Does the Equal Pay Act apply to pay decisions made by an algorithm?

Yes, and it applies more harshly than most employers expect. The Equal Pay Act of 1963 prohibits paying employees of one sex less than the other for substantially equal work in the same establishment. It has no intent requirement — a plaintiff does not need to show the employer meant to discriminate, only that the pay differential exists between comparable jobs. An algorithm that recommends a lower offer for a woman than for a comparable man produces exactly the differential the statute addresses. The fact that a model, not a manager, generated the number is legally irrelevant; the employer set the pay.

How is an Equal Pay Act claim different from a Title VII disparate-impact claim?

The burden structure is different in a way that matters enormously for AI. Under Title VII disparate impact, the plaintiff must first show a statistically significant adverse effect, and the employer can then defend the practice as job-related and consistent with business necessity. Under the EPA, once the plaintiff shows unequal pay for substantially equal work, the burden shifts fully to the employer to prove the differential is entirely explained by one of four affirmative defences: seniority, merit, a system measuring quantity or quality of production, or a factor other than sex. There is no business-necessity defence and no room for a partially explained gap. An employer relying on a model it cannot fully decompose is trying to carry an affirmative burden with an unexplainable variable.

Can prior salary be used as a factor in an AI compensation model?

In much of the country, no. More than twenty states and numerous municipalities have salary-history bans that prohibit asking for or relying on prior compensation, and several federal circuits have held that prior salary is not a permissible 'factor other than sex' under the EPA precisely because it perpetuates historical discrimination. The compliance failure mode with AI is indirect: even when prior pay is excluded as an explicit input, features like current-employer pay band, previous title, tenure-adjusted market position, or a candidate's own stated expectation can act as proxies that reconstruct it.

Does market-rate salary benchmarking data create discrimination risk?

It can. Benchmarking datasets describe what the market currently pays, and where the market pays women or minority workers less for comparable work, a model trained to match market rates will reproduce that gap and present it as an objective recommendation. Aggregate benchmarks are usually safe because they are not segmented by protected class. Risk enters when benchmarks are narrowed by geography, industry sub-segment or job-family in ways that correlate with demographics, or when a model is tuned on the employer's own historical offers, which encode its own past decisions.

What should an AI compensation audit actually measure?

Four things a standard hiring bias audit will not cover. First, a substantially-equal-work grouping — EPA comparators are defined by skill, effort, responsibility and working conditions, not by job title, so the analysis must be run on functional groupings. Second, unexplained residual pay gap after controlling only for the four statutory defences, not for every available covariate. Third, proxy detection for prior salary and negotiation-derived features. Fourth, an assessment of whether the model's discretionary range gives managers enough latitude to reintroduce a gap the model itself did not create.

Home / Blog / AI Pay-Setting and the Equal Pay Act

Your Bias Audit Tested the Hiring Model. Nobody Tested the Offer.

The last three years of AI employment-law attention went almost entirely to screening: resume parsers, video interviews, assessment scoring, the NYC bias audit rule. Meanwhile a second class of model moved into the same workflow with far less scrutiny — the one that decides what to pay. Offer recommendation engines, market-rate benchmarking tools, merit-increase allocators and total-rewards optimisers now set or constrain compensation at a large share of mid-market employers, and they answer to a statute with a materially harsher structure than the one everyone audited against.

The Equal Pay Act has no intent element, no business-necessity defence, and it puts the burden of full explanation on the employer. That is close to the worst possible legal environment for a model whose recommendations you cannot fully decompose.

Why this is not covered by your existing audit: NYC Local Law 144 covers automated employment decision tools used to screen candidates for hiring or promotion. A tool that recommends an offer amount for an already-selected candidate frequently sits outside that definition — and even where it is covered, the audit's four-fifths impact ratio is a Title VII metric that says nothing about whether an individual pay differential is fully explained by a statutory defence.

The burden structure is the whole story

Under Title VII disparate impact, the flow is familiar: plaintiff shows adverse impact, employer defends the practice as job-related and consistent with business necessity, plaintiff may then show a less discriminatory alternative. Employers have built their entire AI governance posture around the middle step — validation studies, job-relatedness documentation, alternative-model analysis.

The Equal Pay Act does not have that middle step. Once a plaintiff establishes that a man and a woman were paid differently for substantially equal work in the same establishment, the employer must prove the entire differential is explained by one of four things:

A seniority system
A merit system
A system measuring earnings by quantity or quality of production
A differential based on any factor other than sex

Three of those four require a system — something structured, communicated and applied consistently, not an ad hoc judgment. And the fourth, the catch-all, has been narrowed considerably: several circuits require the factor to be job-related or serve a legitimate business purpose, and multiple circuits have held that prior salary alone cannot qualify.

The unexplained residual is the liability

A pay-equity regression that explains 96% of the variance is a good model and a bad defence. The 4% it cannot attribute to a statutory factor is precisely what the employer has to account for. Model complexity actively hurts here: a gradient-boosted recommendation with two hundred features may be more accurate than a banded matrix and far less defensible, because "the model weighted these interactions" is not a seniority system, a merit system, a production system, or an articulable factor other than sex.

Where the gap enters an AI compensation stack

Four entry points account for most of the exposure, and none of them require anyone to have done anything deliberate.

1. Prior-salary proxies

Salary-history bans removed the field from the form. They did not remove the information from the system. Current employer, previous title, years in band, equity-refresh eligibility and "candidate expectation" all correlate with prior pay, and a model trained to predict acceptance probability will happily use them. If prior pay is not a valid factor other than sex, a faithful reconstruction of prior pay is not either.

2. Negotiation-responsive pricing

Models tuned to minimise offer cost subject to acceptance probability will systematically offer less to candidates predicted less likely to negotiate. Propensity to negotiate is well documented as correlating with sex, and an optimiser does not need the protected attribute to find the pattern. This is the single most direct path from a benign objective function to a facially actionable pay differential.

3. Historical-offer training data

Fitting on your own past five years of offers means fitting on your own past five years of decisions, including the ones a pay-equity study would have flagged. The model reproduces the gap and, worse, launders it: what was previously a set of individual manager judgments becomes a consistent, documented, system-wide practice. Consistency is usually a compliance virtue. Here it converts scattered anecdotes into a pattern-or-practice exhibit.

4. Discretionary range on top of the recommendation

Most tools output a range and let a manager choose within it. Audit the model and you may find it clean; audit the final numbers and the gap reappears, because placement within the range is unstructured human judgment. Measure the delta between recommendation and final offer, segmented — that residual is often larger than anything the model contributed.

The state layer is stricter than the federal one

The federal EPA is the floor. Several states have materially expanded it in ways that interact badly with algorithmic pay-setting:

Broader comparator standards. California and New York use "substantially similar work" rather than "equal work," widening the pool of employees a plaintiff can compare against — which means the functional groupings your audit uses have to be wider than your job architecture.
Multi-site comparisons. Some states relax the federal same-establishment requirement, so geography-based pay bands become comparable across locations rather than insulated from each other.
Protected classes beyond sex. Several state equal-pay statutes extend to race, ethnicity and other characteristics, importing the EPA's employer-side burden into claims the federal statute never covered.
Pay transparency and reporting. Posted-range requirements and pay-data reporting mean the differential is increasingly visible to plaintiffs, regulators and current employees without anyone filing a discovery request.

What a defensible compensation-model audit looks like

Group by function, not title. Build comparator groups from skill, effort, responsibility and working conditions. Titles are an artifact of your levelling system and will hide the exact comparisons a plaintiff makes.
Regress against the statutory defences only. Run the model with seniority, documented merit ratings and production measures as the permitted explanators. Every additional covariate you add makes the residual smaller and the defence weaker, because you will have to justify each one as a factor other than sex.
Hunt for prior-pay proxies explicitly. Regress each candidate feature against known prior compensation on a historical sample. Anything with meaningful predictive power is a reconstruction risk regardless of intent.
Audit the human layer separately. Compare recommendation to final offer, and recommendation to final merit increase, segmented by group. Report it as its own number.
Test the objective function, not just the outputs. If the model optimises cost subject to acceptance, you have a structural problem that output testing on today's population may not surface until the population shifts.
Run it under privilege, and then act on it. Pay-equity studies are typically conducted under attorney-client privilege for good reason. That protection is worth little if findings sit unremediated — an identified gap you chose not to close is a materially worse fact than one you never measured.
Get vendor decomposition in writing. Before procurement, require the vendor to produce a per-decision explanation adequate to support an affirmative defence, plus contractual commitments on training data provenance and indemnification.

Frequently Asked Questions

Our vendor says their model does not use gender as an input. Is that sufficient?

No. The Equal Pay Act does not ask whether sex was an input; it asks whether the pay differential is fully explained by a permitted factor. A model with no protected attribute can still produce systematically different outputs through correlated features, and the employer still has to explain the resulting gap. Attribute blindness is a starting point for a Title VII conversation and close to irrelevant to an EPA one.

We only use AI for market benchmarking; humans make the final call. Are we covered?

Partially, and possibly worse off. Human final authority does not remove the employer's obligation to explain the differential, and it adds a second, less structured decision layer that is itself unexplainable. Courts have not been receptive to the argument that a human rubber-stamp launders an algorithmic recommendation. Audit both layers and be able to describe each as a system.

How far back does Equal Pay Act liability reach?

Further than most employers assume, because each discriminatory paycheck can restart the clock. The statute of limitations is two years, three for willful violations, and the Lilly Ledbetter Fair Pay Act treats each paycheck reflecting a discriminatory decision as a fresh violation. A pay gap introduced by a model in 2024 that is still reflected in today's salary is not a stale claim.

Does this apply to equity, bonuses and benefits, or only base salary?

The EPA covers 'wages,' which is read broadly to include salary, overtime, bonuses, stock, profit sharing, expense accounts and benefits. Total-rewards optimisers that shift value between cash and equity are squarely in scope, and they are often less audited than base-pay tools because they feel like a benefits decision rather than a pay decision.

We are a small employer. Does the Equal Pay Act apply to us?

Almost certainly. The EPA is part of the Fair Labor Standards Act and does not carry Title VII's fifteen-employee threshold — it reaches employers covered by the FLSA, which is most businesses engaged in interstate commerce. Small employers buying an off-the-shelf compensation tool are often the most exposed, because they have the least capacity to audit what the vendor shipped.

Compensation is the least-audited AI decision in the building

Screening tools got the bias audits because a law named them. Pay-setting tools answer to an older statute with a harder burden and, in most organisations, no audit at all. If you can produce a bias-audit report for your resume screener and nothing equivalent for the model that sets offers, that asymmetry is the finding.

Start with an inventory: every system that produces or constrains a pay number, who owns it, what it was trained on, and whether anyone can explain a single recommendation well enough to defend it.