Why can’t anyone agree on whether AI is taking our jobs? The research has been mixed, which can largely be explained by several factors: inverted AI exposure for high or low skilled jobs, and compositional shifts in hiring rather than volumetric shifts. Basically, it’s a mixed bag depending on the type of job, the timeline, and the specific research question. From AI exposure to adoption rates, wage growth or task shifting, researchers are trying to understand the nuances of how artificial intelligence is influencing labor economics and job market trends.

What academic research says about AI and employment

The question of the day, “will AI take our jobs?”, is being tackled by many researchers around the world (del Rio-Chanona et al., 2025; Salari et al., 2025), and the answers are still being uncovered. Results have been mixed in assessing the impact of AI on the job market, as the question is tackled in different ways. Controlled experiments tend to find large effects. Across randomized trials and field experiments, productivity gains run roughly 15 to 60 percent depending on the setting. But when researchers look at national administrative records instead of individual tasks, the effects mostly vanish. A study linking Danish payroll data to survey-reported AI use found essentially zero impact on earnings or hours worked.

These discrepancies highlight the heterogeneity of the question; yet, several trends remain consistent, which may shed light onto why results differ depending on the context. One is that AI exposure has inverted. Where the automation literature of the 2010s pointed at routine, lower-wage work, measures of generative AI exposure now converge on high-wage, white-collar occupations: the people who write, analyze, and advise. Second, results point towards a composition shift in hiring rather than a volume shift, meaning that the change is showing up less in jobs disappearing than in who gets hired into them. Evidence is accumulating that demand is softening specifically for entry-level and novice roles, the rungs at the bottom of the ladder.

These findings showcase the nuanced answers to a question that is more complex than headlines want to make it out to be, and are not yet represented in aggregate employment statistics. And that is the methodological problem at the center of this literature, which is why much of the research in this space relies on alternative data sources.

Alternative datasets provide valuable insights into the economic impacts of AI and employment research for several reasons:

• Real-life, retrospective, observational data allows you to study unexpected event-related impacts, like the release of ChatGPT and other innovations in generative AI.
• Combining multiple sources of data allows you to examine a complex network of influences, impacts, and relationships.
• Granularity lets you see effects that occur within firms and occupations rather than across them, which is where the current evidence suggests most of the change is happening.
• Timeliness matters when the phenomenon is only a few years old. Official statistics arrive quarterly or annually, while this literature is turning over much more quickly, often in real time.

The job postings data below shows what that looks like in practice.

What job posting data reveals about AI and hiring demand

How are researchers tackling this question today? We recently hosted a seminar on the topic to find out. The LinkUp team walked through how their Job Records dataset is built, and a research team from Georgia State University presented their working paper that applies it (Khosravi & Liu, 2026). Their question: do advanced and developing economies respond the same way, and does the response show up in how much firms hire or in what they hire for?

Why did these researchers find this dataset useful? LinkUp indexes postings directly from employer career sites rather than job boards, with history back to 2007. Because postings are timestamped daily and coded to standard occupations, researchers can measure what a specific firm was hiring for before an event and compare it to what they hire for after.

“One thing I really like about this data set is that it directly collects the job posting from company career pages rather than job boards.” - Faezeh Khosravi

There are two ways generative AI could show up in hiring data. Firms could hire fewer people, or they could hire different people. To separate the two, Khosravi and Liu analyze millions of job postings from thousands of firm-country pairs in hundreds of countries, comparing the years before and after ChatGPT's release. Their measure of a firm's AI exposure comes from the occupations it was hiring for in 2019, well before ChatGPT existed.

“Because this measure is based entirely on pre-ChatGPT hiring patterns, it is fixed before the treatment and is not affected by firms’ subsequent hiring responses" -- Faezeh Khosravi

They found that firms with more AI-exposed hiring profiles moved away from the roles AI can most easily do, and that the gap widened across the three years after ChatGPT's release rather than closing. That points to a change in strategy rather than a reaction to a new tool, as Khosravi describes alongside the event-study figures in the clip below:

‍

The compositional shift showed up across both advanced and developing economies, but overall hiring volume moved less consistently.

"Generative AI is changing what firms hire for more consistently than how much they hire." - Faezeh Khosravi

What made this study possible was granular, timestamped data with consistent occupation coding over a long enough history. While other Dewey researchers are also using LinkUp to examine AI impacts on job market trends (Dong et al., 2025, Shi et al., 2025); there are also several other alternative datasets on the Dewey Platform that can be applied to this emerging research question.

Limitations of Job Posting Data

Job ads represent only one slice of the labor market, and have scope limitations within that. Often the data itself can be messy, containing duplicates, dead or filled or ghost ads, or biased representation of the whole job market. LinkUp’s methodology solves several of these issues by sourcing directly from employer websites rather than job boards, but this limits the scope to larger, tech-forward companies that have their own career opportunities of job board pages. 

Additionally, job ads offer insight into hiring and recruitment, as well as shifts in skill and educational requirements as they relate to salary ranges; but things like job retention, layoff trends, actual compensation, and employee satisfaction are just a few other aspects of the labor market that job ads alone fail to capture. This is another reason why many researchers interested in labor economics are turning to alternative datasets in order to capture different layers and perspectives of how artificial intelligence is influencing the job market.

How alternative datasets on the Dewey Platform open new doors to academic researchers 

No single dataset answers this question, and researchers are utilizing many alternative data sources to evaluate answers from new angles. From job ads to paychecks to pink slips, there are several valuable datasets on the Dewey Platform that can contribute to this research area. 

LinkUp: Job Records, Skills, and AI Skills Tag. 

Postings collected directly from employer career sites back to 2007, with the full description, occupation and sector codes, company identifier, and ticker on each record. Two companion datasets do the text work for you: Skills extracts one row per posting per skill, and AI Skills Tag flags US postings that require AI skills, both joinable to Job Records on the hash field. Extracted Salary adds stated USD compensation for postings active since January 2019. Researchers who prefer to build their own classifier still can, which is the approach Shi et al. (2025) take in developing an LLM-based method for identifying AI jobs.

Lightcast: US and Global Core Labor Market Information. 

The standardized frame: occupation and industry employment, earnings, and demographics mapped to consistent geographies, plus tables that are unusually well-suited to this question. Occupation Hires and Separations reports turnover by occupation and geography, Unemployment by Occupation does the same for joblessness, and Staffing Patterns gives the occupational composition of each industry. Job Postings Skills covers extracted skills for US and global postings. Useful for asking whether occupations flagged as AI-exposed are actually seeing turnover move.

WageScape: Job Postings with Salary. 

US job postings with compensation attached at the posting level, alongside role mapping, titles, and how long each posting stayed open. Advertised pay is a different measurement than realized earnings and captures what firms are willing to offer rather than what they ended up paying. Note the roughly one-year lag, which makes this better for establishing wage levels than for tracking a fast-moving shock.

WARN Database: WARN Layoff Data. 

Mass layoff notices required under the WARN Act, standardized across states through direct pulls and FOIA requests, with company name, state, notice date, effective date, temporary status, and employee count. Layoffs appear here when announced rather than months later in aggregate statistics. Coverage is uneven because the underlying state systems are: Wyoming and Arkansas do not publish notices, and several states report monthly rather than by notice date. A separate Airline Employment Data series is also included.

BrightQuery: employment and occupational wage data. 

Built from IRS, SEC, Department of Labor, and state registry filings rather than surveys, covering public and private companies on the same schema. Employment Data is available monthly, quarterly, and annually at the company level, and Occupational Wage Data adds wage statistics by occupational category. Because records are point-in-time, they reflect how a company stood at each filing date rather than as later revised. Veridion extends firm-level context internationally through a stratified sample of its global company universe.

People Data Labs: Company Insights and workforce dynamics. 

Company-level headcount, growth, churn, and tenure, updated monthly. The detail that matters here is that several of these break out by seniority level: Employee Count by Month by Level and Average Tenure by Level let you watch the shape of a company's career ladder change over time rather than just its total headcount. That makes it one of the few sources on the platform where the entry-level question can be asked directly instead of inferred from posting counts. Individual-level resume and profile data is also available, though under a separate access tier.

These datasets provide new points of entry into evaluating AI impacts as they occur. Several can specifically speak to the hiring composition question from different angles: LinkUp through what firms advertise, People Data Labs through who actually holds junior roles, and WARN through who gets cut. Triangulating across them is closer to an answer than any one alone. That’s the power of alternative data, and the value of a Dewey subscription that includes them all.

Resources

• Explore the Dewey Platform: Alternative Data for Academic Research
• Upcoming seminar: How Alternative Data is Shaping Research Today
• Previous LinkUp Seminar: Is AI already reshaping hiring? What LinkUp Job Postings Data reveals
• LinkUp’s Extracted Salary dataset - unique data attribute article

References

• Khosravi, F. and Liu, E. M., "Generative AI and Firm Hiring Demand: Evidence from Advanced and Developing Economies," April 2026 draft, SSRN 6537038
• Shi, Hanwen and Gupta, Anil and Ding, Waverly W. and Zhang, Kunpeng, Identifying AI Jobs in Online Job Postings: Existing Versus Newly Developed LLM-Based Approaches and A Dataset on Demand for AI Jobs among Russell 1000 Firms, 2018–2025 (November 23, 2025). http://dx.doi.org/10.2139/ssrn.5842102
• Dong, F., Doukas, J. A., & Zhang, R. G. (2025). Strategic technology talent acquisition and firm value: A cross-industry examination. Journal of Financial Research, 1–34. https://doi.org/10.1111/jfir.70038
• del Rio-Chanona, R. M., Ernst, E., Merola, R., Samaan, D., & Teutloff, O. (2025). AI and jobs: A review of theory, estimates, and evidence. arXiv:2509.15265. https://doi.org/10.48550/arXiv.2509.15265
• Salari, N., Beiromvand, M., Hosseinian-Far, A., Habibi, J., Babajani, F., & Mohammadi, M. (2025). Impacts of generative artificial intelligence on the future of labor market: A systematic review. Computers in Human Behavior Reports, 18, 100652. https://doi.org/10.1016/j.chbr.2025.100652
• Humlum, A., & Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56. https://doi.org/10.2139/ssrn.5219933 (circulating at NBER as w33777 under the revised title Still Waters, Rapid Currents)
• Brynjolfsson, E., Chandar, B., & Chen, R. (2025, revised 2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/

‍