Over the past few months, news about the unemployment tribulations of recent college graduates has been one of the biggest talking points on the economy, amplified by how generative artificial intelligence might be contributing to a hiring slowdown. We now have quantitative evidence showing that AI may be having an impact on entry-level employment for those with four-year degrees.

That the job market was softening (for new graduates and others) isn't in dispute. Until now, however, competing theories were bruited about why college graduates were having difficulty finding jobs. Some argued hiring challenges were an artifact of pandemic-era overhiring and/or economic uncertainty relating to the impact of tariffs. Stanford University economist Erik Brynjolfsson and colleagues' new analysis finds that AI is a likely driver of these employment challenges.

Using ADP data, the paper finds that between 2022 and 2025, young workers in highly AI-exposed occupations -- for example, software programming and customer service -- saw employment growth that was 13 percent lower relative to workers in less-exposed jobs. By contrast, experienced workers in the same roles held steady, and in some cases their employment grew by six to nine percent. Earnings, meanwhile, have remained constant, meaning that AI is reducing employment opportunities but not (yet) affecting wages.

Further, the study finds a clear differentiation between jobs where AI automates human tasks (leading to measurable job losses) and those where it augments worker ability (driving job growth). The most automatable jobs include software development roles with tasks like coding, debugging, and routine maintenance, as well as customer service roles involving scripted responses, routine troubleshooting, or text-based interactions. These are the jobs where human employment is declining, while positions higher up the career ladder continue to grow -- an important reminder that fears of mass job destruction have not yet materialized.

The key difference between the pre-AI career ladders and the current ones seems to be an increased need for what are known as "tacit" (i.e., embedded, nonexplicit, and experiential) skills. Unlike repetitive work that can be coded into AI tools, tacit skills are more insulated from automation. As one tech entrepreneur recently said to me, "AIs are stupid. They don't get the context." Human beings, who are ultra sensitive to subtle changes and variations in social and work environments, are smart in ways AI can't currently replicate.

This human superpower for understanding context and emotional signals connects directly to the argument I made in 2018 that we had gone too far in emphasizing narrower technical training at the expense of broader learning and adaptation skills. In our rush for "relevant" education (i.e., that which connects directly to employment and wages), we have underemphasized the importance of the broader knowledge and interpersonal skills required for getting, keeping and advancing at work. The young graduates who were told to "learn to code" are paying the price for a cultural bias in favor of technical skills in an era where tacit knowledge and skills are increasingly at a premium.

The new normal is one in which students need to "reskill," even before they graduate, by developing a blend of legacy technical knowledge (e.g., principles of computer operation and coding), AI fluency, and noncognitive skills. To strengthen demand for human workers, Brynjolfsson calls for tax reforms that eliminate incentives for automating rather than augmenting human labor. Additionally, there is a clear need for granular and timely data on shifts in skill demands, ideally through an early-warning system that would track and categorize which skills and jobs are most automatable, providing lead time for adaptation by schools and students.

As pointed out in Reason Magazine, the Brynjolfsson research is not without limitations. Using the year 2022 as the base year may be skewing the data due to a post-pandemic readjustment in the tech sector coinciding with AI adoption. Moreover, the use of ADP data over represents firms in the Northeast and those that are growing faster than the national average. The study's methodology also excludes nearly 30 percent of the workforce who don't have recorded job titles. While these limitations don't invalidate the study's findings, they do imply it may not fully represent the wider economy.

Brynjolfsson's study makes it clear that AI is probably disrupting the labor market, eroding entry-level jobs, and making it harder for young workers to gain the tacit experience that protects against automation. If we want to maintain an economy in which career ladders still exist for younger workers, we will need a two-pronged strategy: education and training geared toward adaptive, noncognitive skills and AI fluency aimed at strengthening the supply of resilient workers, and tax policies that reduce pro-automation bias. Combined with timely, granular data that can warn schools, students, and employers about which roles are more at risk, these measures can ensure that young people still have pathways into meaningful work.