Contribution
It separates two harms of everyone using one algorithm. The familiar one is correlated failure under a shock. The new one needs no shock: shared rankings correlate errors, so the second firm to hire finds the first firm’s mistakes are its own. The paper gives two conditions on a noise model of rankings under which this produces a Braess paradox. A more accurate algorithm becomes a strictly dominant strategy for both firms and lowers both firms’ utility and total welfare. It then checks which standard ranking models satisfy the conditions.
Key insight
The value of a ranking to someone who chooses second is not its accuracy alone but its independence from the ranking of whoever chose first. “Whenever decision-makers rely on identical or highly correlated evaluations, they miss out on hidden gems.” Each firm, choosing alone, rightly takes the more accurate tool. Together they lose the diversification that independent errors gave.
Assumptions
- Two firms, one hire each, hiring in a uniformly random order, from n candidates with values x₁ > … > x_n drawn from a known distribution D. Offers are always accepted.
- Each ranking comes from a noisy permutation family F_θ (Definition 1): probabilities differentiable in θ; the true ranking has probability → 1 as θ → ∞; and the expected top candidate after any removal does not fall as θ rises, strictly when nothing is removed.
- Human rankings are independent draws from F_{θ_H}. The algorithm’s is one draw from F_{θ_A}, shared by every firm that uses it.
- Definition 2 (preference for the first position): for π, σ independent from F_θ, E[π₁ − π₂ | π₁ ≠ σ₁] > 0 for every θ.
- Definition 3 (preference for weaker competition): for θ₁ > θ₂, σ ∼ F_{θ₁}, π, τ ∼ F_{θ₂}, E[π₁ after removing σ₁] < E[π₁ after removing τ₁].
Key results
- Identity (3), re-derived. With U_AH the second firm’s utility when it uses an independent ranking after a first firm using the algorithm, and U_AA when both use it, U_AH(θ, θ) − U_AA(θ, θ) = E[(π₁ − π₂)·1{π₁ ≠ σ₁}]. So Definition 2 is exactly “at equal accuracy, the second mover prefers independence”.
- Theorem 1. Under Definitions 1–3, for every θ_H there is θ_A > θ_H at which using the algorithm is strictly dominant for both firms and welfare is lower than if both used humans. Re-derived in THEORY-835.
- Theorem 2. Gaussian or Laplacian RUMs with three candidates meet both conditions for every D. Appendix B gives three-candidate noise models violating each condition. Fig. 2: with 15 candidates and Laplacian noise Definition 2 fails for a uniform D.
- Plackett–Luce (Gumbel RUM) has U_AH = U_AA for every θ, by independence of irrelevant alternatives, so it never meets Definition 2, and “monoculture has no effect”.
- Magnitude. Values may be negative, so welfare at the monocultural equilibrium can be negative where the optimum is positive. With three uniform values centred at 0 and Gaussian noise, an instance loses about 4% of welfare.
- Theorem 3. The Mallows model (Kendall tau, θ = φ − 1) meets both conditions for every D. Appendix E reduces Definition 2 to φ^(i−1)(1 − φ^(−n)) > 1 − φ^(−|i−j|−1), strict for some pair when n ≥ 3.
- §4, k > 2 firms. n = 4, k = 3, φ_A = 2, φ_H = 1.75, uniform values: all firms use the algorithm at ≈ 0.551 each against ≈ 0.552 if all used humans. Recomputed exactly here: 0.55111 and 0.55231, and the algorithm is a best response to each profile of the other two (0.55911 against 0.54708 when one other uses it, 0.57129 against 0.55231 when neither does, 0.55111 against 0.54592 when both do).
- Theorem 4. With known sequential order and φ_H ≥ φ_A, all-human is optimal, uniquely if φ_H > φ_A. For φ_A > φ_H, k = 5 and n = 6, all 16 sequences starting with A occur as optimal profiles, ordered like a binary counter as φ_A rises (Fig. 4), observed and not proved.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Monoculture can lower average decision quality without any shock | strong | Theorem 1 with proof; re-derived |
| C2 | A more accurate shared algorithm can be strictly dominant for every firm and still lower each firm’s utility and total welfare | strong | Theorem 1; Mallows and RUM instances; recomputed for Mallows |
| C3 | The effect needs a noise model in which independent re-ranking beats taking the next in one’s own ranking; Plackett–Luce has none | strong | identity (3); the IIA argument |
| C4 | Mallows rankings always meet the conditions; Gaussian and Laplacian RUMs do for three candidates | strong for the stated cases | Theorems 2–3, Appendices C and E |
| C5 | The conditions can fail for other noise models and more candidates | strong | Appendix B; Fig. 2 (simulation) |
| C6 | The Braess effect persists with more than two firms | moderate | one computed instance (§4); no theorem |
| C7 | Optimal sequential profiles follow a binary-counter order | weak | computation for k = 5, n = 6; no proof |
Concepts
- Algorithmic monoculture: many decision-makers relying on one algorithm, so their errors are identical.
- Noisy permutation family F_θ: a θ-indexed distribution over rankings meeting Definition 1.
- Preference for the first position, for weaker competition: Definitions 2 and 3.
- Braess paradox here: adding a better option makes the unique equilibrium worse for everyone.
Connections
The concern it formalizes is Citron and Pasquale’s, by the agricultural analogy (Power and Follett 1987) and its security version (Birman and Schneider 2009). The ranking models are Thurstone’s random utility models, Mallows (1957) and Plackett–Luce (Luce 1959). The paradox is Braess’s (1968). Kleinberg’s small-world paper, LIT-1103, is unrelated.
Bearing on the record
- TERM-064. A25’s “common legibility regimes and loss of organizational diversity” is exactly C2. When institutions share one rule class for reading candidates, each reads more accurately and all read the same errors. The cost falls on the institutions’ joint performance and on the candidates the shared ranking misplaces, which is A25’s point that “the party benefiting from increased legibility need not be the party paying to become legible”, here with the institutions paying too. The paper’s harm is to firms’ utility and welfare; the harm to particular applicants it names and sets aside.
- Scott’s legibility (LIT-1321, unread) and lock-in (LIT-1314, Arthur, unread). This is a static, two-firm result with no adoption dynamics, so it is not lock-in. Each firm’s adoption is a dominant strategy, not a coordination effect.
- The record’s correlated-evidence condition (CLAIM-342, through CLAIM-402). Contexts that share a category’s use are not independent evidence. Here, rankings that share an algorithm are not independent screens, and the loss from treating them as if they were is measurable. The record’s join.
No instruction for machine-learning practice is drawn. The LIT is flagged
anthology-candidate.
Limitations
- Two firms, one hire each, in random order. The k-firm case is computation.
- The conditions are model-dependent. Plackett–Luce kills the effect, and Laplacian noise with 15 candidates violates Definition 2.
- Mallows ignores values. The authors: it “does not depend on the underlying candidate values, so according to this model, monoculture can produce arbitrarily large negative effects”.
- The magnitude is small where measured. About 4% in the one Gaussian instance, and 0.551 against 0.552 for three firms.
- No market response. Wages, applicant behaviour and vendor choice are fixed. With several vendors firms might spread across algorithms, a case the authors leave open.
Open questions
- The authors’: Price-of-Anarchy bounds where welfare is non-negative; more firms and more algorithms; decomposing correlation into shared inputs such as standardized tests.
- Does Theorem 2 extend to Gaussian noise with more than three candidates?
- Is the binary-counter order a theorem?