· 14 min read · worthmydegree.com
Which careers AI touches, and what a slow first year costs
Something real is happening to entry-level hiring, and it is being reported in a way that invites a much larger conclusion than the evidence supports. This guide separates the two: what has been measured, what it is a measurement of, and which parts of a degree's payoff it touches.
What has actually been measured
The strongest evidence is Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen of the Stanford Digital Economy Lab, which reads ADP payroll records covering millions of American workers from November 2022 through June 2026 (the paper, August 2026). Because it is cited far more often than it is read, here are its six facts, in the authors' order and in our words:
- No evidence of widespread, economy-wide job displacement.
- Employment of 22 to 25 year olds in AI-exposed occupations now sits about 19% below where it would be had it kept pace with their less exposed peers. Experienced workers show no comparable gap.
- That divergence has widened steadily since the authors first documented it in August 2025, when it stood at about 15%.
- It runs mainly through reduced hiring of young workers rather than through increased separations.
- The declines concentrate in occupations where AI usage substitutes for human tasks. Where usage complements the worker, employment is flat or rising, especially for experienced workers.
- The adjustment is showing up in employment rather than in base compensation.
Two of those six matter here more than the headline does. The fourth means the gap opens as what the authors call a narrowing entry point for new workers rather than the displacement of workers already employed. The sixth means pay has held so far: among new hires they find no obvious relationship between AI exposure and real starting pay for young workers. What moved is how many were hired, not what the hired ones were paid.
The authors are careful about what the six facts are, and the retellings usually are not. They call these early descriptive indicators rather than causal estimates. The divergence survives excluding technology firms and computer occupations, controlling for exposure to interest-rate increases and for remote work, and swapping in alternative measures of AI exposure. It also attenuates under an education control, shows some divergent trends predating generative AI, and is more pronounced in the ADP sample than in national survey benchmarks.
The wider literature is not unanimous, and one of the paper's own authors is the one who says so. Writing a survey of the field, Bharat Chandar lists four studies of Current Population Survey data, his own among them, that find at most small changes in hiring in AI-exposed jobs, and a Danish study by Humlum and Vestergaard that reaches contradictory results from worker surveys matched to firm data. He also notes that the sample sizes for particular combinations of age and occupation are small (Stanford Digital Economy Lab, October 2025). The entry-level finding does have support from outside the ADP records, from Hosseini and Lichtinger and from Klein Teeselink using Revelio Labs data on American and British firms. Read all of it as an argument in progress rather than a settled result.
That education control deserves its own line in a guide about degrees. When the authors adjust for the share of each occupation held by college graduates, the estimate for the most exposed quintile of 22 to 25 year olds falls from about 18 points to about 9. They publish both and say the pair brackets a range, because occupational college share may be measuring exposure to precisely the codified knowledge generative AI substitutes for, or may be measuring separate shocks to educated labor markets. Quoting the larger figure on its own is quoting half a result.
The mood matches the data. Handshake's outlook for the class of 2026 reports that 61% of graduating seniors describe themselves as somewhat or very pessimistic about their careers, with almost half of those naming generative AI as at least part of the reason, while the share of full-time job postings on the platform that mention generative AI has grown nearly fivefold since 2023 (Handshake). Employer demand for the skill and student anxiety about it are rising together.
The comparison that study is not making
Read the Stanford finding closely and it compares young workers in AI-exposed occupations against young workers in less exposed occupations. It does not compare people who went to college against people who did not. That distinction decides what the number can be used for: it says some fields became harder to enter, and it does not say a degree stopped paying, because the people on both sides of it mostly hold one.
The question this site asks is a different one. Every figure in the calculator measures a career against a debt-free high school graduate who never enrolled, over ten years, after loan payments and taxes. Nothing in the entry-level evidence answers that comparison, in either direction.
There is a second gap between that study and this calculator, and it runs the useful way. The calculator prices what a career pays, and pay is not where the adjustment has shown up so far. So the finding leaves the salary figures here standing. What it puts pressure on is the assumption sitting underneath them, which is that the job starts on schedule.
Where the exposure sits in the careers priced here
The calculator carries an optional AI exposure band for every career, taken from published exposure research and held at the level of the occupation group rather than the individual job. The research itself runs finer than that, down to individual tasks in the Department of Labor's database, and the app deliberately does not: a unique score for one job title would imply a precision the underlying estimates do not carry. Those estimates have since been checked against what people actually do with AI, twice. Tomlinson and colleagues found the Eloundou measures correlate closely with Microsoft Copilot usage during 2025, and the Stanford paper reports that sorting employment by Claude usage matches sorting it by the Eloundou measures (both reported here). Of the 293 careers in this app that typically need a bachelor's degree or more, here is how they fall:
| AI exposure band | Careers | Typical pay |
|---|---|---|
| High | 36 | ~$81,000 |
| Medium | 190 | ~$96,000 |
| Low | 67 | ~$115,000 |
Note: Figures computed from this site's own datasets are rounded and marked with a tilde, in this table and in the others on this page.
The high band is narrower than the conversation suggests, and it is specific: 28 of those 36 careers are business and financial operations, six are legal, two are office support. The low band is mostly hands-on health care, 49 of its 67 careers, plus 15 in community and social service.
Two warnings about that table, because it is the kind of table people screenshot. The bands measure how much of a job's task content overlaps with what current AI tools do, which is not the same as the odds of losing the job. And the pay column describes what these careers pay today, not what the exposure will do to them. Low exposure and high pay landing in the same row is a fact about the present, not a strategy.
The bands, group by group
The score is a property of the occupation group rather than the job, so the groups are the honest unit to list. These are the ten groups with at least five careers at this level, ordered by exposure:
| Occupation group | Exposure | Careers | Median pay | Examples |
|---|---|---|---|---|
| Business & Financial Operations | High (80) | 28 | ~$81,000 | Accountants and Auditors, Financial Risk Specialists |
| Legal | High (80) | 5 | ~$118,000 | Lawyers, Judges |
| Computer & Mathematical | Medium (55) | 19 | ~$106,000 | Software Developers, Data Scientists |
| Arts, Design, Entertainment, Sports & Media | Medium (55) | 21 | ~$75,000 | Graphic Designers, Technical Writers |
| Management | Medium (50) | 28 | ~$130,000 | Chief Executives, Marketing Managers |
| Architecture & Engineering | Medium (45) | 22 | ~$111,000 | Civil Engineers, Mechanical Engineers |
| Life, Physical & Social Science | Medium (45) | 38 | ~$95,000 | Economists, Physicists |
| Educational Instruction & Library | Medium (45) | 59 | ~$78,000 | Elementary School Teachers, Law Teachers |
| Healthcare Practitioners & Technical | Low (30) | 47 | ~$160,000 | Registered Nurses, Pediatric Surgeons |
| Community & Social Service | Low (20) | 15 | ~$60,000 | Healthcare Social Workers, Clergy |
Five smaller groups are left out, eight careers between them. One of those omissions is worth naming: office and administrative support is the highest scoring group in the file at 85, and the two careers it contributes at this level, proofreaders and copy markers and statistical assistants, both pay about $50,000. Exposure and pay are not aligned, inside a band or across the three of them.
A group median also hides a great deal. Healthcare practitioners runs from about $59,000 for exercise physiologists to about $559,000 for pediatric surgeons, every one of them carrying the same score of 30, and the education group runs from about $42,000 for short-term substitute teachers to about $129,000 for postsecondary law teachers. The band tells you something about the kind of work. It tells you nothing about which job inside that work you would hold.
None of this is a forecast, and the Stanford paper is the reason to say so twice. A band describes how much of a group's task content overlaps with what current AI tools do. Overlap on its own does not fix the direction, because the same paper finds employment falling where AI usage substitutes for the work and flat or rising where it complements the work, and a single score cannot tell you which of those you are looking at. That distinction is measurable, and the Stanford authors did not invent it: they take it from the Anthropic Economic Index, which reads millions of Claude conversations against the Department of Labor's own task database and estimates, occupation by occupation, how much of the usage substitutes for the work against how much assists it, splitting 43% to 57% across all usage (Handa et al.). That index covers one company's platform rather than every use of AI, and the table above does not carry it yet. Adding it is the honest next step, because until a second column exists a band reports how much overlap a group has and not which way the overlap runs.
The most exposed work does not need a degree
Everything above covers careers that need a bachelor's degree or more, because that is what this site prices. Widen it to all 825 occupations in the federal file and the shape changes: 59 of the 94 careers in the High band need no bachelor's at all, and 52 of those 59 are office and administrative support, the highest scoring group in the file.
Counting careers understates it, because clerical work is enormous. Weighted by the number of people actually employed:
| Entry education | Jobs | Share in High-exposure groups |
|---|---|---|
| High school diploma | ~54.7M | 30% |
| Bachelor's degree | ~39.0M | 26% |
| No formal credential | ~37.4M | 1% |
| Associate's degree | ~3.3M | 16% |
More high-school-entry jobs sit in the most exposed groups than bachelor's-entry jobs, as a share and in absolute numbers both, roughly 16 million against 10 million. The largest of them are Customer Service Representatives at about 2.6 million jobs and a median near $45,000, Office Clerks at 2.5 million, Secretaries and Administrative Assistants at 1.7 million, and Receptionists at 0.9 million. The Stanford paper names customer service and clerical work in the same breath as software development, so it is pointing at the same place.
This matters more here than it would almost anywhere else, because of who the calculator compares you against. Every premium on this site is measured against a debt-free high school graduate who never enrolled. If the most exposed work in the economy is disproportionately where high school graduates are, then exposure is not something sitting on the degree side of that comparison. It is on both sides, and possibly more on the side without the degree.
Three limits on how far that can be pushed, all of which cut against the convenient reading. The baseline in this model is not built from those clerical occupations: it is an all-occupations median for high school graduates with an age curve on top, so what happens to receptionists does not mechanically move it. Exposure remains task overlap rather than job loss, and that rule does not get suspended because the conclusion flatters the degree. And the Stanford authors' own education control points the other way, since adjusting for how many graduates an occupation employs roughly halves their estimate, which suggests part of what looks like an AI effect is a shock to educated labor markets in particular.
What survives all three is narrower and still worth having. The evidence does not show AI exposure landing on degree holders and sparing everyone else, and a reader deciding whether a degree is worth its price should not assume the alternative is untouched.
What a slow first year costs
Across the 177 careers that typically need a bachelor's degree, the model starts a graduate at a median of about $65,000 and credits them with roughly $740,000 of gross earnings over the following ten years. A missing first year is about 9% of that, which sounds survivable and is the wrong denominator.
Gross earnings are not what this site reports. It reports the premium: what the career earns over ten years minus what a debt-free high school graduate earns across the same ten years, which comes to about $429,000 at those ages. The median premium for a bachelor's-level career is roughly $311,000, so the missing $65,000 is about a fifth of it. Half of these careers land between 16% and 27%.
The asymmetry is what makes it bite. Both sides of the comparison are assumed to work every year, so a year spent searching comes off the career's side and nothing comes off the high school graduate's. For five of these careers, one lost year is enough on its own to turn a positive ten-year premium negative, before any loan is priced. For most of the rest it moves the year the two lines cross rather than which line ends higher, and the calculator will tell you which case you are in.
The Methodology footer states that working assumption in full rather than burying it, next to the related one about underemployment: roughly four in ten graduates work in jobs that do not require a degree at all.
Doing something with it
Price the career rather than the credential. The spread inside a single education level is far wider than the gap between levels, so "is a bachelor's worth it" is a question with no answer while "is this career worth this school's price" has a precise one.
Check the band on the career you have in mind. It sits under Advanced Analysis Settings, off by default, and it names the occupation group and what the research says about it rather than producing a score for your specific job title.
Then move the window. If you expect the first year to be slow, the calculator will show you the year the career passes the high school graduate and stays past them, which is the number a weak start actually moves.
Run your own numbers with the school, the career and the loan you are actually considering. The entry-level market is one input to that arithmetic. It is not the arithmetic.