Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI

Disagreements likely reflect measurement error in registers because firms do not update occupational switches of existing employees.

“The disagreements in Table A.5 likely reflect measurement error in the registers because firms generally do not update occupational switches of existing employees (Groes, Kircher and Manovskii, 2015).”

• cites Groes, Kircher and Manovskii, 2015 sourceDisputed

  Appendix OA9 says miscoding is more likely when a worker switches firms, not because firms fail to update the occupations of employees who stay. Section 2.4.5 describes the occupational data as generally highly reliable, and the paper analyzes recorded occupational switches within firms. The source does not discuss the citing paper’s Table A.5. (LLM, conf=0.84)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Case studies point to notable effects on online labor platforms.

“case studies indicate notable effects on online labor market platforms”

• cites Teutloff et al., 2025 sourceSupported

Randomized trials find productivity gains above 15% in the occupations studied here.

“significant productivity gains-often exceeding 15%-documented by RCTs in our study occupations”

• cites Brynjolfsson, Li and Raymond, 2025 sourceNot checked — no full text of the source
• cites Noy and Zhang, 2023 sourceNot checked — no full text of the source
• cites Peng et al., 2023 sourceSupported

No other new technology has been adopted by workers as quickly as AI chatbots.

“AI chatbots have seen the fastest worker take-up of any new technology”

• cites Bick, Blandin and Deming, 2025Not matched

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

• cites Humlum and Vestergaard, 2025 sourcePartly supported

  The Abstract (page unavailable) reports that 41% of surveyed workers in exposed occupations had used ChatGPT for work about a year after launch, describing its adoption as a landmark event. The Conclusion (page unavailable) also calls the take-up rapid. Neither passage establishes that AI chatbots had the fastest worker adoption of any new technology. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

The 2023 survey found that 39.8% of workers had used ChatGPT for work.

“For example, our 2023 survey found that 39.8% of workers had used ChatGPT for work, with adoption rates 15.9 percentage points lower among women within occupations (Humlum and Vestergaard, 2025).”

• cites Humlum and Vestergaard, 2025 sourceSupported

In the 2023 survey, adoption rates were 15.9 percentage points lower among women within occupations.

“For example, our 2023 survey found that 39.8% of workers had used ChatGPT for work, with adoption rates 15.9 percentage points lower among women within occupations (Humlum and Vestergaard, 2025).”

• cites Humlum and Vestergaard, 2025 sourceSupported

In the 2023 survey round, women were significantly more likely than men to cite lack of training as a barrier to ChatGPT adoption.

“The effectiveness of training aligns with findings from the 2023 survey round, in which women were significantly more likely to cite lack of training as a barrier to ChatGPT adoption (Humlum and Vestergaard, 2025).”

• cites Humlum and Vestergaard, 2025 sourcePartly supported

  The Gender Gaps section says women were more likely to report training-related adoption barriers and that 45% of the affected women said they needed training. It does not explicitly establish statistical significance for that comparison. The Survey section dates data collection to November 2023–January 2024, rather than a distinct 2023 survey round. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Firm-based complementary investment is important for unlocking the work-related benefits of new technologies.

“This highlights the importance of firm-based complementary investment for unlocking the work-related benefits of new technologies (Brynjolfsson and Hitt, 2000).”

• cites Brynjolfsson and Hitt, 2000 sourceSupported

Reported time savings are highest in occupations covered by RCTs, including customer and IT support, marketing and human resources, and software development.

“First, our reported time savings are highest precisely in the occupations covered by RCTs, such as customer and IT support (Brynjolfsson, Li and Raymond, 2025), marketing and human resources (Noy and Zhang, 2023), and software development (Peng et al., 2023).”

• cites Brynjolfsson, Li and Raymond, 2025 sourceNot checked — no full text of the source
• cites Noy and Zhang, 2023 sourceNot checked — no full text of the source
• cites Peng et al., 2023 sourceSupported

Dutz et al. (2025) developed an econometric framework to reweight the sample based on workers' latent willingness to participate.

“Dutz et al. (2025) develop an econometric framework that uses this variation to reweight the sample based on workers' latent willingness to participate; see Humlum and Vestergaard (2025) for its application to our survey.”

• cites Dutz et al. (2025)Not matched

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

• cites Humlum and Vestergaard (2025) sourceSupported

The estimates are not larger for adopters whose employers encourage use.

“are not larger among adopters whose use is encouraged by employers”

In every exposed occupation, a majority of users (64%-90%) report time savings from AI chatbots.

“AI chatbots save time across all exposed occupations (for 64%-90% of users)”

30% of employees have received AI chatbot training.

“30% of employees have received training”

Limited impacts persist in quasi-experimental analysis based on employer policies.

“These limited impacts persist in our quasi-experimental analysis based on employer policies”

Non-users perceive no chatbot-related changes in earnings.

“Direct survey responses from non-users confirm that they perceive no chatbot-related changes in earnings.”

AI chatbots have become widely used.

“AI chatbots are now widespread”

When employers encourage chatbot use, reported benefits are 10%-40% larger.

“the benefits from AI chatbots-time savings, quality improvements, creativity, task expansion, and job satisfaction-are 10%-40% greater when employers encourage their usage”

Most employers encourage their workers to use AI chatbots.

“most employers encourage their use”

The impact of AI chatbots on work quality and job satisfaction varies.

“their impact on work quality and job satisfaction varies”

Occupation-level estimates are close to zero, generally excluding changes greater than 6%.

“At the occupation level, estimates are similarly close to zero, generally excluding changes greater than 6%.”

DiD estimates show no differential trends over time.

“difference-in-differences estimates show no differential trends over time”

Average time savings reported by users are roughly 3% of work hours (2.8% in the main text).

“average time savings of 3%”

• Also stated in: § Abstract · plain

  “users report average time savings of 2.8% of work hours”

38% of employers deploy in-house AI chatbot models.

“38% deploy in-house models”

The economic outcomes of adopters show little change from AI chatbots.

“Our main finding is that AI chatbots have had minimal impact on adopters' economic outcomes.”

Employer encouragement is relatively more important for more intensive usage.

“The relative importance of employer encouragement becomes even more pronounced for more intensive usage.”

Firm-led investments boost adoption, nearly doubling take-up from 47% to 83%.

“These initiatives greatly boost adoption, nearly doubling take-up rates for the typical worker from 47% to 83%.”

As of November 2024, workers overwhelmingly reported no earnings impact.

“workers overwhelmingly report no impact on earnings as of November 2024”

The pass-through elasticity is larger at firms that encourage chatbot use.

“with greater elasticity at firms that encourage chatbot use”

Training initiatives are particularly effective at narrowing demographic gaps.

“with training initiatives proving particularly effective”

Between 3% and 7% of productivity gains translate into higher pay.

“only 3-7% of workers' productivity gains are passed through to higher earnings”

Encouragement from employers reduces the adoption gender gap from 11.9 to 5 percentage points.

“the gender gap in chatbot adoption shrinks from 11.9 to 5 percentage points when firms actively encourage use”

On days when adopters use the tools, average time savings are only 6-7%, substantially below experimental estimates.

“On days when they do use the tools, however, average time savings still amount to just 6-7%, substantially below experimental estimates.”

Software developers are more than twice as likely as journalists to report increased job satisfaction from chatbot use (30.5% vs. 12.6%).

“while software developers are more than twice as likely as journalists to report increased job satisfaction from chatbot use (30.5% vs. 12.6%).”

Reported savings are about half as large among teachers, accountants, and financial advisors.

“In contrast, reported savings are about half as large among teachers, accountants, and financial advisors, whose tasks may be less suited to chatbot assistance.”

The vast majority (80%) of users reallocate time saved from AI chatbots to other job tasks.

“Figure 4.(c) shows that the vast majority (80%) reallocate this time to other job tasks, while fewer than 10% use it for additional breaks or leisure.”

Workers primarily cite time savings in completing job tasks as a benefit of AI chatbots.

“Figure 4 shows that workers primarily cite time savings in completing job tasks (Panel (a)), with an average of 25 minutes saved per day of use (Panel (b)).”

25% of users spend more time on the same tasks they initially saved time on, especially when employers actively encourage chatbot use.

“In addition, 25% spend more time on the same tasks they initially saved time on, especially when employers actively encourage chatbot use.”

The survey indicates that many users do not operate under conditions as favorable as those in RCTs.

“While such estimates provide valuable insight into the workplace potential of AI chatbots, our survey indicates that many users do not operate under similarly favorable conditions.”

The gender gap in AI chatbot adoption persists even in workplaces where the tools are prohibited.

“Second, the gender gap in chatbot adoption persists even in workplaces where the tools are prohibited.”

New workloads are 20-50% more pronounced in workplaces that encourage AI chatbot use.

“Importantly, new workloads are 20-50% more pronounced in workplaces that encourage the use.”

Across 11 occupations and employer policies, average time savings amount to 2.8% of users' total work hours.

“Across our 11 occupations and employer policies, the average time savings amount to 2.8% of the total work hours of users.”

Across all occupations, workplaces that encourage AI chatbot use show greater task creation from these tools.

“Notably, across all occupations, workplaces that encourage AI chatbot use experience greater task creation from these tools.”

Average time savings from AI chatbots vary from 6.8% among marketing professionals whose employers encourage their use to 0.6% among teachers in schools that do not encourage their use.

“In Table 2 Average time savings from AI chatbots vary from 6.8% among marketing professionals whose employers encourage their use to 0.6% among teachers in schools that do not encourage their use.”

Employer initiatives help narrow demographic gaps in AI chatbot take-up.

“We find that employer initiatives-especially those focused on training-substantially increase adoption and help narrow demographic gaps in take-up.”

The benefits from AI chatbots-time savings, quality improvements, creativity, task expansion, and job satisfaction-are all 10%-40% greater when employers encourage their usage.

“Importantly, the benefits from AI chatbots-time savings, quality improvements, creativity, task expansion, and job satisfaction-are all 10%-40% greater when employers encourage their usage.”

Employer bans are associated with a reduction in overall AI chatbot adoption to roughly one-third of the baseline rate.

“Although employer bans are associated with a reduction in overall adoption to roughly one-third of the baseline rate, about 80% of the baseline gender gap remains.”

The paper's time savings estimates align more closely with Bick, Blandin and Deming (2025), who find average reported time savings of 5.4% among users in a nationally representative U.S. survey.

“Our time savings estimates align more closely with those of Bick, Blandin and Deming (2025), who, in a nationally representative U.S. survey, find an average reported time savings of 5.4% among users.”

• cites Bick, Blandin and Deming (2025)Not matched

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

About 80% of the baseline gender gap in AI chatbot adoption remains in workplaces where the tools are prohibited.

“Although employer bans are associated with a reduction in overall adoption to roughly one-third of the baseline rate, about 80% of the baseline gender gap remains.”

Firm-provided training programs reduce the gender gap in AI chatbot uptake to 3.6 percentage points.

“First, firm-provided training programs are particularly effective, boosting overall adoption and reducing the gender gap in uptake to 3.6 percentage points.8”

The age gap in AI chatbot adoption shrinks by approximately 40% when employers promote adoption.

“Similarly, the age gap shrinks by approximately 40% when employers promote adoption.”

Fewer than 10% of users use time saved from AI chatbots for additional breaks or leisure.

“Figure 4.(c) shows that the vast majority (80%) reallocate this time to other job tasks, while fewer than 10% use it for additional breaks or leisure.”

Higher adoption rates among younger, less experienced, and higher-earning workers replicate patterns from the earlier survey round.

“Higher adoption rates among younger, less experienced, and higher-earning workers also replicate the patterns observed in our earlier round.9”

The reduction in adoption gaps is smaller among more experienced workers.

“In contrast, the reduction in gaps is smaller among more experienced workers, suggesting that established work habits may be more resistant to change-even when encouraged by employers.”

Marketers and software developers whose employers promote AI chatbots report time savings of approximately 7%.

“For example, marketers and software developers whose employers promote AI chatbots report time savings of approximately 7%.”

Employer-driven initiatives to support AI chatbot adoption are now common.

“We contribute to this evidence by documenting that employer-driven initiatives to support AI chatbot adoption are now common and play an important role in shaping overall adoption rates and demographic disparities in usage.10”

Even on usage days, adopters spend only 5-6% of their work hours actively using AI chatbots.

“Limited applicability may again be an important explanation: even on usage days, adopters spend only 5-6% of their work hours actively using AI chatbots (see Table B.2).”

Adoption gaps across demographic groups are narrower in firms where management promotes chatbot use.

“Importantly, the demographic gaps in adoption are markedly smaller in workplaces where employers actively encourage chatbot use.”

The average reported time savings for adopters is 2.8%.

“That said, the average reported time savings of 2.8% for adopters may seem small compared to the substantial productivity gains-often exceeding 15%-documented by RCTs in our study occupations.”

Increased job satisfaction is the least commonly reported benefit of AI chatbots.

“In contrast, increased job satisfaction is the least commonly reported benefit.11”

17% of users created new workloads due to AI chatbots.

“created new workloads for 17% of users (Panel (a)), with 4.4 percentage points performing more of the same tasks, 10.9 percentage points taking on new tasks, and 1.7 percentage points doing both.”

Marketing professionals are more than twice as likely as teachers to report that chatbots improve work quality (69.8% vs. 32.1%).

“For instance, marketing professionals are more than twice as likely as teachers to report that chatbots improve work quality (69.8% vs. 32.1%),”

Reported time savings are significantly higher when employers actively encourage chatbot use.

“Second, reported time savings are significantly higher when employers actively encourage chatbot use.”

The gender gap in AI chatbot adoption narrows from 12 to 5 percentage points in encouraged settings.

“In particular, the gender gap narrows from 12 to 5 percentage points in encouraged settings.”

The most common new task relates to integration of AI chatbots into the workplace, accounting for 15%-40% of all new tasks.

“The most common task relates to the integration of AI chatbots into the workplace, accounting for 15%-40% of all new tasks, with the highest shares observed in IT support and software development.13”

The highest shares of AI integration tasks are observed in IT support and software development.

“The most common task relates to the integration of AI chatbots into the workplace, accounting for 15%-40% of all new tasks, with the highest shares observed in IT support and software development.13”

Employer initiatives, especially training, substantially increase AI chatbot adoption.

“We find that employer initiatives-especially those focused on training-substantially increase adoption and help narrow demographic gaps in take-up.”

Firm-provided training programs increase overall AI chatbot adoption.

“First, firm-provided training programs are particularly effective, boosting overall adoption and reducing the gender gap in uptake to 3.6 percentage points.8”

The estimates remain flat even among adopters whose employers actively encourage chatbot use.

“The estimates remain flat even among adopters whose employers actively encourage chatbot use.”

Workers who report substantial benefits, such as daily time savings of more than an hour, do not show significantly better labor market outcomes.

“Second, while some workers do report substantial benefits from AI chatbots (e.g., daily time savings of more than an hour), we cannot detect significantly better labor market outcomes for these workers.”

Encouraged users exhibit even larger earnings premia.

“Panel (b) splits adopters based on whether their usage is employer-encouraged, showing that encouraged users exhibit even larger earnings premia.”

Two years after the fastest technology adoption ever, labor market outcomes at the individual and firm level remain untouched.

“Any account of transformational change must contend with a simple fact: two years after the fastest technology adoption ever, labor market outcomes-whether at the individual or firm level-remain untouched.”

Workers' perceived earnings effects are only weakly correlated with their reported time savings.

“Third, workers' perceived earnings effects are only weakly correlated with their reported time savings.”

Despite rapid adoption and substantial investments, AI chatbots have had minimal impact on productivity and labor market outcomes to date.

“Despite rapid adoption and substantial investments by both workers and firms, our key finding is that AI chatbots have had minimal impact on productivity and labor market outcomes to date.”

Workers in encouraged settings are better able to take on more work in response to chatbot use.

“Consistent with this view, Section 3 shows that workers in encouraged settings are better able to take on more work in response to chatbot use.”

The average perceived earnings impact ranges from 0.04% to 0.2%, depending on employer encouragement.

“First, workers overwhelmingly report that AI chatbots have not meaningfully affected their earnings: the average perceived earnings impact ranges from 0.04% to 0.2%, depending on whether their employer encourages their use.”

Workers reporting enhanced quality, creativity, or new workloads have not fared better in the labor market since AI chatbot introduction.

“Similarly, workers who report enhanced quality or creativity or new workloads from AI chatbot usage have not fared better in the labor market since their introduction.”

Employer initiatives shape the new workloads that AI chatbots create.

“This figure illustrates how employer initiatives influence the new workloads created by AI chatbots.”

• AssumptionRests on an assumption: Workers' self-reported AI chatbot use is accurate.

The paper finds only a weak relationship between workers' time savings and labor market effects from AI chatbots.

“In contrast, we find only a weak relationship between workers' time savings and labor market effects from AI chatbots.”

97% of workers report no change in earnings due to AI chatbots.

“Second, the limited perceived earnings effects primarily reflect that 97% of workers report no change in earnings due to AI chatbots-rather than large positive and negative effects canceling each other out (see Appendix D.2 for a detailed breakdown).”

The paper finds no significant impact on labor market outcomes across its 11 occupations.

“In contrast, we find no significant impact on labor market outcomes across our diverse set of 11 occupations.”

Employer policies influence how workers use AI chatbots.

“This figure illustrates the impact of employer policies on workers' use of AI chatbots.”

• AssumptionRests on an assumption: Workers' self-reported AI chatbot use is accurate.

Employer usage policies influence disparities among workers in adopting AI chatbots.

“This figure illustrates the impact of employer usage policies on worker disparities in AI chatbot adoption.”

• AssumptionRests on an assumption: Workers' self-reported AI chatbot use is accurate.

The earnings differentials entirely predate the arrival of AI chatbots.

“However, leveraging the panel dimension of our administrative data reveals that these earnings differentials entirely predate the arrival of AI chatbots.”

Dynamic difference-in-differences reveal that adopters are on slightly stronger labor market trends.

“The dynamic difference-in-differences in Figure 7.(c) reveal that adopters are on slightly stronger labor market trends.”

Most workers report using time saved by AI chatbots to perform other tasks, including new responsibilities.

“In our survey, most workers report using time saved by AI chatbots to perform other tasks, including new responsibilities”

Only 3-7% of estimated time savings translate into earnings.

“As the slopes in Figure 11 indicate, only 3-7% of their estimated time savings translate into earnings.”

Time savings, pass-through rates, and earnings impacts are significantly larger when employers encourage AI chatbot use.

“Fourth and finally, all these effects-time savings, pass-through rates, and ultimately earnings impacts-are significantly larger when employers encourage AI chatbot use.”

One and a half years after launch, AI chatbots have had minimal average impact on adopters' earnings.

“First, one and a half years after their launch, AI chatbots have had minimal average impact on adopters' earnings.”

Workplaces with higher rates of AI chatbot adoption have not fared differently in total wage bills and employment hours.

“Across the board, workplaces with higher rates of AI chatbot adoption have not fared differently.”

Section 3.2 showed that AI chatbots have already created new workloads for non-users, indicating broader workplace changes.

“Indeed, Section 3.2 showed that AI chatbots have already created new workloads for non-users, indicating broader workplace changes.”

Across all occupations studied, workers encouraged to use AI chatbots have not fared differently in the labor market.

“Across all of our occupations studied, workers who are encouraged to use AI chatbots have not fared differently in the labor market.”

Workers' self-reported earnings effects closely align with difference-in-differences estimates based on administrative data.

“Indeed, workers' self-reported earnings effects closely align with our difference-in-differences estimates based on administrative data.”

There is no evidence that workers initially employed at high-adoption workplaces have been affected differently.

“Once again, we find no evidence that workers initially employed at high-adoption workplaces have been affected differently.”

Quasi-experimental analyses using employer policies closely align with the main difference-in-differences estimates.

“Section 4.2.1 presents the results from these quasi-experimental analyses, which closely align with our main difference-in-differences estimates.”

Adopters earn about 7% more than comparable non-adopters in a difference-in-means comparison.

“The estimates show that adopters earn about 7% more than comparable non-adopters.”

Scaling point estimates with the standard deviation of workplace adoption rates, the study can rule out standardized effects larger than 1%.

“Scaling the point estimates with the standard deviation of workplace adoption rates (20pp after shrinkage; see Appendix G), we can rule out standardized effects larger than 1%.”

Despite employer encouragement boosting work-related benefits, there are no differential changes in earnings for adopters whose employers encourage use.

“Lastly, despite evidence in Section 3 that employer encouragement boosts work-related benefits of AI chatbots, we find no differential changes in earnings for adopters whose employers encourage their use.”

The paper can rule out effects on labor market outcomes comparable to the modest productivity gains of 2.8% time savings per user.

“Although the productivity gains from chatbot use are themselves modest-averaging 2.8% in time savings per user (see Section 3)-we can rule out comparable effects on labor market outcomes.”

Difference-in-differences estimates show no differential changes for adopters following the introduction of AI chatbots.

“When indexing differences to November 2022, the difference-in-differences estimates in Panel (c) show no differential changes for adopters following the introduction of AI chatbots.”

An earnings gap exists between AI chatbot adopters and non-adopters before and after ChatGPT's launch in November 2022.

“This figure shows the earnings gap between AI chatbot adopters and non-adopters, before and after the launch of ChatGPT in November 2022.”

• AssumptionRests on an assumption: Workers' self-reported AI chatbot use is accurate.
• AssumptionRests on an assumption: In the absence of ChatGPT, the earnings trajectories of encouraged and non-encouraged adopters would have followed parallel trends.

Workers in the survey report only modest time savings from AI chatbots.

“Indeed, workers in our survey report only modest time savings from AI chatbots.”

There is no evidence that AI chatbots have differential impacts on intensive users or early adopters.

“Finally, we find no evidence that AI chatbots have had differential impacts on workers who use them more intensively (i.e., daily users for work) or who adopted them early (i.e., within the first year of ChatGPT's launch).”

The estimated effects of encouragement on reported benefits hold up when firm and worker characteristics-firm age, size, productivity, and detailed task mixes-are controlled for.

“We show that the effects of encouragement remain robust when controlling for a range of firm-and worker-level characteristics, including firm age, size, and productivity, as well as workers' detailed task mixes and other attributes.”

Results in Section F.2.2 remain robust when using raw coworker encouragement rates instead of adjusted rates.

“Importantly, our results in Section F.2.2 remain robust when using the raw coworker encouragement rates instead.”

Coworker encouragement has a weaker effect on reported benefits among adopters than self-reported encouragement does.

“By contrast, coworker encouragement has a more muted effect on reported benefits among adopters (Panel (b)).”

Employer encouragement is linked to higher average benefits reported by users of AI chatbots.

“Figure 4 shows that employer encouragement is associated with greater average benefits reported by users.”

• AssumptionRests on an assumption: Workers' self-reported benefits and time savings accurately measure the actual benefits of chatbot use.

Estimates of the impact of employer encouragement on chatbot adoption and perceived benefits remain virtually unchanged after accounting for firm characteristics and worker task mixes.

“Table F.1 and Figure F.1 summarize these robustness checks, showing that our estimates of the impact of employer encouragement on chatbot adoption and perceived benefits remain virtually unchanged after accounting for firm characteristics and worker task mixes.”

Individual-level benefits such as perceived time savings appear more sensitive to whether workers personally feel encouraged, as measured by self-reports.

“However, individual-level benefits-such as perceived time savings-appear more sensitive to whether workers personally feel encouraged, as measured by self-reports.”

Survey responses from November 2024 likewise show no earnings effects of AI chatbots.

“Section 4.3 shows that survey responses from November 2024 similarly report no earnings impacts of AI chatbots.”

Roughly 10% to 15% of teachers who have not used AI chatbots report new workloads from the technology.

“Specifically, Figure C.1.(b)shows that about 10%-15% of teachers who have not used AI chatbots report new workloads from the technology.”

Even customized in-house chatbots are usually versions of ChatGPT.

“The table shows that even customized in-house chatbots are most often versions of ChatGPT.”

Coworker-reported encouragement raises adoption 1.5 to 2 times as much as self-reported encouragement.

“The impact of coworker-reported encouragement on adoption (Panel (a)) is 1.5-2 times larger than the effect of self-reported encouragement.”

Employer encouragement is associated with a doubling of adoption rates.

“the fact that employer encouragement in Figure 2 is associated with a doubling of adoption rates”

Within occupations, employer encouragement is associated with higher AI chatbot adoption across ever, monthly, weekly, and daily usage frequencies.

“Encouraged 0.363 * * * 0.337 * * * 0.285 * * * 0.137 * * *”

• Also stated in: § Adoption rates. · plain

  “Figure 2 shows how employer encouragement is associated with greater chatbot adoption by workers.”

Between 50% and 95% of the new tasks reported are directly linked to AI use.

“Figure C.2 shows that 50% to 95% of these tasks are directly linked to AI use.”

Results are robust to controlling for workers' detailed task mixes within occupations, so effects are not merely driven by task types more amenable to AI chatbot use.

“On the worker side, we show that the results are similarly robust to controlling for workers' detailed task mixes within occupations, ensuring that the effects of employer encouragement are not merely driven by differences in task types more amenable to AI chatbot use.”

Disagreements occur in cells reflecting likely job switches.

“Table A.5 shows the disagreements occur in cells that reflect likely switches, such as (IT Support, Software Developer).”

Workplace outcomes vary with AI adoption rates at the workplace-occupational level after ChatGPT's launch in November 2022.

“This figure shows changes in workplace outcomes by adoption rates, indexed to the launch of ChatGPT in November 2022, at the workplace-occupational level.”

• AssumptionRests on an assumption: Workers' self-reported AI chatbot use is accurate.
• AssumptionRests on an assumption: In the absence of ChatGPT, the earnings trajectories of encouraged and non-encouraged adopters would have followed parallel trends.

Survey and register agree on occupation for 100% of school teachers.

“By contrast, the survey and register data agree on the occupation of 100% of our school teachers.”

AI chatbots have generated new job tasks for many workers.

“Indeed, Section 3.2 shows that AI chatbots have created new job tasks for many workers.”

AI chatbot adopters show different labor market outcomes than non-adopters after ChatGPT's launch in November 2022.

“This figure presents the differential labor market outcomes of AI chatbot adopters relative to non-adopters, indexed to the launch of ChatGPT in November 2022.”

• AssumptionRests on an assumption: Workers' self-reported AI chatbot use is accurate.
• AssumptionRests on an assumption: In the absence of ChatGPT, the earnings trajectories of encouraged and non-encouraged adopters would have followed parallel trends.

Software developers and marketing professionals report time savings of about 7% when their employer encourages AI chatbot use.

“software developers and marketing professionals report time savings of approximately 7% when employers encourage AI chatbot use”

• AssumptionRests on an assumption: Workers' self-reported benefits and time savings accurately measure the actual benefits of chatbot use.

Firm characteristics predict which employer initiatives to promote AI chatbot use are adopted.

“This table examines which firm characteristics predict the adoption of various employer initiatives to promote AI chatbot use.”

• AssumptionRests on an assumption: Occupation fixed effects control for all relevant differences across occupations.

The typical standard deviation of workplace adoption rates within occupations is 20 percentage points.

“The typical standard deviation within occupations is 20 percentage points.”

AI chatbots have led to the creation of new tasks across all 11 occupations studied.

“Figures C.1 illustrate that AI chatbots have led to the creation of new tasks across all 11 occupations in our study.”

Coworker encouragement can fully close the gender gap in AI chatbot adoption.

“Figure F.2.(b) shows that coworker encouragement can entirely eliminate the gender gap in AI chatbot adoption.”

Five percent of non-users report new workloads caused by AI chatbots.

“.(b) shows that 5% of non-users report new workloads stemming from AI chatbots.”

All results remain robust when controlling for firm characteristics such as firm age, size, and productivity.

“On the firm side, we show that all results remain robust when controlling for firm characteristics (Table 1), ensuring that observed differences across workplaces are not driven by variation in firm age, size, or productivity.”

The paper's findings are robust to controlling for a wide range of observable firm and worker characteristics, mitigating omitted-variable concerns.

“First, in Section F.1, we show that our findings are robust to controlling for a wide range of observable firm and worker characteristics, mitigating concerns that omitted variables drive the results.”

Employer encouragement policies have had no measurable impact on workers' labor market outcomes.

“The results confirm that employer encouragement policies have had no measurable impact on workers' labor market outcomes.”

Effects of AI chatbot adoption on earnings larger than 1% are ruled out.

“(a)) rejects effects on earnings larger than 1%.”

Workers whose employers encourage AI chatbot use show different labor market outcomes than other workers after ChatGPT's launch in November 2022.

“This figure presents the differential labor market outcomes of workers who are encouraged to use AI chatbots by their employers, compared to all other workers, indexed to the launch of ChatGPT in November 2022.”

• AssumptionRests on an assumption: Workers' self-reported AI chatbot use is accurate.
• AssumptionRests on an assumption: In the absence of ChatGPT, the earnings trajectories of encouraged and non-encouraged adopters would have followed parallel trends.

Survey and registers agree on occupation for 87% of respondents.

“Table A.5 shows that the survey and registers agree on the occupation of 87% of our respondents.”

• Also stated in: § A.2.1 Representativeness and Response Quality · plain

  “The average agreement rate (diagonal elements) is 87%.”

Microsoft Copilot is an exception: only 39% of its daily users also use ChatGPT.

“An exception is Microsoft Copilot: only 39% of its daily users also use ChatGPT.”

The paper's estimated productivity gains are closer to Acemoglu (2025) than to other estimates.

“Our estimated productivity gains are closer to those ofAcemoglu (2025), though for different reasons.”

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

Results in Section 5.2 remain robust when using raw workplace adoption rates instead of Empirical Bayes shrinkage estimates.

“Notably, our results in Section 5.2 remain robust when using the raw workplace adoption rates instead.”

Users in encouraged settings report higher benefits from AI chatbots (Figure 4).

“Figure 4 shows higher reported benefits among users in encouraged settings is striking.”

Coworker-reported encouragement strongly predicts a worker's own reported encouragement, with a first-stage coefficient of about 1 and an F-statistic of 3,645.

“Table F.2 shows that coworker encouragement strongly predicts an individual worker's own reported encouragement: the first-stage coefficient is approximately 1, with an Fstatistic of 3,645.”

Firm-led complementary investments are important for unlocking the productivity potential of new technologies.

“This underscores the importance of firm-led complementary investments in unlocking the productivity potential of new technologies”

• cites Brynjolfsson, Rock and Syverson, 2021Not matched

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

• cites David, 1990Not matched

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

The study is sufficiently powered to rule out earnings effects as large as the reported time savings.

“we have statistical power to rule out comparable earnings effects”

Randomized trials indicate that chatbot effects can become negative when used on unsuitable tasks.

“RCTs show that the effects of AI chatbots can turn negative if applied to the wrong tasks”

• cites Dell'Acqua et al., 2023 sourceNot checked — no full text of the source
• cites Otis et al., 2024bNot matched

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

The results contradict expectations of an imminent labor market upheaval from Generative AI.

“our findings challenge narratives of imminent labor market transformations due to Generative AI”

Lack of firm-based complementary investments explains the discrepancy: many real-world workers lack favorable experimental conditions, limiting realized productivity gains.

“Second, the lack of firm-based complementary investments: many real-world workers do not operate under the same favorable conditions as those in experimental settings, limiting realized productivity gains.”

Randomized trials show large productivity gains for users.

“randomized controlled trials (RCTs) demonstrate substantial productivity gains for users”

• cites Brynjolfsson, Li and Raymond, 2025 sourceNot checked — no full text of the source
• cites Noy and Zhang, 2023 sourceNot checked — no full text of the source

The average causal effect of AI chatbots on economic outcomes is negligible.

“suggesting that the average causal effect is indeed negligible”

The small effects are not just a short-lived phenomenon.

“indicating that the limited impacts are not merely a very short-term phenomenon”

One reason for the gap is effect heterogeneity: trials target occupations with the biggest gains, whereas this survey covers several exposed occupations with smaller gains.

“First, effect heterogeneity: while existing RCTs concentrate on occupations with the largest productivity gains from AI chatbots, our broader survey reveals that several exposed occupations see more modest gains.”

The gender gap among Norwegian bachelor's students largely reflects male students continuing to use ChatGPT even when explicitly banned.

“This pattern mirrors findings by Carvajal, Franco and Isaksson (2024), who document a similar gender gap in ChatGPT use among Norwegian bachelor's students and show, using a vignette experiment, that the gap largely reflects male students continuing to use the tools even when explicitly banned.”

• cites Carvajal, Franco and Isaksson (2024) sourceNot checked — no full text of the source

Firm-led initiatives play an important role in enabling workers to expand task outputs in response to task-specific productivity gains from chatbots.

“This pattern suggests that firm-led initiatives play an important role in enabling workers to expand task outputs in response to the task-specific productivity gains from chatbots.”

Employer policies offer variation in chatbot adoption that is treated as quasi-exogenous.

“leverage employer policies to isolate quasi-exogenous variation in adoption”

Firm-led initiatives are important for realizing workplace transformations from new technologies.

“This highlights the importance of firm-led initiatives for realizing workplace transformations from new technologies.”

The rise of employer initiatives may explain much of the shift between the 2023 and 2024 surveys, namely increased adoption and a narrowed gender gap.

“For instance, the rise of employer initiatives may explain much of the shift observed between the 2023 and 2024 surveys-namely, increased adoption and a narrowed gender gap-as most initiatives were implemented in the interim.”

• AssumptionRests on an assumption: Most employer initiatives were implemented between the 2023 and 2024 surveys.

We are currently in a phase where substantial resources are being devoted to adapting workflows to AI chatbots.

“This pattern highlights that we are currently in a phase where substantial resources are being devoted to adapting workflows to AI chatbots.”

A similar gender gap in ChatGPT use exists among Norwegian bachelor's students.

“This pattern mirrors findings by Carvajal, Franco and Isaksson (2024), who document a similar gender gap in ChatGPT use among Norwegian bachelor's students and show, using a vignette experiment, that the gap largely reflects male students continuing to use the tools even when explicitly banned.”

• cites Carvajal, Franco and Isaksson (2024) sourceNot checked — no full text of the source

The pattern is consistent with firms being in the trough of a productivity J-curve.

“It is consistent with firms being in the trough of a productivity J-curve (Brynjolfsson, Rock and Syverson, 2021) and helps contextualize the modest time savings we observe in Table 2.”

• cites Brynjolfsson, Rock and Syverson, 2021Not matched

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

Established work habits may be more resistant to change even when employers encourage adoption.

“In contrast, the reduction in gaps is smaller among more experienced workers, suggesting that established work habits may be more resistant to change-even when encouraged by employers.”

Taken together, the widespread creation of new tasks, spillovers to non-users, and stronger effects in encouraging workplaces highlight broader workplace transformations caused by AI chatbots.

“Taken together, the widespread creation of new tasks, the spillovers to non-users, and the stronger effects in workplaces that encourage the tools all highlight the broader workplace transformations caused by AI chatbots.”

RCTs may capture the productivity effects of a well-coordinated and carefully managed adoption of these tools.

“In this sense, RCTs may capture the productivity effects of a well-coordinated and carefully managed adoption of these tools.”

Most exposed occupations do not fall into the extreme category of being highly substitutable by AI chatbots.

“First, while freelance proofreading may be highly substitutable by AI chatbots, most exposed occupations do not fall into this extreme category.”

Findings on limited pass-through of productivity gains caution against extrapolating freelance platform effects to the broader labor market.

“While the flexibility of freelance work makes it useful for identifying short-run changes in labor demand, our findings on the limited pass-through of productivity gains caution against extrapolating these effects to the broader labor market.”

Even if productivity increases, users of AI chatbots may be poorly positioned to bargain for higher pay when employers do not endorse the tools.

“Second, even if productivity increases, users of AI chatbots may be poorly positioned to bargain for higher pay when employers do not endorse the tools.”

Firm-led adaptations play a critical role in unlocking work-related benefits and translating them into labor market outcomes such as earnings.

“This highlights the critical role of firm-led adaptions, not only in unlocking the work-related benefits of the technology (as documented in Section 3) but also in translating these benefits into labor market outcomes such as earnings.”

The limited labor market impacts result from genuinely weak pass-through of time savings and other benefits, rather than workers generally misperceiving their gains.

“This suggests that the limited labor market impacts result from a genuinely weak pass-through of time savings and other benefits, rather than from workers generally misperceiving their gains from AI chatbots.”

Administrative data allows measuring workers' total earnings and hours regardless of how they reallocate time in the labor market.

“A key advantage of our administrative data is that it allows us to measure workers' total earnings and hours, regardless of how they reallocate their time in the labor market.”

Bottom-up adoption without managerial involvement may create principal-agent frictions that weaken the link between productivity and pay.

“The bottom-up nature of adoption-often occurring without managerial involvement-may create principal-agent frictions that weaken the link between productivity and pay.”

Existing literature provides little evidence on whether adopters of AI chatbots have fared better in the labor market.

“The existing literature provides little evidence on whether adopters of AI chatbots have fared better in the labor market, and a core contribution of this paper is to fill that gap.”

Perceived earnings effects collected in November 2024 provide reassurance that zero difference-in-differences estimates are not merely due to a short post-treatment window ending in June 2024.

“Moreover, because these perceived earnings effects were collected in November 2024, they provide reassurance that the zero difference-in-differences estimates in Section 4.2 are not merely a consequence of a short post-treatment window ending in June 2024.”

The study leverages quasi-exogenous variation from employer policies to address unobserved confounding shocks.

“Finally, to address unobserved confounding shocks (e.g., adopters facing sudden setbacks or tailwinds in the labor market after November 2022), we leverage quasiexogenous variation from the employer policies documented in Section 2.1.”

All results remain robust when using raw survey-based adoption rates.

“Notably, all our results remain robust when using the raw survey-based adoption rates.”

Firm-driven investments and workplace reorganizations are critical to unlocking AI's potential: take-up rates and productivity benefits are substantially higher when employers encourage usage, provide training, or deploy in-house models.

“First, consistent with Brynjolfsson, Rock and Syverson (2021), we find that firm-driven investments and workplace reorganizations are critical to unlocking AI's potential: Take-up rates and productivity benefits of AI chatbots are substantially higher when employers encourage usage, provide training, or deploy in-house models.”

• cites Brynjolfsson, Rock and Syverson (2021)Not matched

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

Existing studies tend to concentrate on occupations where productivity effects from AI chatbots are largest.

“This reiterates our observations from Section 3.3 that existing studies tend to concentrate on the occupations where the productivity effects from the tools are the largest.”

The results support the conclusion that AI chatbots have had no causal effect on adopters' labor market outcomes.

“These results support our conclusion that AI chatbots have had no causal effect on the labor market outcomes of adopters.”

No evidence of differential trends over time suggests that the limited effects are not merely a very short-run phenomenon.

“Moreover, we find no evidence of differential trends over time, suggesting that the limited effects are not merely a very short-run phenomenon.”

The study controls for workers' pre-determined characteristics including gender, age, and labor market experience to prevent these factors from driving estimates.

“First, we control for workers' pre-determined characteristics-including gender, age, and labor market experience-to ensure these factors do not drive our estimates (e.g., adopters being younger and naturally on upward earnings trajectories).”

The study uses a two-stage sampling strategy that maximizes coverage of sampled workplaces by first drawing random workplaces and then sampling all their employees.

“First, as described in Section 1.5, we implement a two-stage sampling strategy that maximizes the coverage of our sampled workplaces-first drawing a random set of workplaces, then sampling all their employees.”

This finding reinforces the conclusion that AI chatbots have not caused broader impacts on labor market outcomes such as equilibrium wages or hours.

“This finding reinforces our conclusion that-despite changes in workloads, even among non-users-AI chatbots have not caused broader impacts on labor market outcomes, such as equilibrium wages or hours.”

The absence of differential trends after introduction indicates that the minimal impacts are not merely a very short-term phenomenon.

“Second, these effects show no differential trends after the introduction of AI chatbots, indicating that the minimal impacts are not merely a very short-term phenomenon.”

Observable confounders do not explain the association between employer encouragement and reported benefits.

“These robustness checks show that the association between employer encouragement and reported benefits is not driven by observable confounders.”

The robustness of estimates to rich controls provides reassurance that observable confounders are not driving the results.

“The robustness of our estimates to the inclusion of this rich set of controls provides reassurance that observable confounders are not driving the results.”

The study lacks the statistical power to detect comparable pass-through effects in its difference-in-differences estimates of administrative labor market outcomes.

“We do not have sufficient statistical power to detect comparable pass-through effects in our differencein-differences estimates of administrative labor market outcomes (Section 4.2).”

Survey-reported occupations were cross-checked against administrative registers.

“As an external validation of our survey responses, we cross-check variables workers' reported occupations with those recorded in the administrative registers.”

Survey respondents resemble the survey populations on observable characteristics.

“Table A.2 show that our survey respondents resemble our survey populations on observable characteristics.”

The fact that employer encouragement in Figure 2 is associated with a doubling of adoption rates suggests that these employer initiatives are powerful relative to the amount of individual heterogeneity in net benefits.

“In this light, the fact that employer encouragement in Figure 2 is associated with a doubling of adoption rates suggests that these employer initiatives are powerful relative to the amount of individual heterogeneity in net benefits.”

• AssumptionThe authors state this as an assumption: Adoption exhibits positive selection on benefits, meaning employer encouragement induces adoption among workers with marginally lower idiosyncratic benefits.

The estimates in Table B.1 remain robust when worker characteristics are added as controls.

“The regressions also control for occupation fixed effects, and Table B.1 shows that the estimates are robust to adding controls for worker characteristics.”

• AssumptionRests on an assumption: Occupation fixed effects control for all relevant differences across occupations.

A common finding from RCTs is that AI chatbots tend to yield greater benefits for less experienced and typically younger workers.

“A common finding from RCTs is that AI chatbots tend to yield greater benefits for less experienced (and thus typically younger) workers (Brynjolfsson, Li and Raymond, 2025; Dell'Acqua et al., 2023; Noy and Zhang, 2023).”

• cites Brynjolfsson, Li and Raymond, 2025 sourceNot checked — no full text of the source
• cites Dell'Acqua et al., 2023 sourceNot checked — no full text of the source
• cites Noy and Zhang, 2023 sourceNot checked — no full text of the source

The workplace-based sampling design makes coworker measurements possible by ensuring that most respondents have coworkers who also participated in the survey.

“The coworker measurements are made possible by our workplace-based sampling design (described in Section 1.5), which ensures that most respondents have coworkers who also participated in the survey.”

Because of the negative selection effect, the average reported benefits among users is a lower bound on the causal effect of employer encouragement on individual benefits.

“In summary, the average reported benefits among users provide a lower bound on the causal effect of employer encouragement on individual benefits, due to the presence of a negative selection effect.”

Employer initiatives that reduce adoption costs lower the average benefit among users through a selection effect, because lower costs draw in workers with lower idiosyncratic benefits.

“First, employer initiatives that reduce adoption costs (α c ) lower the average benefit among users due to a selection effect: lower costs induce adoption by workers with lower idiosyncratic benefits (i.e., higher i).”

Employer encouragement mainly raises users' perceived benefits rather than lowering adoption costs, since cost reductions would only generate the negative selection effect.

“It suggests two things: (i) employer encouragement primarily raises perceived benefits (α b ), as reductions in adoption costs (α c ) would only generate the negative selection effect”

The negative selection effect γ is stronger when workers are more homogeneous in their net benefits from using AI chatbots.

“This negative selection effect, γ, is stronger when workers are more homogeneous in their net benefits from using AI chatbots (i.e., when”

Employer initiatives that raise benefits have two opposing effects on the average reported benefits among users.

“Second, employer initiatives that increase benefits (α b ) have two offsetting effects on average reported benefits.”

The results in Section F.2.2 remain when raw leave-one-out means are used instead of shrinkage-adjusted rates.

“Importantly, our results in Section F.2.2 are robust to using the raw leave-out means instead.”

An increase in αb raises the average reported benefits among users if and only if γ<1.

“The net effect of an increase in α b on the average reported benefits among users B is positive if and only if γ < 1.”

The high within-workplace correlation of reported employer encouragement indicates that these initiatives are centralized workplace policies.

“The strong correlation in reported employer encouragement at the workplace level supports our interpretation of these initiatives as centralized policies.”

The condition γ<1 is more likely to hold when workers are relatively homogeneous in their gross benefits but more heterogeneous in their adoption costs.

“This condition is more likely to hold when workers are relatively homogeneous in their gross benefits from the tools B i but more heterogeneous in their adoption costs C i .”