Research

Employers often rely on outside recruiters to find workers, but it is unclear whether human intermediaries add value when employers already have access to algorithmic screening tools. We study a randomized experiment in a large online labor market that assigned human recruiting assistance to job postings. Treated employers received 16% more applications and conducted 35% more interviews but were no more likely to hire than control employers with access only to algorithmic tools. Match quality declined: treated employers spent less on hires and their workers completed fewer hours. We develop a model of delegated recruiting in which recruiters and employers rely on a common noisy signal, which can generate these patterns when their assessments are highly correlated. Consistent with this mechanism, recruited applicants are positively selected on engagement-related characteristics yet do not improve hiring outcomes. These findings suggest that as algorithmic screening improves, the scope for intermediaries to add value shrinks because it becomes harder to access independent information.

In a large randomized experiment, hiring consultants generated more applications and interviews but did not improve hire rates, compared to employers who only used algorithmic screening tools.

Reductions in private search costs due to advances in information technology can improve market efficiency. Although, changes in private search costs can change behaviors, making the welfare implications unclear if that behavior creates negative externalities, as was the case here. We consider the market efficiency effects of the introduction of an AI tool into a labor market. In order to lower their search costs, potential employers were randomly offered AI-written first drafts of their job post. The assistance was widely accepted and treated employers were 19% more likely to post a job; those posting spent 44% less time writing. Despite the substantial increase in job posts, there was no discernible increase in matches. The lack of match formation was mostly due to marginal jobs being posted by employers with lower intent. Up to a fifth of the missing matches were caused by direct impacts on the job posts—in the sense that they were more generic and less informative to jobseekers. This combination of increased congestion and degradation in informativeness wasted jobseeker time as jobseekers applied to jobs they otherwise would not have. Quantifying this waste, the per job post loss to jobseeker welfare is six times larger than the increase to employer welfare from time saving. These negative efficiency outcomes persist after the close to market wide adoption of the technology, showing the reductions in private search costs in this context harmed market efficiency.

AI assisted job posting increased the number of jobs posted but not the number of hires, as lower writing costs flooded the market with low-intent, generic listings that wasted jobseeker time — with losses to jobseekers six times larger than gains to employers.

• Financial Times: The AI Shift: Is hiring becoming less meritocratic?
• New York Times: The Long Term Employed Today: College-grads
• MIT Initiative on the Digital Economy: Using GenAI to Write Job Descriptions? Put Some Effort Into It

Motivated by the potential for large productivity gains from AI, firms are increasingly deploying agentic AI systems capable of independent action. Some firms have also begun formally integrating these agents into their organizational structures—assigning them designated roles and responsibilities, and in some cases explicitly referring to them as employees. This creates new challenges for decisions like when to delegate and how to monitor work. In a survey of 1,261 managers we find that 23% of managers already work in organizations where AI agents have been formally institutionalized on organizational charts. In a randomized experiment we provide those managers with identical documents containing built in errors, where we vary whether the drafts are presented as produced by an AI tool, an AI employee, or a human employee. Average effects to error catching are small. However, in the subgroup of managers whose organizations already have `AI employees’, presenting identical drafts as produced by an AI employee (versus an AI tool) reduces managers’ monitoring intensity by 16%, increases their reliance on additional review from others, and shifts their perceived accountability away from themselves and toward the AI system. The human employee condition shows that this is not simply a response to delegation, in fact, managers do the most careful oversight when they’re told the work came from their human employee. These results suggest that embedding AI agents into formal organizational roles can reduce managerial oversight in AI-mediated work, and should be understood as a governance decision rather than a mere labeling choice.

More companies have 'AI employees' than you think and they occupy a unique position in organizations: treated as delegated producers rather than tools, but not monitored like human employees.

• Wall Street Journal
• New York Times

If and how to regulate online marketplaces is an open question important to both platform designers and policy makers. Using a large field experiment in an online labor market, we analyze the effects of a platform minimum wage. Workers were randomly assigned individual price floors which prevented treated workers from bidding hourly rates below their floor. Workers for whom the floor was likely binding—those historically bidding below the floor—suffered a decline in job-finding probability (30%), but higher wages conditional upon being hired (9%). Treated workers made lower earnings overall, but higher earnings conditional on working at least one hour on the platform. Despite a job being “worth more” if hired, affected workers lowered their search intensity. They did not move to the “uncovered sector”—jobs with a fixed price rather than an hourly wage, nor did they direct their search to better fitting jobs. They were also more likely to exit the platform. After the conclusion of the experiment, the platform rolled out the $3 per hour minimum wage platform wide, allowing us to observe the employment outcomes and job search behavior in equilibrium.

Individual-level wage floors made workers on an online labor market less likely to get hired and more likely to exit the platform entirely.

Generative AI can raise productivity, but realized gains depend on how workers use the tool. We study AI competence as a form of human capital defined by the practical ability to organize work with a model, evaluate its output, and retain judgment while using it to extend one’s existing skills. In a preregistered lab-in-the-field experiment, 332 full-time management consultants whose jobs did not routinely require coding were assigned a difficult Python data-analysis task with or without access to a Gen AI tool. AI access increased scores by 34 percentage points, raised completion by 7 percentage points, reduced time on the task by 12%, and improved debugging efficiency. We open the black box of these gains by combining Gen AI transcripts, code-execution logs, task outputs, and surveys. Using an ethnography-inspired agentic coding procedure, we identify nine distinct AI-collaboration practices that capture how workers frame requests, decompose problems, rely on generated code, edit independently, verify outputs, and manage scope. These practices explain substantial heterogeneity in overall performance. Gains concentrate when workers engage proactively with AI while retaining substantive judgment. Finally, we show that productivity gains and workers’ interpretation of those gains are distinct. AI access does not significantly increase average confidence, trust, enjoyment, or behavioral trust in AI, and belief responses vary substantially by gender and prior coding experience. The results suggest that firms should treat AI training as human-capital investment, teaching workers not only how to prompt, but how to divide labor with AI, evaluate its output, and build confidence in using it well.

As organizations experiment with placing AI agents on org charts as “employees,” new research shows this framing has unintended consequences. In a large-scale experiment, anthropomorphizing AI reduced individual accountability, increased unnecessary escalation, lowered review quality, and heightened employee uncertainty about their roles—without improving adoption. The findings suggest the core challenge is not whether to deploy agentic AI, but how to redesign workflows, roles, and governance so humans remain clearly accountable while effectively supervising increasingly capable systems.

There is a strong association between writing quality in resumes for new labor market entrants and whether they are ultimately hired. We show this relationship is, at least partially, causal: in a field experiment in an online labor market with nearly half a million jobseekers, treated jobseekers received algorithmic writing assistance on their resumes. Treated jobseekers were hired 8% more often. Contrary to concerns that the assistance takes away a valuable signal, we find no evidence that employers were less satisfied. We present a model where better writing does not signal ability but helps employers ascertain ability, rationalizing our findings.

Pre-GenAI algorithmic writing tools like Grammarly make it more likely jobseekers get hired, with no negative impacts on employers, by improving the signal of jobseekers underlying ability.

• Yahoo News: AI can help job seekers get noticed and hired, study finds
• Market Watch: Looking for a new job? Brush up your résumé with a computer algorithm.
• Insider: Writing your résumé with AI makes it easier to find a job

“Upskilling” often refers to the process by which workers acquire and expand their skills, enabling them to perform different types of work as market demands change. This paper demonstrates that while generative artificial intelligence (GenAI) can act as an “exoskeleton,” enhancing workers’ capabilities while they attempt new skills, these gains are dependent on the continued use of the technology. When the “exoskeleton” is removed, little to no knowledge is retained independently, revealing that the newfound capabilities are temporary and reliant on the external support provided by GenAI. We run a randomized controlled trial on “reskilling” with GenAI by providing Boston Consulting Group (BCG) consultants with access and training in using ChatGPT to solve technical problems. We measure their performance on real data science tasks outside their skill sets, which cannot be independently solved by ChatGPT. Treated workers score 49, 20, and 18 percentage points higher than those in the control group on the three tasks and perform close to the level of real BCG data scientists on two of the three tasks. However, treated workers are no better at answering technical questions without the use of ChatGPT post-experiment, suggesting their demonstrated newfound technical capabilities do not imply knowledge acquisition.

In an experiment, non technical consultants given GenAI performed almost as well as actual data scientists on data science tasks, but these gains did not last.

The boundary discontinuity method of causal inference may yield misleading results if a policy’s impacts do not stop at the border of the implementing jurisdiction. We use geographically precise longitudinal employment data documenting worker job-to-job mobility to study policy spillovers in the context of three local minimum wage increases. Estimated spillover impacts on wages and hours are statistically significant, geographically diffuse, and sufficient to create concern regarding interpretation of results even using not-immediately-adjacent regions as controls. Spillover effects appear less concerning with smaller interventions or those adopted in a smaller jurisdiction.

Local minimum wage studies that use nearby areas as controls may understate the true effects, because the policy itself raises wages in those 'control' areas through worker mobility.

Seattle raised its minimum wage to as much as $11 in 2015 and as much to $13 in 2016. We use Washington State administrative data to conduct two complementary analyses of its impact. Relative to outlying regions of the state identified by the synthetic control method, aggregate employment at wages less than twice the original minimum, measured by total hours worked, declined. A portion of this reduction reflects jobs transitioning to wages above the threshold; the aggregate analysis likely overstates employment effects. Longitudinal analysis of individual Seattle workers matched to counterparts in outlying regions reveals no change in the probability of continued employment, but significant reductions in hours particularly for less-experienced workers. Job turnover declined, as did hiring of new workers into low-wage jobs. Analyses suggest aggregate employment elasticities in the range of -0.2 to -2.0, concentrated on the intensive margin in the short run and largest among inexperienced workers.

Seattles minimum wage increases reduced hours worked in low-wage jobs — especially for less-experienced workers — more than they reduced employment outright.

• The Economist
• FiveThirtyEight
• Los Angeles Times
• New York Times
• New York Times (The Upshot)
• Seattle Times
• Washington Post

We study the effects of a large increase in Seattle’s minimum wage on business churn, hours, and revenue using Washington State administrative data. We find the minimum wage affected businesses both at the intensive and extensive margins. At the intensive margin, surviving businesses increased labor costs without decreasing hours and saw no reductions in revenue. At the extensive margin, businesses experienced higher rates of exit and newly opened businesses became less labor-intensive. We find the total effect of the minimum wage to low-wage employment, defined as jobs paying 130% of the minimum wage or less, came from changes to the composition of businesses.