The New York Fed’s 2024 data flips a ranking that made CS the safe credential. What changed is the part of the job that got cheap.

In February, the Federal Reserve Bank of New York updated its labor-market data for recent college graduates. Among 22-to-27-year-olds with a bachelor’s degree, computer science now shows 7.0% unemployment; computer engineering, 7.8%. Philosophy sits at 5.1%. The rate across all majors is 4.2%.

If you saw a version of this chart last year, the CS number was 6.1% and philosophy was 3.2%. That came from the 2023 data vintage. The current figures use 2024 American Community Survey data. The gap between the two majors narrowed from 2.9 points to 1.9. The inverted ranking held.

The by-major table lags by design: it reflects 2024 survey data, the most recent available. The Fed’s monthly series runs closer to now. As of June 2026, unemployment among recent graduates stands at 5.7%, against 4.1% for college graduates overall, a gap that has held all year.

The computer science degree kept working. It wins on pay and on the share of graduates landing jobs that use the degree, and the table shows by how much.

Federal Reserve Bank of New York, “Outcomes by Major,” 2024 ACS data, released February 2026. Figures cover bachelor’s holders ages 22–27.

The claim survives only in narrow form. On the single metric of finding any job at all, a major with almost no vocational pathway is now outperforming the one built as the direct pipeline into software. That narrowness is what makes it worth reading. A broad claim about the humanities coming back would be easy to dismiss; a narrow anomaly in one metric is a hiring signal, and it appeared during the two years generative coding tools went from demonstration to default.

What Got Cheap, and What Stayed Expensive

Producing code that meets a specification got dramatically cheaper. Producing code that meets the right specification did not. Deciding whether the resulting system does what the business needed got harder, because there is now much more output to evaluate and much less time spent producing each unit of it.

Marco Argenti, the Goldman Sachs CIO, put the constraint precisely in a 2024 essay. AI can produce high-quality code, he grants, and then names the limit: “it can work well, but not do what you want it to do.”

That sentence describes a shipping problem, not a benchmark result. Code can be syntactically correct, semantically correct, well-tested, and still solve the wrong problem. Cheap implementation leaves that risk intact and multiplies the surface area it covers.

The scarce skill was never typing. It was knowing what to build, stating it precisely enough to be checked, and telling afterward whether the thing built satisfied the statement. Implementation got cheaper first; those three operations kept their price. They are also the ones a philosophy curriculum happens to drill: define terms, expose hidden premises, follow an argument to where it fails.

The labor market has produced a signal nobody was optimizing for.

Three Shifts Engineering Leaders Now Own

Your junior pipeline is being repriced. In many teams, the entry-level function used to be producing volume under supervision: the junior wrote code, the senior reviewed it. That trade has collapsed on both sides. The volume is cheap now, and the review burden went up. If your junior roles still describe volume production, you are hiring for a job the tools already do. The version that holds is the one where the junior owns a small, bounded problem end to end, including deciding whether it is solved. That is a harder role to write and a harder one to fill.

Evaluation graduated from phase to line item. Most engineering organizations still budget evaluation as a fraction of build cost. When that cost falls by half and output doubles, the fraction produces a shortfall in both directions at once. Evaluation is where the scarce judgment gets applied, and it is the first thing squeezed when velocity metrics improve. If your build time dropped and your evaluation spend did not rise, you have moved risk downstream, where it costs more.

Someone already owns goal definition. When implementation was the bottleneck, the specification arrived from product and engineering executed it, and the boundary was legible. Now the boundary has moved and mostly nobody was told. Increasingly, engineers make product decisions inside prompts. Product managers, in turn, make architectural decisions inside acceptance criteria. Both happen fast, informally, and without review, because the artifact that once exposed them, a slow and expensive implementation, is gone.

The question to take to your next staffing conversation is not whether to hire philosophy majors. It is narrower: who in your organization is accountable for stating what “correct” means for a given system, and does that person have the standing to say a shipped feature failed?

If you cannot name them, the market has already told you what it costs to leave that seat empty. It is telling you in a chart about undergraduate majors, an odd place to read a hiring signal, and a real one.

Where the Numbers Stop

They do not say philosophy is a better degree. The early-career pay gap is large and durable, philosophy’s underemployment rate runs well above the rate across all majors, and both figures are genuine warnings to anyone choosing a major primarily on these signals.

They do not even fully say what they appear to say about employment. Philosophy sends an unusually high share of its graduates on to further study, and someone enrolled full-time is not counted as unemployed in this series. Some of the employment gap is a schooling gap. The same caution applies in reverse to CS, where graduates confident of eventually landing a suitable role may keep searching rather than take work below their training. The search counts as unemployment; the job would have counted as underemployment.

They do not establish causation. The Fed measures outcomes, not employer reasoning, and a window that contains a tooling revolution also contains a tech-sector correction, a hiring freeze, and a surge in CS graduates hitting the market at once. The Fed’s own researchers put a number on a fifth: remote work, they estimate, accounts for 64% of the recent rise in unemployment among young college graduates, and the timing, they argue, points to distance rather than to generative tools. That finding constrains this one. Software work is among the most remotable there is, so distance may be part of what is pressing on CS, and anyone reading this chart as proof of AI displacement is reading past the evidence.

What the numbers do say is that the safest technical credential in the labor market lost its position on one metric, in the specific window when the work it credentialed became partly automatable, and that the major on the other side of the inversion is one built around defining terms and testing arguments.

As a theory of education, that is thin. As a description of the entry-level engineering job, it is sturdy: the work changed, and hiring processes still describe the job it replaced.