble the 22-25 panel in shape, and the 41-49 panel shows a modest post-2022 dip whose pre-trend test rejects parallel trends at the 5% level.

“the 26-30, 31-34 and 35-40 panels resemble the 22-25 panel in shape, and the 41-49 panel shows a modest post-2022 dip whose pre-trend test rejects parallel trends at the 5% level.”

May combine two claims.

The 50+ panels show treated and synthetic series that move together through the post-period; the small positive ATT on 50+ wage emerges from a gradual divergence after 2022H2, which is of comparable magnitude to the visible noise in the synthetic-control match.

“The 50+ panels likewise show treated and synthetic series that move together through the post-period; the small positive ATT on 50+ wage emerges from a gradual divergence after 2022H2, which however is of comparable magnitude to the visible noise in the synthetic-control match.”

• AssumptionRests on an assumption: The synthetic control method provides a credible counterfactual for the treated age bands.

May combine two claims.

The 2019H2 backdating placebo estimates are of comparable magnitude to the real-treatment ATTs in most age bands and are individually significant for 31-34 headcount (+0.046), 41-49 headcount (+0.036) and 50+ headcount (+0.038).

“partially driven by pre-existing trends, as the 2019H2 backdating placebos indicate: they are of comparable magnitude to the real-treatment ATTs in most bands, and individually significant for 31-34 headcount (+0.046 * * ), 41-49 headcount (+0.036 * * ), 50+ headcount (+0.038 * * ).”

May combine two claims.

About 70% of workers in the highest AI-exposure quartile (Q4) remain in Q4 in any given year, with no significant increase after treatment.

“Around 70% of Q4 workers stay within Q4 in any given year, with no significant uptick post-treatment.”

May combine two claims.

Under both the Massenkoff and McCrory (2026) and Eloundou et al. (2024) indices, real-treatment ATTs on headcount are small and positive in the all-ages and below-40 bands, while the 2019H2 backdating placebos are systematically larger and statistically significant in nearly every cell.

“The pattern is consistent across both indices: real-treatment ATTs are small and positive on headcount in the all-ages and below-40 bands, while the 2019H2 backdating placebos are systematically larger in magnitude and statistically significant in nearly every cell.”

• cites Massenkoff and McCrory (2026)Not matched

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

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

May combine two claims.

The pooled panel shows a steady upward pre-trend that the high F flags: high-AI shares within firms have been climbing since 2015, albeit with a pronounced slowdown after November 2022.

“The pooled panel shows a steady upward pre-trend that the high F flags: high-AI shares within firms have been climbing since 2015, albeit with a pronounced slowdown after November 2022.”

May combine two claims.

The pre-period drifts on the unbalanced panel appear similar to the baseline specification, with sharp reversals in trends post-treatment.

“The pre-period drifts appear similar to the baseline specification, with sharp reversals in trends post-treatment.”

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

Between 396 and 399 unique 4-digit STYRK occupations could be matched to a US-SOC code, depending on the exposure index.

“Table 7 reports summary statistics for each index after collapsing to the unique 4-digit STYRK occupations to which we successfully merged a US-SOC code (between 396 and 399 occupations, depending on the source).”

May combine two claims.

The Massenkoff distribution is heavily right-skewed: half of the occupations have an exposure score below 0.032, while the most exposed occupation reaches 0.493.

“The Massenkoff distribution is heavily right-skewed: half of the occupations have an exposure score below 0.032, while the most exposed occupation reaches 0.493.”

• cites Massenkoff and McCrory (2026)Not matched

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

May combine two claims. The quote only partly states this claim.

When the treated set is the above-median half of occupations, the 22-25 headcount ATT flips sign from -0.013 in Table 2 to +0.031 and remains statistically indistinguishable from zero.

“First, the 22-25 headcount ATT flips sign from -0.013 in Table 2 to +0.031 here, and remains statistically indistinguishable from zero.”

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

Returns to social skills are expected to rise relative to analytical skills.

“Specifically, returns to social skills are expected to rise relative to analytical skills”

• cites Freund and Mann, 2026Not matched

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

May combine two claims.

Existing evidence from administrative data suggests that short-run effects are small, heterogeneous across countries, and hard to separate from pre-existing trends.

“Taken together, existing evidence from administrative data suggests that short-run effects are small, heterogeneous across countries, and hard to separate from pre-existing trends.”

May combine two claims.

By 2025 around 15% of Norwegian firms with ten or more employees used at least one GenAI-related technology, compared with an EU-27 average of 9%.

“by 2025 around 15% of Norwegian firms with ten or more employees used at least one GenAI-related technology, compared with an EU-27 average of 9% (Eurostat, 2026a).”

• cites Eurostat, 2026aNot matched

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

May combine two claims.

Overall union density in Norway stood at roughly 50% in 2022, and collective-agreement coverage reached approximately 72% of Norwegian employees once the general-application mechanism is taken into account, up from 64% under directly signed agreements.

“Overall union density in Norway stood at roughly 50% in 2022, and collective-agreement coverage reached approximately 72% of Norwegian employees once the general-application (allmenngjøring) mechanism that extends sector-level agreements is taken into account, up from 64% of employees under directly signed agreements (Dølvik et al., 2022;Kjellberg, 2025).”

• cites Dølvik et al., 2022Not matched

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

• cites Kjellberg, 2025Not matched

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

May combine two claims.

In 2025 Norway had the highest population share of GenAI usage (35%) at work among European countries, well ahead of the other Nordics (21-27%) and an EU average of 15%.

“in 2025 Norway had the highest population share of GenAI usage (35%) at work among European countries, well ahead of the other Nordics (21-27%) and an EU average of merely 15% (Eurostat, 2026b).”

• cites Eurostat, 2026bNot 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 (2025) document a 13% relative decline in employment of 22-25-year-olds in the most AI-exposed occupations.

“Using worker-level data from the largest payroll provider in the United States, Brynjolfsson, Chandar, and Chen (2025) document a 13% relative decline in employment of 22-25-year-olds in the most AI-exposed occupations”

• Also stated in: § 2.3 Empirical evidence · boosted

  “a widely cited study by Brynjolfsson, Chandar, and Chen (2025) using high-frequency U.S. payroll data documents a 13-percent relative decline in employment of 22-25-year-olds in the most AI-exposed occupations”

• cites Brynjolfsson, Chandar, and Chen, 2025Not 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). May combine two claims.

Kauhanen and Rouvinen (2025, 2026) find no statistically significant overall or seniority-specific effects in Finnish population data.

“By contrast, Kauhanen andRouvinen (2025, 2026) find no statistically significant overall or seniority-specific effects in Finnish population data.”

• Also stated in: § 2.3 Empirical evidence · plain

  “Using Finnish population data, Kauhanen and Rouvinen (2025) find no statistically significant changes in wages or employment between more-and less-exposed occupations, and Kauhanen and Rouvinen (2026) find no systematic AI-linked displacement of younger workers.”

• cites Kauhanen and Rouvinen, 2025 sourceNot checked — no full text of the source
• cites Kauhanen and Rouvinen, 2026 sourceNot checked — no full text of the source

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

The net effect of new technologies on employment and wages remains ambiguous.

“The net effect on employment and wages therefore remains ambiguous.”

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

Lodefalk et al. (2026) document a 5.5% relative decline in employment of 22-25-year-olds in high-AI-exposure occupations in Sweden by early 2025.

“Lodefalk, L öthman, Koch, and Engberg (2026) document a 5.5% relative decline in employment of 22-25-year-olds in high-AI-exposure occupations in Sweden by early 2025.”

• Also stated in: § 2.3 Empirical evidence · plain

  “For Sweden, Lodefalk et al. (2026) combine 4.6 million job advertisements with full-population employer-employee register data, and find an accelerating decline in employment of 22-25-year-olds in high-AI-exposure occupations, reaching 5.5 percent by early 2025 relative to less exposed occupations within the same employers”

• cites Lodefalk 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 S3 (load-bearing).

Hosseini and Lichtinger (2025) find a corresponding drop in junior hiring at U.S. firms that adopt Generative AI.

“Hosseini and Lichtinger (2025) find a corresponding drop in junior hiring at U.S. firms that adopt Generative AI.”

• Also stated in: § 2.3 Empirical evidence · plain

  “In a similar setting, Hosseini and Lichtinger (2025) report a sharp drop in junior employment at firms adopting GenAI, which is driven by slower hiring rather than separations or promotions.”

• cites Hosseini and Lichtinger, 2025 sourceNot checked — no full text of the source

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

Eloundou et al. (2024) estimate that around 19% of the total U.S. workforce could see at least half of their tasks affected by large language models alone.

“Eloundou, Manning, Mishkin, and Rock (2024) estimate that around 19% of the total U.S. workforce could see at least half of their tasks affected by large language models (LLMs) alone.”

• Also stated in: § 2.2 Considerations specific to Generative AI · hedged

  “Linking tasks to occupations, they estimate that roughly 19 percent of the US workforce could see at least half of their tasks affected, with exposure generally rising in wages.”

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

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

Freund and Mann (2026) argue that workers with moderate overlap with GenAI may benefit from complementarities and productivity gains, while those with high exposure are harmed due to displacement effects.

“Applied to the ongoing automation wave, Freund and Mann (2026) argue that workers with moderate overlap with GenAI may benefit from complementarities and productivity gains, while those with high exposure are harmed due to displacement effects.”

• Also stated in: § 2.2 Considerations specific to Generative AI · plain

  “They project that workers with moderate AI exposure benefit on average while those with high exposure are harmed, with declining returns to analytical skills but rising returns to social ones.”

• cites Freund and Mann, 2026Not matched

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

Classified as probably not a substantial claim.

Standard errors are clustered at the firm-quartile level, since treatment varies within firm but is correlated within firm-quartile cells over the months of our window.

“Standard errors are clustered at the firm-quartile level, since treatment varies within firm but is correlated within firm-quartile cells over the months of our window.”

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

Additional post-treatment data may yet shift the picture.

“so additional post-treatment data may yet shift the picture.”

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

Acemoglu (2025) argues that based on the current capabilities of AI and given the task-content of contemporary occupations, the macro-economic effects of AI are likely to be modest.

“However, Acemoglu (2025) argues that based on the current capabilities of AI and given the task-content of contemporary occupations, the macro-economic effects of AI are likely to be modest.”

• cites Acemoglu, 2025 sourceNot checked — no full text of the source

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

Massenkoff and McCrory (2026) provide suggestive evidence that hiring of younger workers has slowed down in exposed occupations.

“however, they do provide suggestive evidence that hiring of younger workers has slowed down in exposed occupations.”

• cites Massenkoff and McCrory, 2026Not matched

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

May combine two claims.

Even by the end of 2023 the use of LLMs in European firms remained at around 2 percent on average, only picking up to 5 and 9 percent in 2024 and 2025.

“This is supported by the fact that even by the end of 2023 the use of LLMs in European firms remained at around 2 percent on average, only picking up to 5 and 9 percent in 2024 and 2025 (Eurostat, 2026a).”

• cites Eurostat, 2026aNot 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).

Lodefalk et al. (2026) interpret their findings as evidence that generative AI shifts hiring composition rather than reducing aggregate demand.

“They interpret their findings as evidence that generative AI shifts hiring composition rather than reducing aggregate demand.”

• cites Lodefalk et al., 2026Not matched

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

May combine two claims.

Noy and Zhang (2023) show that access to LLMs reduced completion time of mid-level professional writing tasks by 40 percent and raised quality by 18 percent, while compressing productivity differences between workers.

“Noy and Zhang (2023) show that access to LLMs reduced completion time of mid-level professional writing tasks by 40 percent and raised quality by 18 percent, while compressing productivity differences between workers.”

• cites Noy and Zhang, 2023 sourceNot checked — no full text of the source

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

Eloundou et al. (2024) construct an exposure measure by rating whether access to such technologies could reduce the time required to complete a given task by at least 50 percent.

“Targeting LLM capabilities specifically, Eloundou et al. (2024) construct an exposure measure by rating whether access to such technologies could reduce the time required to complete a given task by at least 50 percent.”

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

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

Mann and Püttmann (2023) document a positive influence of automation on employment in local labor markets: while manufacturing employment declined, this was more than compensated by service sector job growth.

“Using patent data, Mann and P üttmann (2023) document a positive influence of automation on employment in local labor markets: while manufacturing employment declined, this was more than compensated by service sector job growth.”

• cites Mann and Püttmann, 2023Not matched

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

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

We generally consider any results in the following as general equilibrium effects.

“While this attenuates estimated results, we generally consider any results in the following as general equilibrium effects.”

Classified as probably not a substantial claim.

Haapanala et al. (2023) argue that stronger organized labor moderates the unemployment effects of robot adoption across countries.

“Haapanala et al. (2023) argue that stronger organized labor moderates the unemployment effects of robot adoption across countries.”

• cites Haapanala et al., 2023Not matched

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

May combine two claims.

Highly coordinated and protected labor markets such as Norway's should display a muted wage and employment response of incumbents to AI exposure, while adjustments concentrated on the hiring margin might be stronger than in less regulated settings.

“Read together, this evidence implies that highly coordinated and protected labor markets such as Norway's should display a muted wage and employment response of incumbents to AI exposure, while adjustments concentrated on the hiring margin might be stronger than in less regulated settings.”

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

Felten, Raj, and Seamans (2021) map advances in AI capabilities to O*NET abilities, and then link these to occupational groups.

“Felten, Raj, and Seamans (2021) map advances in AI capabilities to O*NET abilities, and then link these to occupational groups.”

• cites Felten et al., 2021 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The theoretical arguments presented for computers and robots largely carry over to any kind of automation technology, including GenAI.

“The theoretical arguments presented for computers and robots largely carry over to any kind of automation technology, including GenAI.”

May combine two claims.

These were early versions with limited output quality and capabilities compared to today, and often they were not widely accessible or easy to use.

“However, these were early versions with limited output quality and capabilities compared to today, and often they were not widely accessible or easy to use.”

May combine two claims.

The treatment timing of November 2022 is well-used in the literature and anchored by the first release of ChatGPT.

“The treatment timing of November 2022 is well-used in the literature and anchored by the first release of ChatGPT.”

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

Teeselink (2025) confirms that highly LLM-exposed firms reduce employment, particularly junior hiring in technical and creative roles, and in high-wage segments.

“In the U.K., Teeselink (2025) confirms that highly LLM-exposed firms reduce employment, particularly junior hiring in technical and creative roles, and in high-wage segments.”

• cites Teeselink, 2025 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Parolin (2020b) finds that individual union membership increases the likelihood that workers in routine occupations remain employed and avoid low earnings.

“Similarly, Parolin (2020b) finds that individual union membership increases the likelihood that workers in routine occupations remain employed and avoid low earnings.”

• cites Parolin, 2020bNot matched

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

May combine two claims.

The Stable Unit Treatment Value Assumption (SUTVA) is hard to argue for here; if AI displaces workers from highly exposed occupations, they might move to less exposed ones, thereby contaminating the control group.

“Lastly, the Stable Unit Treatment Value Assumption (SUTVA) is hard to argue for here; if AI displaces workers from highly exposed occupations, they might move to less exposed ones, thereby contaminating the control group.”

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

Acemoglu and Restrepo (2020) find negative effects on employment and wages from industrial robot adoption.

“Furthermore, Acemoglu and Restrepo (2020) find negative effects on employment and wages from industrial robot adoption.”

• cites Acemoglu and Restrepo, 2020 sourceNot checked — no full text of the source

May combine two claims.

Lodefalk et al. (2026) distinguish effects of a pre-treatment interest-rate hiking cycle from those of GenAI, the latter of which they showed started to materialize only throughout 2024.

“Lodefalk et al. (2026) in their analysis distinguish effects of a pre-treatment interest-rate hiking cycle from those of GenAI, the latter of which they showed started to materialize only throughout 2024.”

• cites Lodefalk 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 four designs used in this study absorb different aspects of moderator-correlated dynamics.

“Beyond that, the four designs used in this study absorb different aspects of moderator-correlated dynamics.”

May combine two claims.

Brynjolfsson, Li, and Raymond (2025) document percent declines in demand for skills such as writing and translation, alongside sharp increases in demand for machine-learning programming and chatbot development.

“Brynjolfsson, Li, and Raymond (2025) percent declines in demand for skills such as writing and translation, alongside sharp increases in demand for machine-learning programming and chatbot development.”

• cites Brynjolfsson, Li, and Raymond, 2025 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

In Annex C we argue for resilience of our results to concurrent shocks.

“Furthermore, in Annex C we argue for resilience of our results to concurrent shocks.”

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

Modern work on labor-saving technology builds on the task-based framework popularized by Autor et al. (2003), who argue that computers substituted workers in routine cognitive and manual roles but complemented non-routine and complex tasks, leading to structural changes to the composition of jobs in the US economy.

“Modern work on labor-saving technology builds on the task-based framework popularized by Autor et al. (2003), who argue that computers substituted workers in routine cognitive and manual roles but complemented non-routine and complex tasks, leading to structural changes to the composition of jobs in the US economy.”

• cites Autor et al., 2003 sourceNot checked — no full text of the source

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

Freund and Mann (2026) develop a general-equilibrium model in which occupations bundle tasks and workers sort by comparative advantage across task-specific skills.

“Beyond measurement, Freund and Mann (2026) develop a general-equilibrium model in which occupations bundle tasks and workers sort by comparative advantage across task-specific skills.”

• cites Freund and Mann, 2026Not matched

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

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

Parolin (2020a) shows that in the United States between 1983-2017, higher union coverage cushioned earnings declines in occupations at greater risk of routine-biased automation, but accelerated the decline of their employment shares.

“Parolin (2020a) shows that in the United States between 1983-2017, higher union coverage cushioned earnings declines in occupations at greater risk of routine-biased automation, but accelerated the decline of their employment shares.”

• cites Parolin, 2020aNot 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). May combine two claims.

Massenkoff and McCrory (2026) combine the previous two approaches, and assign higher weights to observed usage in work-related and automated contexts.

“Massenkoff and McCrory (2026) combine the previous two approaches, and assign higher weights to observed usage in work-related and automated contexts (proxied by whether requests were made via an API).”

• cites Massenkoff and McCrory, 2026Not matched

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

Classified as probably not a substantial claim.

The interpretation that AI exposure causally moderates the post-2022 response requires the stronger factorial parallel trends assumption: that the moderator be mean-independent of potential-outcome trends and of other moderators.

“The interpretation that AI exposure causally moderates the post-2022 response instead requires the stronger factorial parallel trends assumption: that the moderator be mean-independent of potential-outcome trends and of other moderators.”

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

Massenkoff and McCrory (2026) find that actual usage remains below theoretical capabilities, and observe no increase in unemployment for highly exposed workers since late 2022 in the US.

“They find that actual usage remains below theoretical capabilities, and observe no increase in unemployment for highly exposed workers since late 2022 in the US”

• cites Massenkoff and McCrory, 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 S2 (checkable).

Any apparent post-2022 dynamics in high-AI occupations must be carefully judged as to whether they are mere continuations of pre-existing secular trends.

“and suggests that any apparent post-2022 dynamics in high-AI occupations must be carefully judged as to whether they are mere continuations of pre-existing secular trends.”

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

This warrants new types of classifications of jobs into those highly at risk from AI and those less so.

“This warrants new types of classifications of jobs into those highly at risk from AI and those less so.”

May combine two claims.

Labor market adjustment from GenAI is likely to be muted for wages and separations of incumbent workers in Norway, while new hiring could potentially be more constrained than in more flexible settings.

“Therefore, labor market adjustment from GenAI is likely to be muted for wages and separations of incumbent workers, while new hiring could potentially be more constrained than in more flexible settings.”

• AssumptionRests on an assumption: Norway has comparatively strict employment protection, including a just-cause requirement, notice periods of one to six months and routine availability of reinstatement.

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

The null at the worker level is consistent with expectations because Norway's labor markets are highly regulated, with wages often set by collective agreements and labor protection laws restricting layoffs.

“This null at the worker level is in line with expectations, considering that Norway's labor markets are highly regulated; wages are often subject to collective agreements, and labor protection laws restrict layoffs.”

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

The qualitative reading from Section 5.3 is preserved on the shorter window.

“The qualitative reading from Section 5.3 is preserved on this window”

May combine two claims.

By mid-2021, Norwegian employment levels and socioeconomic composition had recovered to pre-pandemic patterns for incumbents, with lasting damage limited to those out of work in early 2020.

“Alstadsæter, Bratsberg, Markussen, Raaum, and Røed (2025) document that by mid-2021 both the level and the socioeconomic composition of Norwegian employment had returned to pre-pandemic patterns for incumbents, with persistent damage confined to those out of work in early 2020.”

• cites Alstadsæter, Bratsberg, Markussen, Raaum, and Røed (2025)Not 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). May combine two claims.

The SDID base panel spans 2015H1 to 2024H2 and includes 396 occupations, distributed as 100 in Q1, 98 in Q2, 99 in Q3, and 99 in Q4.

“The base panel covers 396 occupations spanning 2015H1 to 2024H2 (Q1: 100, Q2: 98, Q3: 99, Q4: 99).”

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

The shorter window restricts the firm panel to cells observed in every month of 2019-2024, which tightens the pre-trend test but reduces the balanced firm sample.

“The shorter window restricts the firm panel to cells observed in every month of 2019-2024 (roughly four pre-treatment years), tightening the pre-trend test at the cost of a smaller balanced firm sample.”

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

The null at the worker level does not rule out any labor-market response.

“It however does not rule out any labor-market response”

May combine two claims.

The within-firm design selects firms that survive a balanced ten-year panel with substantial cross-quartile presence; such firms are large and diversified, so their composition margin is well measured but they may not represent the Norwegian private sector.

“the within-firm design conditions on firms surviving a balanced 10-year panel with substantial cross-quartile presence; these are large, structurally diversified firms whose composition margin is wellmeasured but who may not be representative of the Norwegian private sector”

May combine two claims.

Overall, the authors find no significant decline in the Norwegian labor market due to GenAI, for either early-career workers or incumbents.

“Overall, we observe no significant deterioration of the Norwegian labor market in response to GenAI, neither for early-career workers nor for incumbents.”

Classified as probably not a substantial claim.

As a falsification check, all SDID specifications are re-estimated with the post-period beginning in 2019H2, three years before ChatGPT.

“As a falsification exercise, we re-estimate all specifications with the post-period start moved to 2019H2 (dropping all post-2022H2 data), three years before ChatGPT and immediately before the Covid shock”

Classified as probably not a substantial claim.

A non-zero placebo ATT in the backdating window would indicate that the synthetic-control matching is not adequately absorbing pre-trend differentials.

“A non-zero placebo ATT in this window would indicate that the synthetic-control matching is not adequately absorbing pre-trend differentials.”

May combine two claims.

Across specifications, the evidence on within-firm displacement of 22-25-year-olds is consistent in direction but not conclusive in either case.

“The cross-spec evidence on within-firm 22-25 displacement is therefore directionally consistent but not conclusively identified in either specification.”

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

The shorter window aligns more closely with the shorter studies of Brynjolfsson, Chandar, and Chen (2025) and Lodefalk et al. (2026).

“This window is more in line with the shorter studies of Brynjolfsson, Chandar, and Chen (2025) and Lodefalk et al. (2026).”

• cites Brynjolfsson, Chandar, and Chen (2025)Not matched

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

• cites Lodefalk et al. (2026)Not matched

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

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

The PTA condition here can be interpreted as a flat within-firm Q4-vs-Q1Q3 relative-employment trend in the absence of treatment; we report the joint pre-trend Wald test.

“The PTA condition here can be interpreted as a flat within-firm Q4-vs-Q1Q3 relativeemployment trend in the absence of treatment; we report the joint pre-trend Wald test for βh , h < 2022H1.”

May combine two claims.

Under the Eloundou index, the pre-trend F is clean, but standard errors are wide and the sample size is almost as narrow as in the main specification.

“The pre-trend F is also clean here, but the SE's are wide and sample size is almost as narrow as in the main spec.”

May combine two claims.

The 50+ shift-share finding aligns with the within-firm 50+ result and with evidence from related work.

“This points in the same direction as the within-firm 50+ result, and is consistent with evidence from related studies.”

Classified as probably not a substantial claim.

The authors suggest that longer pre-treatment windows plus several complementary identification strategies are needed to separate AI-period breaks from secular trends in administrative data.

“This indicates that longer pre-treatment windows combined with multiple complementary identification strategies appear essential for distinguishing AI-period breaks from secular trends in administrative data.”

May combine two claims.

Across the three firm- and occupation-level designs, the evidence for declining early-career employment is consistent in direction but not robust, with 22-25 estimates negative yet never individually significant.

“Across all three firm-and occupation-level designs, we find directionally consistent but non-robust evidence of a deterioration of early-career employment: 22-25 point estimates are negative but never individually significant”

May combine two claims.

Across indices, the 22-25 within-firm result is directionally consistent but not conclusively identified under the preferred specification.

“Cross-index, then, the 22-25 within-firm result is directionally consistent but not conclusively identified under our preferred specification.”

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

Taken alone, the 50+ estimate would imply that older workers moved into high-AI occupations after late 2022.

“On its own this would suggest that older workers are concentrated into high-AI occupations after late 2022.”

May combine two claims.

It cannot speak to occupations whose employment varies through firm entry and exit, nor to AI's gradient on occupation-level wages once compositional churn is netted out.

“It cannot speak to occupations whose employment varies through firm entry and exit, nor to AI's gradient on occupation-level wages once compositional churn is netted out.”

May combine two claims.

The within-firm design isolates a reallocation effect within firms, but conditions on them surviving a balanced 10 year panel on narrow age bands.

“The within-firm design of Section 4.1 isolates a reallocation effect within firms, but conditions on them surviving a balanced 10 year panel on narrow age bands.”

May combine two claims.

The HonestDiD bounds and visible pre-period drifts rule out a causal reading of the pooled DiD estimates in Table 10; the apparent wage premium and mobility reductions reflect long-running secular composition trends rather than a discrete labor-market response to the ChatGPT release.

“Taken together with the visible pre-period drifts, these bounds rule out a causal interpretation of the pooled DiD estimates in Table 10: the apparent wage premium and mobility reductions reflect long-running secular composition trends rather than a discrete labor-market response to the ChatGPT release.”

• AssumptionThe authors state this as an assumption: The HonestDiD analysis assumes that violations of parallel trends are bounded by linear pre-trend extrapolations parameterized by M.
• AssumptionRests on an assumption: The ChatGPT release in November 2022 constitutes a discrete treatment shock for AI-exposed occupations.

Classified as probably not a substantial claim.

It is still subject to the parallel trends assumption.

“Furthermore, it is still subject to the parallel trends assumption.”

Classified as probably not a substantial claim.

The authors push the pre-treatment window back to January 2015, about twice as long as in comparable studies.

“we extend the pre-treatment window to January 2015, roughly doubling the length used in comparable studies”

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

The shift-share design requires balanced firm presence over 2015-2024 and at least five age-a workers per firm in 2015, which raises the 22-25 identifying firm count from 45 to 82.

“The firm panel is balanced at the firm level over the ten-year window and restricted to firms with at least five agea workers in 2015, so that the share matrix in equation ( 3) is computed on a non-trivial baseline.”

• Also stated in: § 5.5 Shift-share at the firm level · plain

  “The filters imposed here are balanced presence of firms over 2015-2024 and a 2015 baseline of at least five age-a workers per firm.For the 22-25 cohort this raises the identifying firm count from 45 (within-firm) to 82.”

May combine two claims.

After the balanced-panel and quartile-presence filters, the firm panel has 6,377 firms in the all-ages cross-section and just 45 firms in the 22-25 panel.

“The firm panel that survives the balanced-panel and quartile-presence filters contains 6,377 firms in the all-ages cross-section and shrinks for narrow age bins, reaching only 45 firms in the 22-25 panel.”

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

The SDID design keeps only occupations averaging at least 25 workers per pre-treatment half-year, leaving 89 treated and 261 control occupations in the all-ages run and 46 treated and 142 control occupations in the 22-25 panel.

“We impose the requirement that each occupation averages at least 25 workers per pre-treatment half-year to avoid effects from small and noisy job titles;foot_8 this retains 89 treated (Q4) and 261 control (Q1-Q3) occupations in the all-ages run and narrows to 46 treated and 142 control occupations in the 22-25 panel.”

May combine two claims.

The 22-25 point estimate agrees in direction with the negative US findings of Brynjolfsson, Chandar, and Chen (2025) and the Swedish findings of Lodefalk et al. (2026), but the confidence interval is too wide for a precise Norwegian estimate.

“The point estimate is directionally consistent with the negative findings of Brynjolfsson, Chandar, and Chen (2025) for the United States and Lodefalk et al. (2026) for Sweden; however our confidence interval is too wide to give a sharp estimate for Norway.”

• cites Brynjolfsson, Chandar, and Chen (2025)Not matched

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

• cites Lodefalk et al. (2026)Not 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).

C6 and C7 cluster at the 4-digit occupation (353 / 352 clusters).

“C6 and C7 cluster at the 4-digit occupation (353 / 352 clusters)”

May combine two claims.

Despite drift in the exposure measure through aging, the design still captures a relevant margin, since firms with many early-career workers in highly AI-exposed jobs should see overall employment fall if AI reduces demand for the tasks those workers typically perform.

“Thus, this design still captures a relevant margin: for example, firms with many earlycareer workers highly exposed to AI should see a decline in their overall employment if AI reduces demand for the tasks typically performed by such workers.”

May combine two claims.

The SDID time weights concentrate on the most recent pre-treatment observations, yet the synthetic control tracks the treated series across the full 2015H1-2022H1 pre-period.

“Note also that the SDID time weights concentrate on the most recent pre-treatment observations, yet the synthetic control tracks the treated series across the full 2015H1-2022H1 pre-period.”

• AssumptionRests on an assumption: The synthetic control method provides a credible counterfactual for the treated age bands.

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

Hui et al. (2024) find that on a large online platform the release of ChatGPT, DALL-E 2 and Midjourney reduced both employment and earnings of freelancers in highly affected occupations.

“Hui et al. (2024) find that on a large online platform the release of ChatGPT, DALL-E 2 and Midjourney reduced both employment and earnings of freelancers in highly affected occupations.”

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

May combine two claims.

A 15-percent average productivity gain in customer-support jobs is strongly concentrated among less experienced and lower-skilled workers.

“find a 15-percent average productivity gain in customer-support jobs, strongly concentrated among less experienced and lower-skilled workers.”

May combine two claims.

The SSB quarterly tables aggregate over the universe of employees by occupation, age and quarter rather than restricting to private-sector full-time workers, at the cost of coarser age granularity.

“Second, the SSB tables aggregate over the universe of employees by occupation × age × quarter rather than restricting to private-sector full-time workers, at the cost of coarser age granularity.”

• cites Statistics Norway, Table 11658Not matched

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

Classified as probably not a substantial claim.

Treated and control workers are reweighted using entropy balance on pre-2022 means of age, log wage, immigrant share and firm size, before re-estimating the baseline.

“Treated and control workers are reweighted using entropy balance on pre-2022 means of age, log wage, immigrant share and firm size, before re-estimating the baseline.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

All observed covariates other than the employment outcome are time-invariant and therefore fully absorbed by unit fixed effects.

“All other observed covariates are time-invariant and hence fully absorbed by unit fixed effects.”

May combine two claims.

The cross-index agreement on the qualitative pattern suggests the index choice is not the binding constraint on what can be identified, even if the 22-25 magnitude remains too noisy to support a strong causal reading.

“The cross-index agreement on the qualitative pattern suggests the index choice is not the binding constraint on what we can identify, even if the 22-25 magnitude remains too noisy to support a strong causal reading.”

May combine two claims.

The pooled coefficient on log wage in Table 10 captures the average level shift between the seven pre-treatment years (mostly negative coefficients) and the two posttreatment years (small positive coefficients), most of which is built up through pre-existing trend rather than a discrete break in 2022.

“The pooled coefficient on log wage in Table10therefore captures the average level shift between the seven pre-treatment years (which carry mostly negative coefficients) and the two posttreatment years (which carry small positive coefficients), most of which is built up through pre-existing trend rather than a discrete break in 2022.”

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

Age is typically correlated with job seniority and therefore with task content.

“However, age is typically correlated with job seniority and hence task content.”

May combine two claims.

The post-2022 employment decline among the youngest cohort is concentrated in the highest-exposure quartile while Q1-Q3 remain broadly stable, so pooling Q3 with Q4 in the treated group dilutes the Q4 signal.

“This is consistent with the descriptive evidence in Figure 1: the post-2022 employment decline among the youngest cohort workers is concentrated in the highest-exposure quartile, while Q1-Q3 remain broadly stable, so pooling Q3 with Q4 in the treated group dilutes the Q4 signal.”

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

Standard errors are computed via placebo bootstrap with 500 replications.

“Standard errors via placebo bootstrap (500 replications).”

May combine two claims.

This indicates that compositional effects are at play, motivating the firm- and occupation-level analyses.

“This indicates that compositional effects are at play here, motivating the firm-and occupation-level analyses.”

May combine two claims.

For the older cohorts, Q3 occupations track the pre-2022 trajectory of Q4 most closely, consistent with Q3's higher mean wage and older workforce.

“Q3 occupations appear to track the pre-2022 trajectory of Q4 most closely for these older cohorts, consistent with Q3's higher mean wage and older workforce.”

Classified as probably not a substantial claim.

Standard errors are clustered at the firm level (5,793 clusters).

“Standard errors clustered at the firm level (5,793 clusters).”

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

Lodefalk et al. (2026) flag the analogous Riksbank tightening as a potential confound for AI-period employment trends in Sweden.

“Lodefalk et al. (2026) flag the analogous Riksbank tightening as a potential confound for AI-period employment trends in Sweden.”

• cites Lodefalk et al. (2026)Not matched

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

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

Dropping the 2025H1 half-year avoids biasing headcount through seasonality, because the sample ends in March 2025 and the period would otherwise contain only its first three months.

“The 2025H1 half-year is dropped because our sample ends in March 2025 and would otherwise include only the first three months of the period, possibly biasing headcount due to seasonalities.”

May combine two claims.

Together with the main-spec results, the SSB panel reinforces the conclusion that apparent post-2022 employment dynamics in high-AI occupations are largely a continuation of pre-existing trends, and that any clean break attributable to ChatGPT is small, even when extending the analysis to early 2026.

“Together with the main-spec results, the SSB panel reinforces the conclusion that the apparent post-2022 employment dynamics in high-AI occupations are largely a continuation of pre-existing trends, and that the magnitude of any clean break attributable to ChatGPT is small, even when extending the analysis to early 2026.”

• AssumptionRests on an assumption: The SSB quarterly panel is representative enough to reinforce conclusions about high-AI occupations.

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

All specifications include worker fixed effects, firm-by-year fixed effects (year fixed effects only for the months-employed outcome), and a quadratic in age.

“All specifications include worker fixed effects, firm-by-year fixed effects (year fixed effects only for the months-employed outcome, see Section D), and a quadratic in age.”

Classified as probably not a substantial claim.

The raw A-ordningen microdata can contain multiple employment-relationship rows per worker within a given year and month.

“This step is necessary, since the raw A-ordningen microdata can contain multiple rows (employment relationships) per worker within a given (year, month).”

May combine two claims.

The cross-index pattern reinforces the reading that no SDID-significant cell replicates across both indices, so the three Massenkoff-significant cells are not attributed to a labour-demand response.

“The cross-index pattern reinforces our reading from the main text: there is no SDID-significant cell that replicates across both indices, so we do not attribute the three Massenkoff-significant cells to a labour-demand response.”