Abstention as a Fairwashing Channel: Auditing Jointly Trained Reject Pipelines

Many deployed classifiers are allowed to abstain. When the model is unsure, it hands the case to a human reviewer instead of deciding. This reject option is usually presented as a safety feature, and newer methods explain their refusals ("I am abstaining because this case sits close to the decision boundary"). We argue that the reject option is also an unaudited place to hide discrimination. Suppose an organization's model systematically denies qualified members of a protected group. It can route exactly those cases to "human review" and attach a plausible uncertainty explanation to each one. The automated decisions left behind then look demographically fair, rejection rates can be matched across groups, and every refusal comes with a respectable reason.

This is fairwashing (making a biased model look fair through its explanations or reported metrics) carried out through non-decisions rather than decisions. We propose to build this attack and to build an audit against it. The audit rests on one observation: a rejection makes a checkable claim, namely that the case is genuinely uncertain. An auditor can estimate that uncertainty from data alone and test whether, among equally uncertain cases, group membership still predicts who gets rejected.

Core Claim

Existing fairness checks on selective classifiers look at three things: fairness of the accepted decisions, parity of rejection rates, and the plausibility of per-case reject explanations. We claim all three can be satisfied at once by a laundering pipeline, especially one trained "one-pot."

The term is borrowed from chemistry. There, one-pot synthesis runs several reaction steps in a single vessel without isolating and purifying the intermediates. Examples include cascades triggered once (Bioinspired Total Synthesis of (+)-Kopsiyunnanine B) and multicomponent assemblies that form several bonds together (One-pot double annulations to confer diastereoselective spirooxindolepyrrolothiazoles). The ML analogue is a pipeline in which the predictor, the rejector and the reject explainer are optimized jointly, so the intermediate each auditor would normally inspect in isolation never exists as a separate, frozen object. We use the chemistry only for this structural contrast between stepwise and joint construction. No chemical experiments are proposed.

Our testable claims are two. First, one-pot laundering achieves a larger apparent fairness gain at equal detectability than a stepwise laundering pipeline, in which the black box, the rejector and the explanations are each fit and frozen in sequence. Second, a conditional-independence audit catches both kinds of pipeline at a controlled false-alarm rate on honest pipelines. Let $R$ be the rejection indicator, $A$ the protected attribute, and $U$ an uncertainty estimate the auditor computes without querying the owner's internals. An honest uncertainty-driven rejector should then satisfy

$$ P(R=1 \mid A=a,\, U=u) = P(R=1 \mid U=u) \quad \text{for all groups } a, $$

and the audit statistic is the group gap in rejection rate within bins of $U$, averaged over bins.

Prior Work and Where This Sits

The gap. Two lines of work establish it. A Fair Selective Classifier to Put Humans in the Loop shows that plain uncertainty-based rejection hurts protected subgroups. It proposes FSC, which combines subgroup checks, local situation testing (comparing a case with similar people from a reference group) and confidence into a four-way routing rule. Topics in Selective Classification likewise finds that group-agnostic abstention magnifies minority rejection. Both treat the owner as honest. Neither asks what an adversarial owner could do with the reject channel, and that is our question.

The explanation surface. L2loRe: a method for explaining the reject option supplies the explanation we attack. It rejects cases whose nearest counterfactual (the closest input that flips the prediction) lies within distance $\tau$, and it emits rule explanations. Counterfactual Explanations Can Be Manipulated shows that counterfactual search can be steered by warping the loss landscape, producing recourse 20.1× cheaper for a favored group. We combine the two: a jointly trained owner can make L2loRe distances small for exactly the minority cases it wants to defer, so the explanation is faithful to a manipulated model and still misleading.

The attack template. Fairwashing: the risk of rationalization supplies this through LaundryML. LaundryML searches rule lists that trade fidelity to a biased black box against apparent unfairness, and Characterizing the risk of fairwashing shows such rationalizations transfer. Both assume every instance receives a forced prediction. Our attack instead uses abstention as the degree of freedom.

Constraints on the defense. Three papers limit what the defense can promise. Washing The Unwashable : On The (Im)possibility of Fairwashing Detection warns that detecting fairwashing is hard in general. We therefore target a narrower signature: rejection asserts uncertainty, and that assertion can be checked against data. Predictive Multiplicity in Classification shows that many near-equally accurate models can disagree on individual cases. The auditor's $U$ must therefore not come from a single model the owner could argue against. We build $U$ from a set of near-optimal auditor models, so that legitimate multiplicity counts as legitimate uncertainty. Robust ML Auditing using Prior Knowledge motivates our threat model: the owner controls what it returns during an audit. For that reason the test uses only the owner's reject decisions plus data the auditor holds.

The uncertainty decomposition. Causes of Rejects in Prototype-based Classification Aleatoric vs. Epistemic Uncertainty motivates splitting $U$ into two parts. Aleatoric uncertainty covers overlap between classes. Epistemic uncertainty covers novelty, meaning inputs unlike the training data. The split matters because an owner might claim that minority cases are "novel." That claim is checkable too, with density or distance-to-training estimates.

Research Program

The attack. We first construct laundering pipelines on tabular benchmarks with protected attributes. The owner starts from a deliberately biased black box. The one-pot attacker jointly trains a predictor, a rejector and an L2loRe-style explainer. The joint objective rewards four things: agreement with the biased black box on the full population (the owner keeps its behavior), a small demographic-parity gap on accepted cases, a small gap in group rejection rates, and a high fraction of rejects whose counterfactual distance falls below $\tau$ so that the explanation reads as "ambiguous." Coverage is held at a fixed level.

The stepwise attacker pursues the same goals with isolated stages. It freezes the black box, fits a rejector to its outputs in the manner of LaundryML search, and then runs L2loRe post hoc without landscape warping. Comparing these two attackers tests the first half of the core claim.

The audit. Second, we build the auditor. From audit data alone it fits a set of near-optimal models and computes two quantities per case. Aleatoric uncertainty is the mean predicted entropy over the model set. Epistemic uncertainty combines disagreement across the set with a nearest-neighbor novelty score. The auditor queries only the owner's accept/reject decisions and computes the binned statistic above, using a permutation test within bins to set the threshold.

The adaptive attacker. Third, we study an attacker who knows the audit and is penalized whenever its rejects are not auditor-uncertain. The central scientific output is the resulting trade-off curve. It shows how much apparent accepted-set fairness the owner can still buy while passing the audit. If that curve stays high, the audit is weak and we will report it as such.

Baselines

Honest reference pipelines. The naive reference point is an honest uncertainty-based selective classifier (USC) that rejects the least confident 20% of cases. Alongside it we run honest FSC and honest L2loRe, each across several model families and seeds. These pipelines calibrate false alarms: any audit must pass them. They also anchor the fairness damage attackers must hide. For Adult at 80% coverage, FSC reports a false-negative rate for Black women of 0.67 under USC versus 0.49 under FSC (A Fair Selective Classifier to Put Humans in the Loop). L2loRe reports COMPAS accuracy rising from 0.72 to 0.827 on accepted samples (L2loRe: a method for explaining the reject option). We reproduce both numbers before building on them.

Standard audit checks. The strongest existing audit is the composite check a careful regulator would run today. It combines accepted-set demographic parity, rejection-rate parity, and plausibility of L2loRe explanations. We run it with the same audit sample and budget as our test. Our audit beats it if it detects laundering pipelines that the composite check passes, with detection area under the curve clearly above the composite's and a false-alarm rate at or below 5% on honest pipelines.

Auditor-side situation testing. A second strong competitor is FSC's situation testing, run by the auditor on rejected cases. It compares each rejected minority case with its nearest majority neighbors. This is the closest existing tool to our test. It differs by conditioning on feature similarity rather than on auditor-estimated uncertainty. We give it the same audit data and neighbor budget.

Ablations. One ablation computes $U$ from a single auditor model instead of the near-optimal set. This isolates whether handling multiplicity is what keeps false alarms low. A second ablation drops the epistemic component. This tests whether owners can evade by framing minority rejects as novelty.

Why not just…?

Why not just check that rejection rates are equal across groups? Rate parity constrains how many cases are rejected, not which ones. The laundering objective explicitly includes a rate-parity term. The owner satisfies it by also rejecting genuinely uncertain majority cases while choosing minority cases for their predicted denial. Experiment 1 below directly measures whether pipelines passing rate parity still fail the conditional test.

Why not just demand forced predictions on the rejected cases and audit those? If the owner complied honestly, this would work, and it is a real alternative. But the owner controls those outputs, and a one-pot system can return rationalized forced predictions, just as LaundryML returns rationalized surrogates (Fairwashing: the risk of rationalization). We include an honest-compliance variant as an upper bound. We also report how our test compares with it when the owner does not tamper, so readers can see what is lost by not trusting the owner.

Why not just use FSC-style situation testing? Situation testing flags a rejected minority case when similar majority cases are accepted. An owner can defeat it by rejecting the similar majority neighbors too, which also serves rate parity. Situation testing also cannot say whether the shared rejection is justified. The conditional test asks the question that rejection itself raises: were these cases actually uncertain? The head-to-head against the auditor-side situation-testing baseline settles which catches more. If situation testing matches us, we will say so and shrink the claim to the one-pot result.

Why not just project explanations onto the data manifold, as with tangent-space projection? Fairwashing explanations with off-manifold detergent defends gradient saliency maps. Its tangent estimation cost 48 hours on four P100s for CIFAR10. It also does not address which cases a rejector selects, since those selections occur on the data manifold. It is the wrong tool here, and we do not run it as a baseline.

Proposed Experiments

We use Adult, COMPAS and German Credit, where FSC, L2loRe and LaundryML already provide reference behavior, plus Communities and Crime, where landscape manipulation was demonstrated. For each dataset we train about 30 honest pipelines and 30 laundering pipelines of each type, at coverage levels of 70–90%.

Experiment 1 asks whether one-pot and stepwise laundering both pass the composite audit while restoring accepted-set parity. We measure the accepted-set parity gap, rejection-rate gap, explanation plausibility, and the full-population false-negative rate for the protected group. The last measure is the true harm and is only visible to us.

Experiment 2 compares detection area under the curve for our test, the composite check and auditor-side situation testing, and plots detection power against audit sample size.

Experiment 3 tests the first core claim: at matched detectability under the composite audit, does the one-pot attacker achieve a lower accepted-set parity gap than the stepwise attacker?

Experiment 4, the decisive check, runs the adaptive attacker. It traces the curve of achievable apparent fairness gain against audit pass rate. The proposal succeeds if forcing rejects to be auditor-uncertain removes most of the laundering gain. It fails if the owner can reshape which cases are uncertain to the auditor's reference models. That could happen by exploiting features the auditor's model set treats as ambiguous, and we will characterize when it does.