GenAI Adoption Reduces Social-Skill Demand Within Roles

The standard prediction for generative AI at work is a compositional one: if AI takes over cognitive tasks, employers should redesign jobs around interpersonal coordination, persuasion, and other social capabilities. This study finds a different pattern in U.S. public-company hiring. Across job postings from firms that adopted GenAI between 2022 and 2024, social-skill demand falls relative to comparable non-adopting roles after ChatGPT’s launch, while cognitive-skill demand does not show the expected decline.

That result matters because it challenges an increasingly common inference from task-automation arguments to skill demand. A role can absorb a tool without becoming more socially intensive. The authors’ account is that GenAI enables self-service: employees outside administrative functions take on some functional administrative work themselves, adding those skills to postings rather than shifting toward more cross-functional interaction.

Core Contribution

The paper’s contribution is a skill-centric framework for separating several possible consequences of GenAI adoption within an occupation or role. Rather than treating “cognitive” and “social” skills as a fixed trade-off, it asks how the mix of requested skills changes when firms introduce GenAI. That distinction is consequential: a stable cognitive-skill measure alongside lower social-skill intensity has a different implication from broad cognitive substitution.

The empirical setting covers roughly seven million job postings from 595 U.S. public firms. The analysis compares GenAI-adopting and non-adopting roles around the launch of ChatGPT, using difference-in-differences estimates. The central outcome is skill intensity, with a complementary analysis of the number of skills listed. This design is aimed at within-role changes in advertised demand, not at whether particular workers used GenAI or became more productive.

Technical Approach

The authors classify the skills named in postings into cognitive, social, and functional administrative categories, then track their relative prominence over time. Their identification strategy asks whether the change after ChatGPT’s launch is different for roles associated with GenAI adoption than for other roles, while adjusting for contemporaneous changes in the comparison group.

The mechanism test is more informative than a simple before-and-after trend. The paper traces the social-skill result to the addition of functional administrative skills in roles outside the administrative functions where those skills would ordinarily be concentrated. In other words, the declining social-skill share need not mean employers stopped valuing communication; it can arise because postings acquired a larger administrative component. The paper reports that this pattern is concentrated in non-managerial roles, which is consistent with employees using GenAI to complete support tasks directly rather than relying on intensified managerial coordination.

Figure 4 presents the difference-in-differences comparisons for skill intensity before and after ChatGPT’s release.

Results and Analysis

The headline estimate is a 3.4% relative decline in demand for social skills in GenAI-adopting roles. The paper finds no corresponding evidence that cognitive-skill demand increased or decreased relative to the comparison roles. That combination is the paper’s strongest corrective to the social-skills-reallocation story: the data do not support the claim that GenAI-adopting jobs broadly compensate for automated cognition by requesting more social capability.

The descriptive evidence and time-series estimates are useful checks on the interpretation. Figure 2 breaks GenAI-related posting patterns out by occupation and industry, while Figure 5 separates changes in skill intensity from changes in skill counts. Separating those measures matters because a lower share of social skills could reflect additions elsewhere rather than removals of social requirements. The authors’ administrative-skill explanation addresses precisely that issue.

Figure 5 reports the time-series estimates for both skill intensity and skill count.

The evidence supports a narrow conclusion: among the sampled firms and advertised roles, adoption coincides with a modest but directionally clear reweighting away from social skills. It does not establish that GenAI caused every observed change, that workers’ actual tasks changed in the same way as postings, or that the result generalizes to private firms and labor markets outside the United States. Still, the scale of the posting data and the explicit comparison design make this a more useful test than claims derived from task taxonomies alone.