On May 20-21, 2026, Wharton Human-AI Research welcomed scholars, practitioners, and policymakers to campus for its third annual AI and the Future of Work Conference. If the inaugural gathering two years ago framed AI as a transformation on the horizon, this year’s tone made clear the horizon had closed in fast. 

“A year ago, some of those ideas still felt speculative,” said Nancy Rothbard, deputy dean and David Pottruck Professor of Management at the Wharton School during opening remarks, “but today they’re actually shaping how we work, learn, communicate, and make decisions in real time.” 

A self-described childhood science fiction fan, Rothbard noted that the worlds once imagined by Asimov, Heinlein, and Clarke now seemed less like fiction than forecast. But she was careful to keep the focus human, pointing to her own recent research on the tension between technology’s productivity gains and the strain they place on employees — and a phenomenon she and her co-authors call “coddling work,” the increasingly intensive support managers must provide to help people absorb accelerating expectations. “Ultimately the future of work is still about people,” she said.  

When 17x More Code Doesn’t Mean 17x More Software 

This year’s AI and the Future of Work Conference consisted of more than 40 research presentations across two days on Wharton’s campus, tackling a wide range of issues related to AI’s impact on work and research. Leon Musolff, assistant professor of business economics and public policy at Wharton, took on a question lurking behind much of the AI productivity conversation: when developer output explodes, where does all that gain actually go? 

In their paper, Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools, Musolff’s research team studied roughly 100,000 GitHub developers, merged with internal Microsoft telemetry, to trace the productivity effects of successive generations of AI coding tools — from 2022-era autocomplete, to synchronous agents that work alongside a developer in real time, to the newer asynchronous agents that take a written task and return finished code from the cloud. 

The headline numbers were striking, but the story was in their decay. Adopting the full stack of tools multiplied a developer’s lines of code by roughly 17 times. Yet that effect compressed steadily as it moved up the software production pipeline: about four times as many files touched, three times as many commits, and, at the far end, only about a 40 percent increase in actual software releases. 

The bottleneck, Musolff argued, is human review, which does not scale with AI output. “I don’t have time to read all of this stuff that’s being generated,” he said, describing a feeling familiar to anyone collaborating with AI-assisted colleagues. When human and AI inputs are complements, the gains attenuate downstream — AI can help you write far more code without helping you ship proportionally more product. 

Invoking the famous Solow paradox (that the computer age was visible everywhere but in the productivity statistics), Musolff asked whether developer-level gains show up in the aggregate. The answer was a qualified yes: coding activity on GitHub has measurably accelerated since 2025, while downstream software releases have risen more modestly and unevenly across marketplaces. He was candid about the limits of observational, public data, and about the maintenance and code-complexity questions his measures can’t yet capture. 

Building Public Goods in a Day 

Closing the loop between research and practice, Daniel Rock, assistant professor of Operations, Information and Decisions at Wharton, presented results from an experiment run in partnership with the Gates Foundation: the Wharton-Gates Public Goods Build-a-thon. The premise was simple — gather some 30 people of varied expertise from academia, industry, and nonprofits, hand them the conference’s own animating question of how AI is reshaping jobs, and ask them to build something useful in a single day. 

Five teams formed, drawing participants from Arizona State, Columbia University, CUNY, MIT, Microsoft, Jobs for the Future, and Opportunity@Work, among others. The projects ranged from JobShock, which compares job postings against the work actually being done, to a faculty tool for updating syllabi toward greater AI relevance, to a scenario-planning tool built on jobsdata.ai that pairs AI exposure with demand elasticities to model which roles are most at risk.  

The standout came from a team led by Xinlan Emily Hu, a computational social scientist at MIT, with Andreas Haupt, David Holtz, and Prashant Raganathan. Their system, SimulaCrew, builds digital twins of team members to simulate a collaboration before it happens, predicting outcomes, optimizing who should be in the room, and surfacing people who are missing. They field-tested it on the Build-a-thon itself, predicting what each team would produce from nothing more than participant profiles and a seed document of ideas. In one unsettling case, the simulation predicted Emily’s own team would abandon its starting idea and pivot to an “AI hiring audit” tool — closely mirroring a project she was separately, and privately, already working on. “How did it know this?” she asked. “I didn’t even put anything about hiring into the seed information.” 

Rock framed the broader payoff in human terms: pair people who know engineering and “vibe coding” with people who deeply understand a problem domain, give them half a day, and they learn an enormous amount from one another. The most common reaction afterward, he said, was simply: “I didn’t know this was possible in such a short amount of time.” 

Or, as he put it after SimulaCrew’s demo: we’ve all felt that a meeting could have been an email. “Now we can see a meeting could have been a fake meeting.” 

Looking Ahead

Across two days, the conference’s throughline was the one Rothbard named at the outset: the pace of adaptation now demanded of workers, managers, researchers, and institutions alike. The research presented here, some of it still preliminary, some of it built in real time — is part of how Wharton intends to not only study that adaptation, but help shape it. 

To view conference presentations and learn more about the presenters, click here.