The Resume Trap.
The résumé never reaches a human. It hits an algorithm that scores criteria it won't disclose, against benchmarks it won't share, using weights it changes without notice. The candidate optimizes harder. The algorithm optimizes harder. Pe=16.8. The arms race has no end because only one side can see the board.
Void dimensions
ATS criteria hidden, scoring weights undisclosed, no feedback on rejection. Candidates optimize for unknown signal. Each optimization attempt shoots at a concealed target.
System does not respond to candidate self-correction. Rejection produces no learnable signal. The arms race has no target because the target is invisible.
Job market existential stakes create high engagement, but coupling is partial — candidates can exit to other employers. Rated 2 not 3 for this reason.
| HR tech deployment | O | R | α | V | Pe |
|---|---|---|---|---|---|
| ATS with hidden scoring (Workday, Taleo) | 3 | 3 | 2 | 7.2 | 16.8 |
| AI video interview (HireVue) | 3 | 3 | 3 | 9 | >>4 |
| ATS with disclosed scoring rubric | 2 | 2 | 2 | 6 | 3.8 |
| Structured interview + blind CV | 1 | 1 | 2 | 4.4 | 1.2 |
| Internal referral only (no ATS) | 2 | 2 | 1 | 5.4 | 2.1 |
Deep dive
Candidates learn to stuff keywords. ATS vendors sell keyword optimization tools. Companies update their criteria secretly. The market settles at a mutual opacity equilibrium: neither side can see the other clearly, both optimize harder.
This is the void arms race. The framework predicts that Pe stabilizes above drift threshold when O=3 because the feedback loop required for self-correction (candidate learns rejection reason → adjusts → reapplies with improved signal) is structurally blocked. The ATS arms race is not a market failure — it is the thermodynamic equilibrium of maximum opacity.
ATS vendors have monetized both sides of this asymmetry: they sell hiring software to employers and resume optimization tools to candidates. The opacity that creates the problem generates a secondary market for expensive workarounds. This is not incidental — it is the revenue model of O=3.
Field experiments (Bertrand & Mullainathan 2003; Kline, Rose & Walters 2022) show a 50% callback rate reduction for identical CVs with minority-signaling names. ATS does not "intend" discrimination — it proxies on proxies: zip code → school district → race; name → ethnicity; graduation year → age.
The discrimination is structural, not intentional. This matters legally and practically. Intentional discrimination requires a discriminatory actor. Structural discrimination requires only an opaque algorithm and a training corpus that reflected historical bias. The ATS reproduces historical hiring patterns because it was trained on them.
The void framework identifies this as an O=3 prediction: when the mechanism is hidden, discriminatory proxies cannot be identified, challenged, or removed. Transparency is not merely a procedural good — it is the precondition for any form of accountability. At O=3, discrimination is an expected output of the architecture.
The 1.6× CV whitening success rate (the probability uplift from removing ethnic-signaling information) is a direct measurement of the gap between O=3 and O=1 outcomes. That gap is the cost of opacity.
Annex III §4 of the EU AI Act classifies HR systems affecting recruitment decisions as high-risk. This includes ATS screening, CV ranking, and automated interview assessment. High-risk classification triggers the full compliance stack:
- Art. 13 — Transparency: Meaningful explanation of system outputs. Current ATS deployments fail this — rejection produces no explanation of criteria applied.
- Art. 14 — Human oversight: Meaningful human review capability. "A human clicks approve" does not satisfy this if the human sees only the ATS output without underlying rationale.
- Art. 10 — Data governance: Training data must be assessed for bias. Systems trained on historical hiring data inherit historical discrimination.
- Art. 9 — Risk management: Ongoing monitoring of discriminatory outcomes by protected characteristics.
Pe=16.8 predicts Art. 13 violation directly: opacity (O=3) is the mechanism that causes both the arms race and the regulatory non-compliance. Reducing O reduces Pe and moves the system toward compliance simultaneously. The framework and the regulation are pointing at the same mechanism.
D1 — Agency attribution: Candidate attributes rejection to personal inadequacy rather than algorithmic opacity. The opacity of the mechanism makes self-assessment impossible. Candidates cannot distinguish between "unqualified for this role" and "failed the ATS keyword match on this specific criteria set." Both feel identical from the outside. The result is inaccurate self-models: skilled candidates conclude they are less qualified than they are.
D2 — Boundary erosion: Resume inflation — credentialing that doesn't reflect skill, keyword gaming, authenticity erosion. Candidates reshape their self-presentation to match inferred ATS criteria, not actual job requirements. The resume becomes a keyword optimization document, not a representation of capability. At scale, this corrupts the signal value of the credential system itself.
D3 — Harm facilitation: Structural hiring inefficiency, discrimination reproduction, talent misallocation. Companies can't find qualified candidates while qualified candidates can't get through ATS. Labour market mismatch at scale. The harm is not individual failure — it is systemic inefficiency caused by opacity preventing accurate signal transmission in both directions.
- P1 Companies with transparent ATS criteria (published scoring rubrics, disclosed weighting) show lower time-to-hire and higher quality-of-hire metrics than opacity-matched controls — testable via HR analytics datasets with disclosed vs. undisclosed criteria conditions.
- P2 Pe correlates with EEOC complaint rates across HR platform deployments at ρ > 0.70 — platforms scoring higher on O should show higher per-hire EEOC complaint rates, testable against EEOC administrative data linked to ATS vendor market share.
- P3 Adding rejection reasons (reducing O from 3 to 2) reduces re-application keyword optimization arms race, measurable as reduction in resume keyword density and increase in application-to-interview conversion rate within 6 months of implementation.
- P4 Structured interview + blind CV outperforms ATS on predictive validity (correlation of hiring decision with 12-month performance review) — substantial supporting evidence base exists; replication with Pe-scored ATS variants would directly test the opacity → validity degradation prediction.
- P5 EU AI Act Art. 13 compliance audits will show higher non-compliance rates for ATS platforms scoring O=3 than O≤2 — testable as enforcement actions begin under the high-risk AI system provisions (Annex III §4) post-2025 enforcement date.