Social Mobility and Class Dynamics

In the spring of 2024, two researchers at the University of Washington sat down to run an experiment that any hiring manager might have assumed would come back clean. Kyra Wilson and Aylin Caliskan took the kind of large language models that companies now routinely bolt onto their recruiting software, fed them more than five hundred real résumés spread across nine occupations, and asked them to do the most ordinary task in modern hiring: rank the candidates. The only thing they changed was the names at the top of each résumé. They swapped in 120 first names that Americans reliably associate with white or Black men and women — the same qualifications, the same experience, the same words, a different name.

The results were not subtle. Across roughly three million comparisons, the models preferred résumés with white-associated names 85.1 percent of the time and Black-associated names just 8.6 percent (Wilson & Caliskan, AIES 2024). Résumés carrying men's names were favored 51.9 percent of the time; women's names, 11.1 percent. And in the single starkest finding of the study, when a résumé with a Black man's name went head-to-head against an otherwise identical résumé with a white man's name, the Black candidate essentially never won. The models preferred the white male candidate close to 100 percent of the time. Zero percent is not a tendency or a lean. It is a wall.

This is the documented core of what this chapter is about. The technology that was sold as the cure for human prejudice in hiring — objective, tireless, blind to the things humans can't help noticing — has been measured, repeatedly and rigorously, reproducing the oldest prejudices we have, and doing it faster and more quietly than any human ever could. To understand why, and to understand what it means for the possibility of climbing from a hard start to a better life, we have to follow the machine backward to what it learned from, and forward to what it decides.

What the evidence actually shows

The Wilson and Caliskan study is the most cited, but it is not alone, and the convergence of findings matters more than any single number. When researchers affiliated with Brookings examined intersectional bias in language-model résumé screening, they found the same directional pattern: white names advantaged, Black names penalized, and the disadvantage compounding when race and gender stacked — Black men and, in some tests, Black women faring worst of all (Brookings, 2024). A separate line of work found that people who watch an AI system make biased recommendations tend to absorb and mirror that bias in their own subsequent judgments, meaning the damage does not stop at the algorithm; it trains the humans standing next to it (University of Washington, 2025).

Comparison in the UW audit Rate the favored group was preferred
White-associated names (overall) 85.1%
Black-associated names (overall) 8.6%
Men's names (overall) 51.9%
Women's names (overall) 11.1%
White male vs. Black male (head-to-head) ~100% white male preferred

Two things are worth holding onto here. First, the magnitude. A gap of a few percentage points would be a fairness problem worth fixing; a preference that runs to 85 percent one way and a head-to-head result of effectively 100–0 is not a bug at the margin. It is the machine's central tendency. Second, the mechanism the researchers uncovered when they tried to fix it: stripping the names off the résumés did not solve the problem, because subtler signals — the wording of a bullet point, the name of a school, the phrasing of an achievement — still leaked identity back to the model. You cannot easily blind a system that has learned to read between the lines.

Why the machine learns to discriminate

The reason is almost banal, which is what makes it dangerous. An AI hiring tool is a pattern-matcher trained on history. Companies feed it examples of the people they hired and promoted in the past, and it learns to recognize the shape of a "successful candidate." If a firm spent decades hiring mostly white male graduates of a handful of universities — not out of stated policy but through the ordinary accretion of referrals, unconscious preference, and institutional habit — then the pattern the model extracts is precisely that shape. It does not know it is discriminating. It has no concept of race or fairness. It is optimizing for resemblance to a biased past, and it executes that optimization thousands of times a day.

graph LR
  A[Historically biased<br/>hiring decisions] --> B[Training data encodes<br/>who was hired before]
  B --> C[Model learns 'successful<br/>candidate' pattern]
  C --> D[Résumés scored for<br/>resemblance to pattern]
  D --> E[Under-represented<br/>candidates filtered out<br/>before human review]
  E --> A

The diagram closes into a loop for a reason. Every cohort the biased system filters becomes the next cohort of "successful hires," which becomes the next batch of training data, which sharpens the same pattern. Human prejudice, however corrosive, was at least inconsistent — one recruiter's blind spot was another's cause. An algorithm applies the same learned bias to every applicant, at scale, with a consistency no bigoted human could match. It is the difference between a leak and a pipe.

There is an old name for the analog version of this. When banks drew red lines around neighborhoods and refused loans to the people inside them, the pretext was actuarial neutrality — they were only following the data. Algorithmic hiring is redlining rebuilt in code, with the added feature that no one involved has to intend it or even know it is happening. The candidate gets an automated rejection and assumes she wasn't good enough. The company sees a diverse applicant pool at the top of the funnel and a homogeneous one at the bottom, and concludes the market simply delivered fewer qualified minority candidates. The bias is laundered through the machine until it looks like merit.

The gap that opens before anyone applies

By the time an algorithm rejects a résumé, much of the damage has already been done upstream, in the schools that decide who ever assembles a competitive résumé in the first place. Here the evidence is newer but pointing in one clear direction. In England, the Sutton Trust surveyed more than 10,000 teachers in April 2025 and found an AI divide opening between private and state schools that mirrors every other advantage gap in education. Teachers at private schools were more than twice as likely to have received formal training in AI tools than their state-school counterparts — 45 percent against 21 percent — and the informal gap was wider still, 77 percent against 45 percent. In schools rated "outstanding," teachers were more than three times as likely to have had formal AI training as those in schools rated "requires improvement" or "inadequate" (35 percent versus 11 percent) (Sutton Trust, 2025).

The number that matters is not how much students use AI but how well. A well-resourced school with trained staff and a considered policy teaches AI as a thinking instrument — something to interrogate, to check, to argue with. An under-resourced school, frightened of cheating and short on training, tends to do one of two things, both bad: ban the tools outright, so students never develop fluency, or permit them unsupervised, so the tools become a way to avoid thinking rather than a way to sharpen it. The result is a divergence in exactly the skill the labor market is starting to price most heavily. The wealthy student learns to work with the machine. The poor student learns to hide from it or to lean on it, and either way arrives at the hiring funnel less prepared for a world that now assumes fluency.

This is the front end of the same pipeline whose back end the résumé screeners guard. The student who never gained AI fluency writes a weaker résumé, applies to a system tuned to reward fluency's signals, and gets filtered — a disadvantage that began in a classroom compounding into a disadvantage sealed by an algorithm.

A ladder that was already rotting

It would be a mistake to blame AI for the decline of social mobility, because the decline is decades old and well documented. The landmark measurement comes from Raj Chetty and colleagues, who tracked the odds that an American child would grow up to earn more than their parents. For children born in 1940, the answer was 92 percent — near-universal upward mobility, the statistical spine of the American Dream. For children born in 1980, the figure had fallen to roughly 50 percent, a coin flip (Chetty et al., "The Fading American Dream," Science, 2016). Their counterfactual simulations found that rising inequality, not slowing growth, was the main driver: when they gave the 1980 cohort the income distribution of 1940, most of the mobility gap closed.

So AI arrives not at the start of this story but well into its third act. The ladder was already missing rungs before the first résumé screener booted up. The honest question — and this is where the book has to be careful to separate what we know from what we reasonably fear — is whether AI is an independent new problem or an accelerant of the old one. The available evidence points to accelerant. The mechanisms AI introduces (algorithmic filtering, unequal fluency, entry-level displacement) all operate on the same axis as the inequality Chetty identified: they concentrate advantage among those who already have it. That is not a new force so much as a powerful new servant of an old one.

The vanishing first rung

Of all the mechanisms in this chapter, the disappearance of entry-level work may be the most consequential, because entry-level jobs are not merely jobs — they are the mechanism by which a person without a pedigree earns the experience that leads to everything after. And the entry level is precisely where AI is biting first.

The clearest evidence comes from a 2025 Stanford study by Erik Brynjolfsson and colleagues, bluntly titled "Canaries in the Coal Mine," which analyzed ADP payroll records for millions of workers. Since the widespread arrival of generative AI, workers aged 22 to 25 in the most AI-exposed occupations experienced a 16 percent relative decline in employment, even as employment for workers over 30 in the very same firms and occupations held steady or grew by 6 to 12 percent (Stanford Digital Economy Lab, 2025). Young software developers were hit hardest, with employment for 22-to-25-year-olds falling nearly 20 percent while their older colleagues' numbers rose. Crucially, the declines concentrated in roles where AI automates the work rather than assisting it; where AI augments a junior worker's tasks, entry-level employment held up.

Read that pattern against the mobility question and its meaning sharpens. The junior analyst role, the entry coding job, the first rung — these existed because someone had to do the routine, learnable work while they absorbed the tacit knowledge of a profession. AI is very good at exactly that routine, learnable work. So the firm keeps its seniors, who supply judgment, and quietly stops hiring the juniors, who used to supply cheap labor and, in exchange, receive an education and a career. Remove the first rung and you have not made the ladder harder to climb. You have made it impossible to reach. The person born into advantage still has the family network, the internship secured through a parent's colleague, the buffer to work unpaid for a season. The person without those things has lost the one door that historically did not require them.

Skills, credentials, and the illusion of the fair test

The favored answer to all this, in policy circles and among the hiring-software vendors themselves, is skill-based hiring: stop looking at where someone went to school or who they know, and look only at what they can demonstrably do. It sounds like the meritocratic correction the moment demands. In practice it tends to rebuild the same hierarchy through a side door.

The problem is that skills do not grow in a vacuum; they grow in conditions. The candidate who spent four university years building a portfolio of projects and the candidate who spent those same years working forty-hour weeks to keep the lights on are not equally positioned to "demonstrate" competence, and the difference between them is not effort. It is the presence or absence of free time, mentorship, tools, and the security to fail at something without going hungry. A skill-based evaluation sees only the output — the polished portfolio, the certification, the side project — and reads prior advantage as present merit. The algorithm cannot perceive potential; it can only perceive evidence of development, and evidence of development is itself a luxury good.

Credentialing compounds the trap. As the bachelor's degree loses its monopoly on professional doors, it is being replaced not by open competition but by a shifting thicket of micro-credentials, certifications, and portfolio conventions whose value is legible mainly to those already inside the relevant professional networks. Knowing which credential actually carries weight — and which is worthless paper sold to the desperate — is itself a form of inherited knowledge, passed through alumni networks and family connections distributed sharply along class lines. The AI screener, tuned to recognize the credential signals its training data associates with success, filters out the candidate who chose the wrong certificate as surely as it filters out the one who had none. The game's rules keep changing, and the people most likely to know the current rules are the people who least needed the help.

Geography as destiny

Class is not the only axis AI is hardening. Location increasingly decides outcomes too, and the two reinforce each other. AI investment, venture capital, and the dense informal networks where opportunities actually circulate are concentrating in a handful of metros — the Bay Area, Seattle, Boston, New York — a dynamic this book examines in detail in Chapter 1.4.2. Proximity to those ecosystems buys the things that never appear in a job posting: the internship, the chance hallway conversation, the tacit knowledge of an industry that no one has written down.

The intuitive fix is to move toward the opportunity. But relocation is only neutral for people who can afford it. It costs money for deposits and moving trucks; it costs social capital, the friend who helps you find an apartment and the contact who makes an introduction; and it costs the support network — family, community, the people who watch your kids when the shift runs late — that provides the thin margin of stability a low-income worker cannot replace on arrival. For the professional-class worker with savings and connections, moving is a decision. For the worker without them, it is a cliff. So a talented young person in rural Mississippi and an equally talented one in Palo Alto do not compete on equal terms, and when both applications reach the same screener, the algorithm favors the Palo Alto candidate — not because it weighs geography directly, but because geographic advantage has already expressed itself through every proxy the model does weigh: school quality, credential type, the shape of the extracurricular profile, the network embedded in the résumé.

How the disadvantages compound

The temptation, faced with all this, is to treat each mechanism as a discrete problem with a discrete fix — audit the algorithm, fund the schools, subsidize relocation, reform the credentials. But the defining feature of the system is that the mechanisms do not operate independently. They chain. A child in an under-resourced school never gains AI fluency; she therefore builds a weaker portfolio; she applies from a disadvantaged zip code to a screener trained to reward the fluency and signals she lacks; the entry-level role that might once have let her prove herself has been automated away; and the credential that might have compensated was one she couldn't afford and wouldn't have known to choose. No single link in that chain is dramatic. Each is a modest, plausible, deniable disadvantage. But they multiply rather than add, and a person carrying several of them at once — poor, Black, rural, first-generation — faces not the sum of those penalties but their product.

This is why the emerging structure is best understood as a stratification, not a spectrum. At the top, an AI-fluent elite: people who grew up with access, learned to wield the tools in well-resourced schools, and can afford both the latest models and the credentials that certify their command of them. They use AI to amplify already-high productivity and earnings, pulling further ahead. In the middle, a broad AI-adjacent population with partial access and partial skills — adapting, holding on, but not advancing, their jobs fragmenting under automation and their wages flat. And at the bottom, a growing AI-excluded group without reliable access, fluency, or the economic margin to acquire either — filtered by the screeners, displaced from the entry-level roles that were their foothold, and priced out of the credentials that might let them compete. What distinguishes these strata is not effort or talent. It is position at the moment the technology arrived.

The window for correction

There is a temporal warning buried in all of this. Systems like these tend to harden. Each biased hiring cycle feeds the next model; each cohort excluded from AI fluency raises less AI-fluent children; each metro that captures AI investment becomes more attractive to the next round of it. The disadvantages are not just compounding across mechanisms at a moment in time — they are compounding across time itself, which means the cost of correction rises the longer it is deferred. There is a plausible point, though we cannot mark it precisely, past which the class barriers AI is encoding become structural enough that no technical fix — no better audit, no debiased model — can dislodge them, and only broad policy intervention could. Whether we are approaching that point or have already crossed it is genuinely unknown, and the honesty of that "unknown" is important: it is the difference between a problem we can still design our way out of and one we will have to legislate our way out of.

What could actually be done

None of this is technologically inevitable, and it is worth being concrete about the levers, along with their limits.

The most direct is regulation of algorithmic hiring, and the law is beginning to move. In the United States, Mobley v. Workday has become the first major legal test: in May 2025, a federal judge granted conditional certification of a nationwide collective action under the Age Discrimination in Employment Act, allowing job applicants over 40 who were rejected through Workday's AI screening since 2020 to join a disparate-impact claim (Holland & Knight, 2025). The court dismissed the intentional-discrimination theory but let the disparate-impact theory proceed — a meaningful signal that "the algorithm did it" is not a defense. On the statutory side, New York City's Local Law 144 now requires employers using automated hiring tools to commission an independent bias audit annually, publish its results, and notify candidates in advance; the European Union's AI Act classifies recruitment and selection systems as "high-risk," subjecting them to conformity assessments as its provisions phase in through 2026 and 2027.

The limits are as instructive as the rules. A 2026 audit by New York City's own Comptroller found enforcement of Local Law 144 largely ineffective — the mandate existed, but the oversight to make it bite did not (DLA Piper, 2026). Transparency requirements create the possibility of accountability without guaranteeing it. Mandatory bias auditing only helps if the audits are rigorous, independent, and enforced, and if "fairness" is defined in a way that closes rather than launders the disparity. Human-review mandates — requiring a person to look at candidates an algorithm filtered — can catch what the machine misses, but only if the human is not simply mirroring the machine's bias, which the research suggests is exactly what tends to happen.

On the education side, the necessary condition is treating AI fluency as a public good rather than a private luxury. That means not just hardware and bandwidth for under-resourced schools but the expensive, unglamorous parts — curriculum, teacher training, ongoing support — precisely the resources the Sutton Trust data shows are most unequally distributed. Universal, free-at-the-point-of-use AI tutoring could blunt the advantage now accruing to students with private tutors; the evidence on what equitable integration looks like exists, and the gap is one of implementation and will, not knowledge.

And beneath all of it sits the distributional question this book keeps returning to: who captures the enormous productivity gains AI generates, and through what mechanism do any of those gains reach the people being filtered, displaced, and priced out. Bias audits and school funding treat the symptoms. Whether the gains are shared at all — through redistribution, through public investment outside the coastal hubs, through some renegotiation of the bargain between labor and capital — is the underlying question, and the one on which there is, as yet, no political answer at the necessary scale.

What we don't yet know

Intellectual honesty requires marking the edges of the evidence. The résumé-screening audits, powerful as they are, test a specific class of systems — large language models applied to résumé text — under experimental conditions, not the full, proprietary, constantly-updated pipelines that real employers run behind closed doors. It is reasonable to infer that commercial systems share the underlying vulnerability, since they are built from the same materials, but "reasonable inference" is not "measured fact," and vendors' refusal to open their systems to independent audit is itself part of the problem. We know the mechanism is real and the direction is consistent; the precise prevalence across the whole hiring economy is not something the current evidence can pin down.

Larger still is what we cannot yet know because it hasn't happened. No one has measured the lifetime effect of an early AI-access gap on eventual earnings, because the children now diverging in fluency are still children. Answering that would require longitudinal studies that follow cohorts for decades — the kind of patient evidence that, by definition, arrives too late to guide the decisions being made right now. And there is a genuine open question about whether the AI-stratified class structure this chapter describes is a durable new configuration or a transitional artifact of an early, chaotic deployment phase that better tools and rules might yet soften. The two scenarios would look similar today and diverge sharply over time; distinguishing them empirically is work that has barely begun. The intellectually honest position is that we can see the machinery clearly, we can measure its immediate effects, and we cannot yet see where it settles.

Summary

AI did not invent the barriers to social mobility, but it is automating and hardening them, and the evidence that it does so is now specific rather than speculative.

The documented core is the bias in résumé screening. In the most rigorous audit, AI models preferred white-associated names 85.1 percent of the time and Black-associated names 8.6 percent, favored men's names over women's by roughly five to one, and — most starkly — preferred a white male candidate over an identically qualified Black male candidate close to 100 percent of the time. The mechanism is a pattern-matcher trained on a discriminatory past, executing that past at scale, and the bias survives even when names are removed because subtler signals leak identity back.

Around that core, the disadvantages compound. An AI-fluency gap opens in schools before anyone applies, with private and well-rated schools far ahead of under-resourced ones. Entry-level jobs — the historical first rung of mobility — are being automated away fastest, with young workers in exposed occupations down 16 percent while their seniors hold steady. Skill-based and credential-based hiring, sold as meritocratic corrections, reward demonstrated competence that is itself a function of prior advantage. Geographic concentration turns location into destiny. And these forces chain rather than add, producing a three-tier stratification — an AI-fluent elite, an AI-adjacent middle, and an AI-excluded bottom — sorted less by talent than by position when the technology arrived.

The trend of declining mobility is decades old: the odds an American child out-earns their parents fell from 92 percent for the 1940 cohort to about 50 percent for the 1980 cohort. AI is best understood as an accelerant of that decline, not a separate one. The outcomes are not inevitable — bias audits, transparency mandates, human-review requirements, and public investment in AI education access are all real levers, as the Mobley v. Workday litigation and laws like NYC's Local Law 144 and the EU AI Act demonstrate. But their limits are already visible, enforcement is weak, and the deeper distributional question remains unanswered. The window in which technical fixes can still work is closing; past a point we cannot precisely mark, correction will require structural policy rather than better code.

Sources

Last updated: 2026-07-18

V2 (in progress) Previous: V1