Statistical Guarantees in the Search for Less Discriminatory Algorithms

URL SCAN: Statistical Guarantees in the Search for Less Discriminatory Algorithms

FIRST LINE: "U.S. discrimination law can impose liability on firms that fail to adopt a less discriminatory alternative (LDA)..."

The Dissection

This paper is a compliance optimization exercise dressed as legal scholarship. It takes as given the premise that U.S. discrimination law obligates firms to actively search for "less discriminatory algorithms" (LDAs) in high-stakes domains—employment, lending, housing—and then asks the purely technical question: when is the search for LDAs algorithmically "sufficient" to shield a firm from liability?

The authors formalize LDA search as an optimal stopping problem: how many random-seed retrainings should a developer run before they can certify—soundly, to a court—that further search won't yield meaningfully better disparate impact reductions?

The contribution is an adaptive stopping algorithm with high-probability bounds on attainable disparate-impact gains from continued retraining. They validate on credit and housing datasets.

The Core Fallacy

The paper assumes the legal/procedural layer is the operative constraint, when in fact it is cosmetic theater.

The entire framework rests on a fiction: that if firms just search harder, retrain smarter, and document their good-faith effort, they can satisfy anti-discrimination law through algorithmic refinement. The assumption smuggled in is that the problem is implementation—that current models are suboptimal discriminators and better search will surface fair enough alternatives.

This is the same cognitive trap as believing you can make a coal plant "green enough" by installing better scrubbers. You're optimizing the margins of a dying paradigm.

Under the Discontinuity Thesis, the operative question isn't "can we find less discriminatory algorithms?" It's "does the productive participation substrate even exist for this discrimination to matter in the way the law imagines?"

The law assumes a world where:
- Hiring, lending, and housing decisions gate access to economic participation
- Those decisions can be meaningfully audited and corrected
- Disparate impact reflects a tractable harm with a just remedy

AI-driven displacement collapses all three assumptions simultaneously. When mass unemployment eliminates the wage-work access that housing and credit gatekeep, disparate impact auditing becomes performing ethics on a corpse.

Hidden Assumptions

1. Model multiplicity is a feature, not a symptom of fundamental instability. The paper treats the fact that different random seeds produce materially different disparate impacts as an opportunity for ethical optimization. In reality, this sensitivity is evidence that the model space is unprincipled—these are not discoveries of better policies but lottery tickets in a high-dimensional gamble.

2. Firms have incentive to genuinely reduce discrimination. The "good faith" framework assumes firms want to find LDAs and are being strategic about when to stop. In practice, firms will use this framework to certify minimal compliance—the stopping algorithm will become the ceiling, not the floor.

3. Courts and regulators can verify "sufficient search." The authors want to give developers a certificate they can show "to a court." But courts have no technical capacity to audit the bounding algorithm's assumptions or verify the distributional claims. This is a compliance document looking for legal cover.

4. The domain (employment, lending, housing) remains the operative economy. This is the massive unstated assumption. If AI displacement hollows out the employment sector that this entire framework is designed to protect, the "high-stakes decisions" become decisions about an increasingly small pool of economic participants.

Social Function

This is transition management. Specifically, it's elite self-exoneration infrastructure for the technology sector. It does several things simultaneously:

• Creates procedural legitimacy for ongoing AI deployment in high-stakes domains by giving firms a "scientific" answer to "did you try hard enough to be fair?"
• Deflects structural critique by framing discrimination as a search problem rather than a systemic output of maximization-based training on systems that are structurally extractive
• Serves regulatory capture by offering regulators a technically complex solution that requires deep ML expertise to audit—expertise the regulators don't have
• Provides legal cover for continued deployment while genuine structural alternatives (mandatory employment floors, wealth redistribution, economic participation guarantees) are off the table

It's the algorithmic equivalent of installing better locks on the Titanic. Technically not nothing. Structurally irrelevant.

The Verdict

This is sophisticated compliance theater for a regime that is structurally dying.

The paper's technical contribution may be real—it does have genuine mathematical content about bounding improvement gains from repeated retraining. But the entire framework is premised on a legal and economic order that the Discontinuity Thesis says is being dismantled from below by AI-driven productive participation collapse.

The framework asks: how can firms certify they've searched hard enough for fairness? The unasked question is: does the economic substrate that this fairness framework is designed to protect remain solvent?

When employment is displaced by AI, the discrimination in hiring decisions matters dramatically less—the harm isn't that the AI discriminates in selecting who works; it's that there are fewer jobs for anyone to be selected for. Disparate impact in lending becomes moot when the credit-worthy population contracts. Housing discrimination becomes abstract when the economic viability of the housing market itself depends on employment that is evaporating.

The paper is legally sophisticated, computationally legitimate, and structurally irrelevant to the actual dynamics of economic collapse under AI displacement.

Its function in the transition is to give technology firms and their legal teams a credible document to wave at courts while the economic foundation the law assumes continues to rot beneath it.

Lag-Weighted Assessment

The paper's authors are solving the wrong problem with the right tools. This is useful for lawyers. It is not useful for people facing economic displacement. It is a document for the transition intermediary class, not the transition casualty class.

Structural verdict: sophisticated hospice care for a patient who died when productive participation became optional for capital.