Skip to the review

INTERNATIONAL RESEARCH / WORKING PAPER

Is young worker employment falling in AI-exposed jobs?

In US payroll records, how did young worker employment change in occupations with more tasks that AI can perform? Employment among workers aged 22-25 stood 19% below the level it would have reached by keeping pace with their peers in less exposed occupations. Total headcount may not describe the prospects of people starting work.

ACADEMIC WORK REVIEWED

Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence

ResearchersErik Brynjolfsson, Bharat Chandar, Ruyu Chen

Stanford Digital Economy Lab; August 2026; Updated version, first circulated August 2025; Working paper · Source language: English

ENGLISH REVIEW AND COMMENTARY: DIJITALPI

We explain the paper in plain English, discuss what it may mean for organisations and identify our own examples separately.

English review: 16 November 2026 · The source date appears in the citation above.

Editorial review draft

The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.

LET'S READ THE RESEARCH TOGETHER

We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.

01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?

How does young worker employment in more AI-exposed occupations differ from employment among less exposed peers?

The researchers set out to describe US employment after the spread of generative AI by age and occupation. They examined how headcount changes relate to hiring, departures and base pay.

Explaining the context · DijitalPi commentary

DijitalPi commentary: A small business’s total headcount may not show the work available to new entrants. We read this study as a reason to consider task allocation and opportunities to gain experience together when planning hiring. Its findings do not determine how many employees a company needs.

02 / HOW WAS THE RESEARCH CONDUCTED?

How did they test the question?

The researchers used ADP’s anonymised US payroll records. Firms observed each month from January 2021 through June 2026 account for 3.5 to 5 million employees per month. The study covers full-time workers under 70 with positive earnings. A separate firm sample beginning in January 2018 helps examine earlier trends. The researchers use separate measures of tasks’ suitability for AI and Claude usage related to those tasks.

Understanding the method · DijitalPi commentary

AI exposure describes how suitable an occupation’s tasks are for AI or the use associated with those tasks. Occupations are divided into five exposure groups, called quintiles. The researchers compare headcounts across these groups within each age band. Age does not directly measure experience. An occupational measure also does not establish that a particular ADP firm adopted AI.

A hire is counted when a worker appears at a firm where they were absent in the preceding month. A departure is the reverse. Using task classifications from Claude conversations, the authors distinguish automation, where AI performs the work, from complementary use, where it supports the worker. This is not an experiment randomly assigning AI to businesses.

03 / RESEARCH FINDINGS

What changed in young worker headcounts?

Between November 2022 and June 2026, total employment in the ADP sample rose by about 6%, while the most exposed quintile grew by about 4% across all ages. Among workers aged 22-25, employment fell by about 11% in the two most exposed quintiles and rose by about 10% in the three least exposed quintiles. The difference between growth rates is 21 percentage points. Employment in exposed groups stood 19% below the level it would have reached by keeping pace with less exposed peers. This is a relative shortfall, not a percentage-point difference. A separate calculation, the unadjusted occupational regression in Table 1, compares the most exposed quintile with the least exposed using firms observed since January 2018. Its coefficient is -0.179, described by the authors as about 18 percentage points. This coefficient is not the 19% shortfall measure. The employment divergence primarily reflects weaker hiring of young workers; departures did not rise. Declines are concentrated in occupations with high automation shares. Young recruits’ starting base pay shows no obvious relationship with exposure.

Young employment in the most exposed groups
About 11% declineWorkers aged 22-25 in the two most exposed quintiles; headcount change from November 2022 to June 2026.
Young employment in the least exposed groups
About 10% growthThe three least exposed quintiles for the same age group and period.
Difference between growth rates
21 percentage pointsThe direct difference between employment growth rates in more and less exposed groups of young workers.
Shortfall against keeping pace
19%The relative gap below the employment level exposed young workers would have reached by keeping pace with their less exposed peers.

Source: Main findings and employment comparisons ↗

04 / DIJITALPI'S EXPLANATION

How should we interpret these findings?

DijitalPi commentary: When assessing an entry-level vacancy, write down both the tasks to be automated and those through which a recruit will gain experience. Specify the experienced colleague’s review responsibilities separately. The study does not establish that replacing a young employee with a tool would improve a company’s results.

DijitalPi commentary: Marketing, sales and customer service occupations appearing in the exposure classification does not mean the paper provides a separate decline rate for each. Do not turn a quintile result into a staffing reduction target for your team. Internal productivity and cost outcomes require their own measurement.

05 / CONCLUSION AND OPEN QUESTIONS

What did we learn, and what do we still not know?

The text examined is a working paper and contains no information about peer review. The authors say they do not estimate AI’s causal impact. These findings describe observed changes. Controlling for college-graduate share roughly halves the primary sample coefficient and weakens statistical support; in the other firm sample, significance disappears. With college share, interest-rate exposure and suitability for working from home entered together, the Table 1 coefficient becomes -0.080 and is statistically indistinguishable from zero. Occupational trends already differed before ChatGPT. The national ACS survey comparison has a consistent direction, but the magnitude appears specific to ADP. New and small businesses are not well represented; the rates cannot be transferred to Türkiye or a particular small business. The study does not measure internal productivity gains or total staffing cost savings. Base pay excludes bonuses, overtime, commissions, equity and tips. ADP provided the data and supports the Stanford Digital Economy Lab through its Corporate Affiliate Program and a sponsored research project unrelated to this paper. ADP has the right to review the paper to prevent disclosure of confidential information. These data and financial ties are an interest connection relevant to the study. Payroll records are not public. The authors do not predict that the trends will continue indefinitely.

Our open question at DijitalPi: if entry-level hiring is postponed, who will take over those tasks, and how will opportunities for new employees to gain experience be preserved?

DIJITALPI'S APPLICATION COMMENTARY

How can we use this in marketing?

We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.

EXAMPLE 01

What will a new marketing employee do?

An illustrative small manufacturer uses AI for content drafts while considering hiring a marketing assistant.

If drafts are automated, what work will the new employee do?

Open the recommendation for example 01

Separate draft production from opportunities to learn

DijitalPi commentary: List product fact-checking, gathering customer questions and reviewing copy with an experienced colleague as distinct tasks. Specify which tasks will help the new employee gain experience. The study gives no separate employment percentage for young marketers and does not test the productivity of this task allocation.

EXAMPLE 02

Which records should guide sales hiring?

An illustrative wholesaler postpones filling a sales assistant vacancy after introducing AI for quotation preparation.

Does hiring less show that AI is sufficient?

Open the recommendation for example 02

Record why hiring was delayed alongside the task allocation

DijitalPi commentary: Write down why the vacancy was postponed. Record changes in order demand and the responsibilities of the experienced colleague taking over the work separately. Stable headcount does not show that the sales team offers opportunities for new entrants. These records help explain your decision; they cannot establish AI’s employment impact on their own.

EXAMPLE 03

How will customer service experience be passed on?

An illustrative online retailer uses AI drafts for common questions while planning training for a new support employee.

How will the new employee learn to handle customer problems when answers are already drafted?

Open the recommendation for example 03

Define the tasks an experienced colleague will review

DijitalPi commentary: Set out when the new employee should verify a response and consult an experienced colleague. Identify the customer problems they will review together during training. The study did not test this training arrangement. The presence of support occupations in the exposure classification does not establish that this business can provide the same service with fewer people.

TRY IT WITH YOUR TEAM

How can tasks and learning be planned before hiring?

DijitalPi commentary: For a planned entry-level role, we suggest listing its tasks, the work AI will perform and the experienced colleague’s review responsibilities in the same table. Record which tasks will help the recruit gain experience. Assess any reason for postponing hiring alongside this task plan. This proposal has not yet been implemented.

Data Analytics Services

Source and ownership

Erik Brynjolfsson, Bharat Chandar, Ruyu Chen
Erik Brynjolfsson: Stanford University and NBER. Bharat Chandar: Stanford University. Ruyu Chen: Stanford University.

Stanford Digital Economy Lab; August 2026; Updated version, first circulated August 2025; Working paper. Open the original publication · Find on Google Scholar

The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.

← Return to all research articles

ASK AI ABOUT DIJITALPI

What does DijitalPi do?

Opens your chosen assistant with a ready research prompt. It reads the site live and answers.

Message us on WhatsApp