Will generative AI change employment, or will it restructure work first?

A preprint working paper by Humlum and Vestergaard linked a representative survey of 11 high-exposure AI occupations in Denmark to administrative records on wages, working hours, workplace, and job history. The 2024 survey included approximately 25,000 individuals and 7,000 workplaces, comparing changes between users and non-users, and between workplaces with and without implementation. Within two years after ChatGPT's public release, the differences in average wages, recorded working hours, and hourly wages were nearly zero, rejecting an average effect exceeding 2%.[1] No differences in employment, total wage bills, hiring, or turnover were detected at the workplace level.

Even when surface indicators remain quiet, work is not static. Eighty-five percent of AI users redirected their saved time to other tasks. In workplaces without proactive employer measures, about 8% of users took on new business tasks, rising to about 17–18% in workplaces with such measures. New business tasks include not only content generation but also quality verification of AI outputs, ethics and compliance, and integration into workflows.[1] Initial organizational impacts may manifest not by reducing headcount but by altering the allocation of work creation versus supervisory work.

• Whether reduced work time was redirected to other customer service or analysis
• Whether new burdens from verifying, explaining, and recording AI outputs have increased
• Whether supervisory or control work has spilled over to non-users
• Whether roles and evaluation criteria have changed despite job titles remaining the same

Hui et al. analyzed monthly employment histories of 79,322 text-based freelancers on Upwork from January 2022 to April 2023. After ChatGPT's public release, in high-exposure occupations related to writing, monthly job counts decreased by 2.0% and monthly income by 5.2% compared to low-exposure occupations. The probability of obtaining at least one job dropped by 1.2 percentage points, and among months when employed, job counts fell by 4.7%.[2]

Another analysis using the release of image generation AI observed a 2.1% decrease in job counts and a 5.2% decrease in income for image-related occupations.[2]

These results do not imply that generative AI reduced employment across the entire economy by the same ratio. Online outsourcing is a market where short-term contracts are easily renewed and buyers can quickly switch to in-house production, AI, or other talent. The study does not observe income outside the platform nor whether individual clients actually adopted AI.[2] Nevertheless, this serves as a warning that substitution pressure may appear faster in outsourcing than in existing employees' wages where adjustment costs are low.

• Work with small outsourcing units, such as writing, image generation, research, and data formatting
• Work with clear deliverable specifications, making it easy for clients to compare with AI
• Work where demand is re-determined per project rather than through continuous employment
• Work where it is difficult to differentiate based on context understanding, responsibility, and trust beyond price

A preprint working paper by Tucker used the Quarterly Workforce Indicators, covering about 95% of U.S. private employment, to compare employment and hiring for ages 22–24 by AI exposure by industry and state. In the highest exposure quintile, regression-adjusted employment fell relatively by about 12% over 10 quarters after ChatGPT's public release, and hiring dropped by about 9% immediately after the release. The difference in wage growth was small, and the main pathway for employment reduction was weak hiring.[3]

However, this study did not directly observe corporate or individual AI use. Information was lost in the process of converting occupational exposure to industry levels, and pre-existing trends related to the COVID period, remote work, and enrollment rates are also visible. Financial policy shocks may explain up to about one-quarter of the employment differences.[3] Therefore, the 12% cannot be called a causal effect of generative AI. A safe interpretation is that changes may appear first at the entry point of youth hiring rather than in existing employees' wage adjustments, suggesting that training pipelines should be monitored separately.

• Track youth hiring numbers and applicant counts separately from total headcount
• Verify whether the disappearance of entry-level work will lead to a shortage of mid-career talent in the future
• Distinguish between problems AI will substitute and the foundational experience needed to supervise AI
• Include alternative explanations such as economic conditions, interest rates, remote work, and enrollment simultaneously

Even if total headcount remains unchanged, future composition differs depending on whether hiring decreases and turnover decreases simultaneously, whether entry-level hiring is halted to compensate with internal mobility, or whether outsourcing is brought in-house. The recommended personnel ledger separates beginning headcount, hiring, turnover, internal mobility, replacement with outsourcing, and new AI supervisory tasks, linking them by the same job family. This is not a new estimate of labor market effects.

This is a hypothetical example. If total headcount is flat but youth hiring decreases and existing employees shift to AI verification, short-term employment stability and a contraction of the training entry point occur simultaneously. Conversely, if outsourcing projects decrease but employee workload increases, it does not necessarily mean a simple disappearance of demand. We observe not just the difference in headcount but who enters through which entry point, through what tasks, and where they move.

The recommended comparison involves dividing time among creation, verification, customer adjustment, and exception judgment within the same job title. We distinguish between cases where only the work changes while the title remains the same and cases involving job mobility. We do not attribute unmeasured time to automation benefits; if business records and self-reports conflict, we suspend judgment. Even when personnel stock is quiet, we continue to update flow and role composition.

Studies addressing existing employees, outsourcing markets, and youth hiring provide material for observing multiple layers that may adjust at different speeds, but they do not prove that generative AI impacts always appear in the same order.[1][2][3] 'Outsourcing and hiring move first' is a hypothesis to be verified in personnel planning and should not be transferred as a definitive general rule to organizations with different countries, contracts, economic conditions, and implementation methods.

As competing hypotheses, we can posit pathways where AI complements existing employees and hiring increases, only short-term outsourcing decreases, hiring is suppressed while internal redeployment advances, or all layers increase due to demand expansion. This is not a prediction but an example of not fixing the same observations into a single narrative. For each hypothesis, we determine the indicators that move first, changes that would falsify it, and external explanations such as economic conditions.

In the recommended review, we do not switch to the most convenient hypothesis each quarter. We fix the observation period and falsification conditions at the start and verify which of hiring, turnover, outsourcing, mobility, or role time moved as expected. If multiple hypotheses remain, we do not determine the cause nor convert it into a headcount reduction rate. We choose reversible hiring and placement decisions, carrying uncertainty into the next period.

• Hypothesis pathways: Separate complementarity, outsourcing substitution, hiring suppression, and demand expansion
• Leading indicators: Specify hiring, outsourcing, mobility, and time that move first for each pathway
• Falsification conditions: Pre-define changes contrary to expectations, no change, and external factors
• Response: Record reversible hiring and placement decisions that can be taken while hypotheses remain

The labor impact of generative AI transmits through four layers: task time, how existing employees work, corporate hiring, and the overall labor market. Even if tasks shorten, if saved time shifts to other work, total working hours do not change. Even if existing employees' time does not change, if outsourcing orders or youth hiring are suppressed, demand changes at the entry point. Mixing layers leads to directly linking productivity gains to headcount reductions or misreading unchanged wages as no impact.

In practice, we track not only labor costs but also hiring, outsourcing, redeployment, AI supervision, and education in the same personnel ledger. If entry-level work is eliminated for short-term efficiency, we also design who will acquire exception judgment and review capabilities. The basic policy consistent with current evidence is not to set a uniform reduction rate across the company, but to observe work redistribution and place early warning indicators on outsourcing and youth hiring, where negative changes are more likely to appear first.

The workforce ledger separates role, business unit, and observation period from decision measures: early-career entry, outsourcing dependence, post-redeployment retention, exception-handling capacity, and AI supervision time. If entry, retention, or exception capacity falls below its minimum, or outsourcing dependence or supervision load exceeds its maximum, AI expansion in that workflow stops for workforce-plan review. Restart requires the HR and workflow owners to jointly confirm recovery of every criterion across two monthly reviews after training and placement changes.

• Task layer: Required time, quality, rework, AI supervision time
• Employee layer: Total working hours, overtime, wages, job mobility
• Corporate layer: Hiring, turnover, outsourcing costs, redeployment, training investment
• Market layer: Job-specific job postings, youth employment, income distribution, new roles