Screening & Exclusion

Somewhere underneath every hiring decision is a question neither side can fully answer: is this claim true? But “neither side” undersells how lopsided the problem actually is. A candidate researching an employer has real infrastructure to work with — Glassdoor alone hosts more than 70 million reviews of pay, benefits, culture, and interview process, drawing over 100 million visitors a month, with roughly 80% of the Fortune 500 covered. An employer researching a candidate has nothing close to an equivalent. There is no Glassdoor for people. And for one specific category of information — what a past employer actually thought of you — American law has made verification close to impossible on purpose.

What It Is

Three groups facing measurable, often invisible exclusion — workers with disabilities, caregivers returning after a gap, and people with criminal records — and the single legal mechanism, the reference-check paradox, that makes adverse selection worse for everyone regardless of background.

How It Works

Workers with disabilities face a gap that’s driven more by labor-force participation than by a hiring-funnel failure. The labor force participation rate for people with disabilities sat at 24.3% in February 2026, against 67.6% for people without disabilities — a massive structural gap. About 75% of people with a disability weren’t in the labor force at all in 2025, versus roughly 32% of people without one, meaning most of the “unemployment” problem for this group happens before the hiring pipeline is ever engaged. There’s a genuine bright spot: disability employment has risen rapidly since the pandemic, driven largely by expanded remote-work availability — a concrete, small-scale case study of a specific arrangement (mandatory in-office presence) directly suppressing employment, and its relaxation directly helping.

Caregivers re-entering after a gap face a documented, controlled penalty: correspondence-audit research finds all forms of caregiving-related employment breaks yield significantly lower interview and hiring probability than continuous employment — though recruiters largely don’t differentiate between types of break (childcare versus eldercare) when applying the penalty. Usefully, the caregiving-gap penalty is somewhat smaller than the penalty for a comparable ordinary long-term unemployment gap, which argues against a “caregiving is uniquely punished” framing in favor of a broader “any gap is punished” one. The mechanism is perceived skill loss and perceived reduced commitment — perceptions, not verified facts, which is the theme running through this entire section. One tested intervention actually works: rewriting a resume to list years-worked totals instead of exact employment dates increased callbacks by about 8% for candidates without gaps and about 15% for candidates with gaps — evidence the penalty is partly a presentation artifact, not an immovable judgment.

People with criminal records are covered by fast-moving reform: more than 80% of the US population now lives in a jurisdiction with some form of ban-the-box or fair-chance policy, and 15 states mandate removing conviction-history questions from initial applications outright. Employment is strongly associated with lower recidivism, giving fair-chance hiring a public-safety case, not just a fairness one. And the same false-negative blindness that runs through this whole article shows up here directly: employers report treating a criminal record as a strong negative signal of future performance, but the evidence suggests these concerns are often exaggerated relative to actual on-the-job outcomes — because, as with any rejected candidate, no employer can observe the counterfactual performance of someone they didn’t hire.

The reference-check paradox

This is a self-contained illustration of adverse selection created entirely by US legal structure. The setup: employers face negligent-hiring liability if they fail to investigate a candidate whose “dangerous propensities” could have been discovered and that candidate later harms someone — real legal pressure to reference-check thoroughly. The trap: employers simultaneously face defamation liability risk if they share negative information about a former employee with a prospective new employer. Legal counsel routinely advises sharing only dates of employment and title — nothing evaluative, positive or negative. The net result is that most employers won’t share anything beyond basic verification, because their own counsel has told them additional information creates more legal risk than value. The reference check has become a compliance ritual rather than an actual information-transfer mechanism — compounded by salary-history bans in numerous states that, in some cases, extend to barring compensation inquiries during reference checks specifically. Eighty-seven percent of the discrepancies background checks actually find relate to employment and education verification — exactly the category reference checks are supposed to catch, and mostly don’t, because the other side of the exchange has been legally silenced.

This is the clearest possible illustration of the structural-versus-contrived distinction from the thesis: both sides needing to verify claims about each other is a universal, structural adverse-selection problem. But the asymmetry between the two sides is where “structural” stops applying and “contrived” starts. Nothing about labor markets in general prevents an employer-side Glassdoor from existing — the demand is obviously there, since the candidate-side version thrives. What’s actually blocking it is the specific defamation-and-negligent-hiring bind described above: a distinctly American legal equilibrium that pulls employers in opposite directions simultaneously and doesn’t have a clean analog constraining what candidates can say about a company. Other legal environments don’t necessarily generate this exact trap, and the fact that the equivalent tool exists and works fine for the other direction is itself evidence this gap is a specific, alterable legal choice rather than an unavoidable feature of hiring.

This dysfunction predates AI by well over a decade. Peter Cappelli’s 2012 book Why Good People Can’t Get Jobs argued the “skills gap” narrative of that era was substantially a myth — employers were chasing unrealistically over-specified “purple squirrel” candidates rather than facing an actual shortage of qualified workers, and applicant tracking systems were already making it harder, not easier, to find qualified candidates — a direct ancestor of today’s 14% AI-screening shortlist overlap problem (see Macro Labor Data).

That history makes one widely-praised “fix” especially instructive: when firms drop formal degree requirements from job postings, the measured effect on actual hiring outcomes is startlingly small — about 0.14 percentage points, or roughly 97,000 workers out of 77 million annual US hires, fewer than 1 in 700 (Burning Glass Institute + Harvard Business School, Feb 2024). Nearly all of that small effect occurred among the 37% of companies that changed both the posting and their actual evaluation process; companies that only edited the posting language saw almost no difference in who was actually hired. Meanwhile 70% of employers in NACE’s 2026 survey claim to practice skills-based hiring, up from 65% the year before — a stated-practice figure sitting in stark tension with the measured outcome. Widely-praised, bipartisan-endorsed reforms can be pure theater unless they touch the underlying evaluation mechanism, not just the posting language.

Why It Matters