AI Skills Improve Job Prospects: Causal Evidence from a Hiring Experiment

The growing adoption of artificial intelligence (AI) technologies has heightened interest in the labor market value of AI-related skills, yet causal evidence on their role in hiring decisions remains scarce. This study examines whether AI skills serve as a positive hiring signal and whether they can offset conventional disadvantages such as older age or lower formal education. We conducted an experimental survey with 1,725 recruiters from the United Kingdom, the United States and Germany. Using a paired conjoint design, recruiters evaluated hypothetical candidates represented by synthetically designed résumés. Across three occupations of graphic design, office assistance, and software engineering, AI skills significantly increase interview invitation probabilities by approximately 8 to 15 percentage points, compared with candidates without such skills. AI credentials, such as university or company-backed skill certificates, only lead to a moderate increase in invitation probabilities compared with self-declaration of AI skills. AI skills also partially or fully offset disadvantages related to age and lower education, with effects strongest for office assistants, for whom formal AI certificates play a significant additional compensatory role. Effects are weaker for graphic designers, consistent with more skeptical recruiter attitudes toward AI in creative work. Finally, recruiters’ own background and AI usage significantly moderate these effects. Overall, the findings demonstrate that AI skills function as a powerful hiring signal and can mitigate traditional labor market disadvantages, with implications for workers’ skill acquisition strategies and firms’ recruitment practices.

Introduction. The demand for AI skills is growing. In major labor markets such as the US or the UK demand for AI skills has approximately tripled between the mid-2010s and the mid-2020s (Stanford HAI, 2025). As firms across industries are increasingly integrating AI systems into production, coordination, and decision-making processes, they are seeking talent skilled with AI, since unlocking AI's potential depends less on the technology itself than on the human skills needed to leverage it (Brynjolfsson et al., 2018). Consistent with basic economic theory, rising demand has translated into increasing salaries. A growing body of empirical work documents substantial wage premia associated with AI skills, with estimates suggesting that workers listing AI-related competencies command significantly higher salaries than otherwise comparable workers (Bone et al., 2025; Stephany and Teutloff, 2024).

Despite growing evidence that AI skills are associated with higher wages and improved labor market outcomes, most existing studies rely on observational data. These studies typically infer the value of AI skills from correlations in job postings, résumés, or wage data. While informative, such approaches face well-known limitations. Workers who possess AI skills may differ systematically from those who do not along unobserved dimensions, such as motivation, access to elite educational institutions (Alekseeva et al., 2021), or workforce adaptability. Indeed, AI proficiency might simply proxy for multiskilling – a trait long theorized to reduce resistance to technological change (Carmichael and MacLeod, 1993) and drive firm productivity (Kim and Park, 2003). Similarly, firms that demand AI skills may differ in ways that independently affect wages and hiring practices (Garcia-Lazaro 2025). As a result, it remains difficult to isolate whether AI skills themselves causally influence labor market outcomes or merely proxy for these other valued characteristics.

This paper provides causal evidence on the economic value of AI skills by investigating whether AI skills improve hiring prospects. We conduct a conjoint survey experiment with 1,725 professionals who have hiring experience in the UK and the US, collecting 22,195 CV comparison tasks. We present participants with pairs of fictitious CVs and ask them to choose which candidate they would invite for an interview. Each participant evaluated candidates for one of three roles: office assistant, graphic designer, or software engineer. These roles span administrative, creative, and technical work, each requiring different types of AI competencies ranging from general AI literacy to specialized technical skills. The CVs were programmatically generated to systematically vary the presence of five AI skill treatments: no AI skills mentioned, self-reported AI skills, company-certified AI skills, professional certification, and university certification. We also varied candidate characteristics including age and educational attainment to examine whether AI skills can offset traditional hiring disadvantages, such as the "technological obsolescence" stereotype often applied to older workers (Hudomiet and Willis, 2022). This experimental approach is inspired by previous literature in labor economics that use discrete choice experiments and vignette studies to elicit employer preferences (e.g., Humburg & van der Velden, 2015; Piopiunik, Schwerdt, Simon, & Woessmann, 2020).

Our experimental approach allows us to address three questions. First, do AI skills on resumes increase the probability of being invited to an interview, and does this effect differ across occupational contexts? Second, inspired by recent work on skills versus degrees (Bone et al., 2025), we investigate whether AI skills can offset traditional hiring disadvantages such as lower formal education or being older. Third, drawing on job market signaling theory (Spence, 1973) and recent work on "muddled information", that is disorganized, confusing, or ambiguous data where key details are mixed with irrelevant factors, leading to a loss of clarity. As AI might make CVs less informative of a candidate’s actual ability (Frankel and Kartik, 2019), we examine whether more credible signals of AI competence carry greater weight in hiring decisions. If employers view AI skills with uncertainty, costly and verifiable credentials should provide stronger signals than self-reported claims.

Our study is one of the first to provide causal evidence for the growing literature on the labor market value of AI skills1. The diffusion of generative AI has expanded demand for workers who can develop, deploy, and effectively use AI systems (Bone et al., 2025; Stanford HAI, 2025; OECD). For instance, Teutloff et al. (2025) show how ChatGPT's release shifted demand on freelancing platforms toward AI-complementary skills, illustrating that new technologies create opportunities for workers whose capabilities complement them.

Related work. Our second contribution speaks directly to signaling theory. While Spence (1973) originally focused on a single dimension of ability, modern theory must account for the "muddling" of signals in the generative AI era. Frankel and Kartik (2019) propose a "Muddled Information" framework, suggesting that as signal manipulation becomes cheaper, signals become less informative about natural ability and more indicative of "gaming ability" (AI proficiency). This shift drives the labor market toward a "cheap talk" environment (Galdin and Silbert, 2025), where the ease of generating polished materials – that may appear to attribute the associated skills to their creators – makes the skills themselves a noisy signal that fails to separate high- and low-ability workers. Indeed, Cui et al. (2025) provide evidence that AI acts as a substitute for pre-existing writing ability rather than a complement, effectively narrowing the gap between workers of different initial skills. This concern is closely related to debates about the proliferation of low-quality, AI-generated content – often described as “AI slop” or “work slop” – that mimics competence without substantive effort (Niederhoffer et al., 2025). While some workers genuinely augment their capabilities through AI, others may use it as a substitute for effort, producing lower-quality output (del Rio-Chanona et al., 2025). Our study shows that, despite these concerns, AI skills function as a strong market signal even when self-reported. However, these returns are not uniform; echoing Firpo et al.’s (2025) finding that AI signal value is highly context-dependent, we observe that certification matters most for conventionally disadvantaged candidates, while skepticism toward AI is most pronounced among recruiters hiring for creative roles, where sentiment toward AI skills is markedly more negative.

Method. To causally isolate the effect of AI skills and their interactions with candidate attributes on hiring prospects, we design a paired-comparison conjoint experiment (Figure 1). This methodology allows us to estimate the causal influence of specific attributes that are often confounded in observational data (Piopiunik et al., 2020; Humburg & van der Velden, 2015). The design achieves three key objectives: it mimics the real-world comparative process recruiters use when screening candidates, it allows for clean identification of causal effects by constructing pairs that differ only on manipulated attributes, and the forced-choice format circumvents respondent-specific scale-use bias while reducing social desirability bias through revealed preference (Hainmueller et al. 2014).

The experimental design was structured around three key factors, which were systematically varied across the CV pairs.

Job Profile. The CVs were created for three distinct occupations representing a range of skill requirements and output verifiability: Office Assistant (administrative), Graphic Designer (creative), and Software Engineer (technical). For each role, we varied the contract type between a 6-month fixed-term contract and a permanent (open-ended) contract, allowing us to test whether employers' time horizons affect their valuation of AI signals.

Candidate Attributes and Disadvantages. Each CV was built around a core profile defined by gender, age, and education level. In a large subset of cases, one candidate (Candidate B) was presented with a potential hiring disadvantage relative to the other (Candidate A). These disadvantages were operationalized as: (a) Age: an older candidate (approximately 60 years old) paired with a younger one (approximately 32 years old); and (b) Education: a candidate with a lower level of education (e.g., an Associate's Degree) paired with a candidate with a higher level (e.g., a Bachelor's Degree).

AI Skill Treatment. To test our main assumptions, the disadvantaged candidate (or one of the candidates if no disadvantage occurred in the pair) was randomly assigned one of five AI skill treatments: (1) No AI skill (control), where the CV contained no mention of AI skills; (2) Self-reported AI skill, where the candidate listed AI skills in their skills section with supporting language in their job descriptions; (3) LinkedIn certification, where the candidate held an AI skill certificate from LinkedIn Learning; (4) University credential, where the candidate possessed an AI skill micro-credential from a university (e.g., "Online Certificate in Generative AI for Business Professionals, Oxford University"); or (5) Company credential, where the candidate had an AI skill certification from a major technology firm (e.g., "Generative AI for Business Professionals, IBM - Coursera certificate").

The selection of Office Assistant, Graphic Designer, and Software Engineer is grounded in a rigorous tripartite framework of administrative, creative, and technical labor, effectively capturing the diverse impacts of the cognitive industrial revolution where all three roles exhibit 100% exposure to LLMs yet face distinct displacement or augmentation trajectories (Eloundou et al., 2024; Felten et al., 2021). For the Office Assistant, a role historically susceptible to automation-driven displacement, the specific selection of productivity-enhancing skills such as AI-powered scheduling, automated reporting, and automating workflows with Zapier is justified by the urgent need for these workers to transition from routine information processing to human-in-the-loop coordination, thereby signaling enhanced efficiency to counteract projected employment stagnation (Bureau of Labor Statistics, 2024).

Discussion. This finding has profound implications for labor market efficiency. It suggests that the AI skill premium documented in this study might be conditional on the digital literacy of the hiring gatekeeper. Firms whose recruiters do not themselves use generative AI may be systematically undervaluing AI competencies in candidates, creating a friction between the nominal demand for AI skills and the actual ability of hiring processes to reward them.

Beyond the binary metric of adoption, the perception of AI’s utility – captured through the sentiment analysis detailed in Annex C – offers a critical explanatory layer for the observed premium. While recruiter usage of AI tools is positively correlated with the likelihood of selecting AI-skilled candidates, the qualitative evidence suggests this relationship is driven by an underlying belief in the technology's value proposition. As shown in Table C1, recruiters who expressed positive sentiment toward AI were significantly more likely to prioritize AI literacy, with a positive interaction effect of 5.2 percentage points (Model 2). Crucially, this "enthusiasm effect" remains robust (4.8 percentage points, Model 3) even when controlling for the recruiter's own frequency of AI usage, indicating that the premium is not merely a function of technical familiarity, but is shaped by the specific value they ascribe to the technology.

This perception of value is not uniform, however; it is mediated by the specific demands of the occupation (Table C2). The sentiment-driven premium is most pronounced for administrative roles, where positive recruiter sentiment increases the probability of selecting an AI-skilled Office Assistant for an interview by 7.1 percentage points (Model 2). The effect is lower for Software Engineers (3.5 percentage points) and Graphic Designers (4.1 percentage points). This heterogeneity implies that for administrative tasks, AI is perceived as an unambiguous efficiency multiplier, automating routine friction. Conversely, for code-related and creative occupations, the lower translation of sentiment into interview preference suggests a more complex valuation, where AI is seen as a tool requiring careful oversight rather than a direct substitute for human expertise.

However, the valuation of these skills is heavily contingent on the "gatekeeper effect" we identified: a recruiter’s own familiarity with AI dictates how they value a candidate’s AI skills. Recruiters who utilize Generative AI in their daily workflows assign a significantly higher premium to these skills compared to non-users. This heterogeneity implies that the labor market integration of AI is currently asymmetric; as AI tools become more ubiquitous among hiring managers, the penalty for lacking these skills may increase, potentially deepening the divide between the AI-literate and the AI-illiterate workforce.

Conclusion. This study bridges the gap between the surging demand for AI skills and the scarcity of causal evidence regarding their value in hiring. By deploying a paired conjoint experiment with 1,725 recruiters in the UK and US, we isolate the causal premium of AI credentials from confounding factors typical of observational data. Our findings confirm that AI skills act as a powerful signal of employability, increasing the probability of receiving an interview invitation by approximately 8 to 15 percentage points across diverse occupations. This effect persists regardless of whether the skills are self-taught or formally certified, suggesting that in the current labor market, the demonstrated ability to use Generative AI is valued as a proxy for adaptability and potential productivity, independent of the prestige of the credentialing institution.

Beyond the aggregate premium, our results highlight the potential of AI upskilling to act as an equalizer in the labor market. We document a significant substitution effect where AI credentials partially or fully offset the penalties associated with lower formal education and older age. For older candidates, possessing AI skills effectively neutralizes ageism, signaling that the applicant is technically current and adaptable. Similarly, for candidates without university degrees, AI proficiency serves as a compensatory mechanism, allowing them to compete with university-educated peers. This suggests that AI adoption could facilitate social mobility and extend the economic relevance of the workforce in an aging society, provided that access to training is equitable.

Theoretically, our findings contribute to the economics of education and signaling theory by distinguishing between the signaling value of traditional degrees and emerging technical competencies. While traditional education signals long-term persistence and general cognitive ability, AI skills appear to signal immediate readiness for technological disruption. This distinction is crucial for understanding how labor markets adjust to rapid technological change.

Limitations. Finally, while our experimental design allows for causal identification, it is not without limitations. First, our results rely on stated preferences in a hypothetical hiring scenario; while conjoint experiments are proven predictors of real-world behavior, actual hiring decisions may involve additional constraints. Second, our analysis is limited to three specific white-collar occupations, and the magnitude of the AI premium likely varies across other sectors. Future research should examine whether the hiring premium translates into wage premiums and actual on-the-job productivity gains, as well as how the signaling value of AI skills evolves as these technologies move from novel differentiators to standard job requirements.