Three CMU researchers join Google’s AI economy cohort to study hiring and automation

Three Carnegie Mellon University researchers have joined Google’s inaugural AI & Economy cohort, the university said October 14. Their projects examine different parts of AI’s economic impact: what happens when employers use systems with similar hiring errors, how to measure benefits beyond workplace productivity, and why some tasks resist being split between people and machines.

The eight-member international cohort gives each researcher $100,000 in unrestricted funding, dedicated computing resources and early access to frontier AI systems. Its remit covers labor markets and organizational productivity. CMU’s participants include two professors from its Heinz College of Information Systems and Public Policy and one from its College of Engineering.

Brian Jabarian, an assistant professor of economics and technology at Heinz College, is examining AI in hiring. His work includes a field experiment with PSG Global Solutions, which handles outsourced recruitment. The team randomly assigned roughly 70,000 job applicants to one of three groups: interviews with human recruiters, interviews with an AI voice agent, or a choice between the two.

Applicants offered that choice favored the AI interviewer. Jabarian’s cohort project turns to a different question: what happens if employers rely on AI systems that make similar mistakes? The focus is on errors shared across hiring systems, rather than applicants’ willingness to speak with an automated recruiter. That is a research question, not a finding that such shared errors have occurred.

78%Applicants choosing the AI interviewer

CMU says this share applies to applicants offered a choice of interviewer—not the full experimental sample.

Avinash Collis, an associate professor of digital economy at Heinz College, is looking beyond hiring and worker output. He studies benefits that conventional economic statistics, including gross domestic product, may miss. His argument is that workplace productivity captures only part of AI’s value, particularly when people use tools without paying for them or use them outside their jobs.

Collis is developing metrics to quantify those user benefits and give policymakers a fuller account of AI’s costs and benefits. He is especially interested in how the balance changes with income and geography, including differences between wealthy and poor countries. Where Jabarian’s project examines a risk in employer decisions, Collis’s work asks how to count benefits that users receive beyond measured workplace output.

Christophe Combemale, an assistant research professor of engineering and public policy, is studying the division of labor itself. Rather than treating an entire occupation as vulnerable or resistant to AI, he examines the tasks within it. His preliminary research suggests that some tasks can be economically separated between people and AI, allowing partial automation. Others are harder to pull apart.

• Information can bind tasks together. A lawyer following a business dispute may need to track how it develops across many communications. Combemale identifies that information structure as a potential cost of separating the work.
• Physical requirements can bind tasks together too. A technician’s manual operation may require vision, judgment, dexterity and strength at the same time, rather than as independent steps that can be assigned separately.

Under those constraints, Combemale says, performing one task adequately is not enough: AI must perform the linked tasks simultaneously to deliver a productivity gain. He hopes the research will help employers assess AI’s productivity in specialized sectors.