Algorithmic Gender Prediction Is Illegitimate, But Gender Imputation Can Yield Valid Measurements

AI summary

Overview

Research area: AI ethics and algorithmic fairness, specifically demographic data imputation, gender prediction, and the conceptual foundations of disparity auditing. The paper draws on transfeminist philosophy, critical HCI, science and technology studies, and measurement theory.

Technical level: Intermediate. This is a conceptual and normative paper rather than a technical one: it introduces no new models, datasets, or benchmarks. Its difficulty lies in the philosophical distinctions it builds, not in mathematics or engineering.

Scope: One sentence: the paper argues that all algorithmic gender prediction is illegitimate because it restricts gender self-determination, but that the narrower practice of gender imputation can still produce valid measurements of sexism that targets women and femininity, while remaining illegitimate and harmful for sexism that targets transgender and nonbinary people.

What This Paper Is About

Measuring discrimination requires demographic labels, but those labels are often missing because of legal restrictions, privacy concerns, or refusal to disclose. Practitioners fill the gap by predicting gender from faces, names, or images, which fairness researchers then use to quantify disparities and build debiasing methods. Transgender and critical scholars argue this practice is outright wrong, creating a bind: gender prediction is said to be both untrustworthy and capable of revealing genuine gender disparities. The paper tries to resolve that bind by separating two different senses of "wrong" — being illegitimate versus being invalid.

Key Contributions

1. A legitimacy/validity distinction applied to gender prediction. The authors translate existing critiques of gender prediction into two separate questions: whether a prediction is illegitimate (lacks normative authority and harms gender self-determination) and whether it is invalid (fails to measure what it claims to measure). They argue these are distinct qualities that prior debates have conflated.

2. A distinction between traditional sexism and oppositional sexism. Drawing on Julia Serano's (2007) analysis, the paper separates sexism that targets women and femininity (traditional sexism) from sexism that targets gender deviance, including transphobia, homophobia, and cissexism (oppositional sexism), and the compounding form Serano calls trans-misogyny. This separation is what allows the authors to claim imputation can be valid for measuring one form while being illegitimate with respect to the other.

3. Definitions that separate gender prediction from gender imputation. Gender prediction is defined as inferring gender from inputs, treating gender as the output that is not known a priori. Gender imputation is a subset of prediction distinguished by application (auditing or describing existing structures or processes) and interpretation (drawing aggregate, descriptive conclusions or motivating structural action).

4. Recommendations for harm minimization and a call for inclusive methods. The authors argue gender imputation should be deployed only when it achieves anti-discrimination benefits that no other reasonable means can achieve, with harms minimized as far as possible, and they conclude by recommending the development of methods that address all kinds of sexism, not only traditional sexism.

Main Findings

• All gender prediction is illegitimate to some degree. The authors anchor legitimacy on a negative criterion: a gender prediction is illegitimate if it restricts individuals' agency and capacity to self-determine their gender. Self-determination is not merely internal; it includes aligning external expression and social recognition with identity. Under this criterion, no gender prediction escapes illegitimacy, though a restriction can be permitted if a specific purpose warrants it and only insofar as that purpose requires.

• Illegitimacy arguments fall into three kinds. The paper distills prior critiques into arguments against (1) the act of predicting gender, (2) the reinforcement of gendered norms, and (3) the classification system of gender labels. The first includes direct harms of misgendering and administrative and data violence; the second includes bioessentialist biometrics and the way algorithmic systems fix gender norms that human interaction could otherwise contest; the third includes restrictive, insufficiently expressive gender schemas that erase or homogenize nonbinary identities.

• Illegitimacy does not entail invalidity. The authors argue that in the imputation case specifically, an illegitimate prediction does not necessarily produce an unusable measurement. They give a hypothetical: an algorithm used by an oppressive state to classify people in order to mistreat them may be valid while being entirely illegitimate in authority.

• Imputation can be valid for measuring traditional sexism. Discrimination against women and femininity is often based on perceived gender or gendered social position rather than gender identity. A nonbinary person misperceived as a woman may still experience catcalling, and people perceived as not conforming to gender norms can experience harms regardless of whether they are cisgender. Because these forms of discrimination are by definition constructed by inference, they can in principle be imputed and measured.

• Imputation is illegitimate and unsuitable for oppositional sexism. Transgender and nonbinary people cannot be accurately classified by existing gender prediction systems, so imputation will likely fail to detect discrimination against them and implicitly privileges one form of discrimination over another. Imputation always carries oppositional-sexist harms, and the authors state it should face heavy scrutiny even when valid.

• Gender imputation is often invalid for the same reasons it is illegitimate. Gender classification systems are not only harmful but inaccurate; gender norms are less stable than model designers assume; and nonbinary people cannot be accurately classified. The authors state that naïve imputation via conventional prediction models will almost necessarily be invalid.

• Validity is judged against three criteria. The paper highlights content validity (measuring the appropriate dimensions of the concept), convergent validity (correlating with other established validated measurements, which gives an empirical way to validate disparity measurements where self-reported gender is available), and consequential validity (weighing how a measurement shapes the quantities studied, the conclusions drawn, and the world).

• Denying any possible validity can cause greater harm. Citing Boyd and Sarathy (2022), the authors note that disputes over statistical correctness can become a political weapon. They point to the CFPB's use of imputed race being attacked as statistically invalid by corporations and political actors seeking to curtail anti-discrimination regulatory powers.

• On the CFPB precedent. Bayesian Improved Surname Geocoding (BISG), a naive Bayes model of race using Census name and location data (Elliott et al. 2009), was used in 2013 by the United States Consumer Financial Protection Bureau to detect racial disparities in auto lending, motivating real anti-discrimination regulation.

• A concrete motivating example. The paper argues that the experience of feeling like a gender minority in a computer science conference room full of perceived men is grounded in aspects of gender such as perception and social position, even without knowing anyone's actual gender identity — and that rejecting these feelings dismisses real experiences of gender discrimination.

• On the underlying pay-gap motivation. The paper notes that quantifying the U.S. gender pay gap in 2024 — that women earned 85% of what men earned — requires data on workers' genders.

Methodology in Plain English

This is an argumentative and conceptual paper, not an empirical one. The authors work through a large body of existing literature — transfeminist philosophy, critical HCI, algorithmic fairness, and measurement theory — and reorganize it into a cleaner set of distinctions. Their main move is analytical: they split the single question "is gender prediction wrong?" into two independent questions about legitimacy and validity, then further split the concept of sexism into traditional and oppositional forms. They define their terms carefully (gender prediction versus gender imputation, dominant versus resistant definitions of gender, identity versus perceived gender versus gender as social position) and then trace how each reason for illegitimacy does or does not undermine the validity of a disparity measurement. They propose recommendations modeled on the principle of beneficence from the 1979 Belmont Report — maximize possible benefits and minimize possible harms. The paper is meant to include three case studies (auditing gender bias in generative image models, measuring gender disparities in film, and imputing gender from personal names), but the truncated content provided does not include the case study details or the remainder of the recommendations beyond the beginning of Section 5.1, so their specific findings are not reported here.

Why This Matters

Impact on research. The paper gives fairness researchers a vocabulary for a debate that has largely been conducted as a standoff. Instead of choosing between "gender prediction is harmful" and "we need labels to measure harm," it offers criteria for when imputation might be defensible and insists that the harms it cannot capture — those to transgender and nonbinary people — be treated as a first-order research problem rather than a footnote. The authors specifically criticize works that relegate this harm to a mere acknowledgment or footnote.

Real-world applications:

• Financial regulation and lending: the CFPB's use of imputed race via BISG to detect disparities in auto lending shows imputation already underpins real anti-discrimination enforcement.
• Settings where demographic collection is legally barred: the paper notes that in the U.S., many financial institutions and government agencies such as the Internal Revenue Service and the Patent and Trademark Office are often not legally permitted to collect demographic data such as race and sex, and that the Equal Credit Opportunity Act originally restricted lenders' collection of demographic data to prevent intentional discrimination.
• Jurisdictions with race-blind policy: France and many other European countries do not collect racial statistics or codify racial categories, so imputation is used as a proxy.
• Auditing and representation measurement: the paper's named case studies target gender bias in generative image models and gender disparities in film, both settings where self-reported labels are unavailable.

Industry relevance. Any organization that trains or audits models on demographic parity needs labels it may not legally or practically be able to collect. The paper's guidance — deploy imputation only when it achieves benefits no other reasonable means can achieve, tailor predictions to narrow gender concepts rather than gender as a whole, and avoid enabling downstream misuse such as new surveillance infrastructure — applies directly to model design, auditing pipelines, and the governance decisions behind them.

Future Directions

• Developing methods for oppositional sexism. The authors state that the illegitimacy of imputation is itself a reason researchers should develop different measurements and techniques to study discrimination against transgender and nonbinary people, rather than relying on imputed binary labels.
• Empirically validating disparity measurements via convergence. Convergent validity offers a path: compare imputation-based disparity measurements against other established validated measurements in settings where self-reported gender is available.
• Establishing when no reasonable alternative exists. The recommendation to use imputation only when benefits cannot be achieved by other means requires operational criteria for what counts as a reasonable alternative — for example, joining records with self-identified sources, which the authors cite as one such alternative.
• Working out the case studies' practical implications. The paper promises concrete guidance from three case studies (generative image models, film, and personal names) on how model design and downstream application factors change the legitimacy and validity calculus; the truncated content does not report those details.

Target Audience

Fairness and AI ethics researchers who need demographic labels they cannot obtain; practitioners building auditing or debiasing pipelines that rely on imputed demographic data; policy and regulatory analysts working in jurisdictions with legal restrictions on demographic data collection; and transgender studies and critical HCI scholars who want a precise articulation of where their critiques bite and where a narrower, narrowly scoped use of imputation might still be defended. Readers looking for empirical benchmarks, model architectures, or quantitative results will not find them here — the paper's contribution is conceptual.

Authors’ abstract

Machine learning ethics researchers and critical HCI scholars have argued that algorithmically predicting gender is wrong. At the same time, other researchers rely on predicted gender labels to study gender disparities and develop algorithmic fairness techniques. How do we reconcile these two seemingly contradictory intuitions? We differentiate two ways gender prediction may be wrong: being illegitimate, thereby contributing to harm; and being invalid, thereby producing unusable measurements. Our analysis translates arguments against gender prediction into these terms of legitimacy and validity and shows how gender imputation applied for fairness purposes can be illegitimate yet still yield valid disparity measurements. We clarify this bind by drawing upon transfeminist literature to distinguish sexism that targets women and femininity from sexism that targets transgender and nonbinary people. While gender imputation can produce valid measurements for the former, it is illegitimate and harmful for the latter. We argue that practitioners should deploy gender imputation only when it would achieve anti-discrimination benefits that cannot be achieved through other reasonable means, while harms are minimized to the extent possible. We examine this tension in three case studies: auditing gender bias in generative image models, measuring gender disparities in film, and imputing gender from personal names. By disentangling legitimacy from validity, and differentiating these two forms of sexism, we show how debates over gender prediction have conflated distinct concerns, obscuring both the settings in which gender imputation can support fairness efforts and the harms towards transgender and nonbinary people that it fundamentally cannot capture. We conclude by recommending the development of more inclusive methods that address all kinds of sexism.