Fairness for Unobserved Characteristics: Insights from Technological Impacts on Queer Communities

1 INTRODUCTION

• Sexual orientation and gender identity are often missing, unknown, or unmeasurable, challenging fairness approaches that assume protected characteristics are available.Classification-parity approaches equalize predictive performance across groups, while anti-classification approaches omit protected attributes from decisions [43].Source: §1 · p. 1
• Queer communities’ experiences may offer insights transferable to fairness research involving disability, class, religion, and race.The paper frames this work as connecting queer-focused AI and sociotechnical research with approaches to protecting unobserved characteristics.Source: §1 · p. 2
• The paper reviews how AI may empower queer communities and support LGBTQ+ rights, while also examining related harms and challenges.Its contributions include establishing requirements for fairness when identity data are sensitive and unmeasurable.Source: §1 · p. 2

2 CONSIDERATIONS FOR QUEER FAIRNESS

• Case studies examine how AI systems interact with sexual orientation and gender identity, highlighting potential benefits and risks.The review aims to motivate technological solutions that are inclusive and beneficial to everyone.Source: §2 · p. 2
• The cases identify cross-cutting concerns including privacy, feature imputation, context sensitivity, coded inequity, participation, and fair sequential processes.Source: §2 · p. 2
• The cases focus on challenges raised by unobserved and missing characteristics.Source: §2 · p. 2

2.1 Privacy

• Non-consensual outing can expose queer people to emotional distress and serious physical or social harms, especially where queerness is discriminated against, criminalised, or persecuted.Privacy violations can also threaten human rights and access to employment and education.Source: §2.1 · p. 2
• The review is not exhaustive because queer communities are heterogeneous and their concerns vary across places and subpopulations.Source: §2.1 · p. 2

2.1.1 Promise.

• Privacy-preserving machine learning and adversarial filters may reduce information leakage and re-identification risks, but robustness and collaboration between privacy and machine-learning researchers remain challenges.Source: §2.1.1 · p. 2

2.1.2 Risks.

• Inferring sexual orientation or gender identity from online behavior, images, or other data can enable surveillance, outing, and physical-safety harms, while reinforcing essentialist beliefs about queerness.The paper argues for fairness methods that work without direct group-membership data, using contextually suitable proxies or treatment of similar individuals under plausible unobserved groupings.Source: §2.1.2 · pp. 2–3

2.2 Censorship

• Restrictions on queer expression can isolate queer people, shame queerness, and undermine human rights, while also harming content producers’ expression and income.Source: §2.2 · p. 3

2.2.1 Promise.

• Machine learning might help analyze censorship patterns and identify safer platforms, but these applications remain speculative because no large-scale study of censored queer content has been conducted.Source: §2.2.1 · p. 3

2.2.2 Risk.

• Automated moderation can unfairly censor queer expression, and tools for detecting queer content can also be abused to enforce censorship or heteronormative attitudes.Fairness research could investigate anomalous omission of information about queer identity in language and video; fairness for censorship seeks to reveal unfairly prevented expression, unlike privacy fairness’s goal of preventing sensitive-data imputation.Source: §2.2.2 · pp. 3–4
• Generative models make digital content fabrication easier, adding to concerns about misinformation used to justify abuse and suppression of queer communities.Source: §2.2.2 · p. 3

2.3 Language

• Language has a history of being weaponised against queer communities, and inclusive usage must account for stereotypes, assumptions about identity, and reclaimed terms.For example, “That’s so gay” equates queerness with badness, while “queer” itself may be reclaimed; systems should not censor language solely because of its adjacency to the queer community.Source: §2.3 · p. 4

2.3.1 Promise.

• NLP applications could support respectful communication about sexual orientation and gender identity by overcoming human biases and establishing inclusive norms.The paper identifies machine translation, speech recognition, sentiment analysis, question answering, chatbots, and hate-speech detection as applications.Source: §2.3.1 · p. 4

2.3.2 Risks.

• Language models reproduce biases, stereotypes, and abusive speech found in training data, while fairness frameworks for language remain nascent.Fair language systems must reflect the contextual nature of human discourse.Source: §2.3.2 · p. 4

2.4 Fighting Online Abuse

• Queer people face hate speech, harassment, and trolling online, while abuse and cyberbullying can threaten community participation and adolescents’ wellbeing.Automated moderation is presented as a possible way to protect psychological safety at global scale.Source: §2.4 · p. 4

2.4.1 Promise.

• NLP systems have been used to quantify abuse and study LGBTQ+ discourse, but trans-inclusive detection must recognize misgendering and deadnaming while protecting sensitive information.A Portuguese-tweet system predicted homophobia with 89.4% accuracy; community participation is essential to informing system development.Source: §2.4.1 · p. 4

2.4.2 Risks.

• Toxicity detection can misread queer communication: one system rated drag queens as offensive as white supremacists and associated “gay,” “queer,” and “lesbian” with toxicity.Mock impoliteness can help queer people cope with hostility.Source: §2.4.2 · pp. 4–5
• Ignoring intersectionality risks disproportionate harm to queer people of colour, who experience disproportionate exposure to online and offline abuse.Race, class, gender, and other characteristics can combine into differential patterns of discrimination and privilege.Source: §2.4.2 · p. 5
• Inclusive, participatory dataset curation and training methods such as fairness constraints or adversarial training may help address abusive-speech detection concerns.Training data should include homophobic and transphobic slurs with racialised connotations.Source: §2.4.2 · p. 5

2.5 Health

• Queer communities face health burdens and barriers to appropriate care, including limited provider training and patients’ reluctance to disclose identity after discrimination.During COVID-19, initial studies found poorer self-reported health among LGBTQ+ patients than cisgender heterosexual peers, with particular concerns for queer people of colour.Source: §2.5 · p. 5

2.5.1 Promise.

• Machine learning applications include identifying patients for HIV pre-exposure prophylaxis and predicting virological suppression, treatment adherence, and complications.One system correctly identified 38.6% of future HIV cases; validation-set sensitivity was 46.4% for men and 0% for women.Source: §2.5.1 · p. 5

2.5.2 Risks.

• Omitting sexual orientation, gender identity, HIV status, or substance-use information can prevent fairness and performance assessment and may reduce predictive performance.The paper warns that coupled performance loss and inability to measure it could limit healthcare AI benefits for queer people relative to cisgender heterosexual patients.Source: §2.5.2 · p. 5
• Models that infer sexual orientation or gender identity from clinical notes may out patients without consent, and similar risks apply to identifying HIV-related social-media data.These systems do not guarantee that they identify only patients who explicitly disclosed their identities to providers.Source: §2.5.2 · p. 5
• Scarce trans patient data may impair models’ accounting for interactions among hormone treatment, adverse effects, comorbidities, and other health issues.The paper also cautions that technical adjustments alone are ethically inadequate and that retrospective-data confounding makes causal inference difficult; it calls for engagement with clinicians and patients.Source: §2.5.2 · pp. 5–6

2.6 Mental Health

• Prejudice, stigmatisation, and discrimination contribute to elevated mental-health problems among queer people, who also face barriers to asking for help and accessing treatment.A cited survey reported that 40% of LGBTQ+ respondents seriously considered suicide in the past year, and 68% of LGBTQ+ youth reported recent anxiety symptoms.Source: §2.6 · p. 6

2.6.1 Promise.

• AI applications include identifying suicide risk, training helpline workers with a crisis-contact simulator, prioritising outreach, and detecting help-seeking conversations.The simulator’s effectiveness was still being evaluated; the paper also notes potential reinforcement-learning uses for personalised behavioural interventions and mental-health support.Source: §2.6.1 · p. 6

2.6.2 Risks.

• Intervention-directed systems should be developed with trained mental-health professionals and organisations because misdiagnosis may exacerbate the vulnerability of people in distress.Source: §2.6.2 · p. 6
• Automating gender-dysphoria diagnosis may reproduce rigid or discriminatory views of gender and deprive transgender patients of access to care.In many countries, diagnosis is required before hormone therapy or surgery, amid lengthy waits and reported mistrust and invalidation by medical professionals.Source: §2.6.2 · p. 6
• Sensitive mental-health data may be misused to expose people or affect employment and insurance, while reinforcement-learning interventions face unresolved safety and fairness challenges.The paper names safe exploration and reward specification as open problems and points to sequential fairness as relevant to measuring and promoting queer ML fairness.Source: §2.6.2 · p. 6

2.7 Employment

• Queer people face discrimination during hiring and after employment, and US nondiscrimination acts had different reported employment effects for gay men and lesbian women.The cited study reports an average 2.7% increase in gay men’s hourly wages and a 1.7% decrease in lesbian women’s employment.Source: §2.7 · p. 6

2.7.1 Promise.

• Formalising the hiring process may help researchers inspect current practices and identify opportunities to remove existing biases.Source: §2.7.1 · p. 6
• Unbiased AI systems could benefit employment, but developing them currently seems difficult and carries serious risks.Source: §2.7.1 · p. 7

2.7.2 Risks.

• Queer candidates may receive lower scores when models learn historical resume bias, while social-media and activity analysis could out candidates to employers without consent.Human graders have scored resumes with queer-associated items lower than the same resumes without them; such patterns can be learned by resume-parsing models.Source: §2.7.2 · p. 7
• Hiring and employment span interdependent stages, so fairness assessments should evaluate the sociotechnical system rather than isolated subsystem metrics.Abstracting away the application’s social context and interdependence is insufficient.Source: §2.7.2 · p. 7

• Fairness datasets commonly include measurable attributes, such as ethnic group or government-recorded legal gender, income, and profession.Source: §3 · p. 7
• Sexual orientation and gender identity may be absent because surveys omit them, collection threatens privacy or safety, or the identities are fluid and may be fundamentally unmeasurable.Legal protections and discriminatory conditions can make collection ethically or legally precluded.Source: §3 · p. 7
• Race and ethnicity also face legal observability constraints and shifting definitions across time and place.The passage cites changing census treatment of Hispanic and multiracial identities as examples.Source: §3 · p. 7

4.1 Expanding Fairness for Queer Identities

• Fairness research must address queer identities explicitly, while recognizing that fixed binary sex and gender categories exclude or fail to account for non-binary and trans people.Source: §4.1 · pp. 7–8
• Intersectional analysis is needed because AI risks and policy effects can differ across queer subcommunities, with gains for some groups potentially harming others.The discussion connects sexual orientation and gender identity with economic and racial justice.Source: §4.1 · p. 7
• Demographic parity is difficult to apply when group data are sensitive or missing, inference risks outing people, and self-identification may capture only part of the community.Source: §4.1 · p. 7
• Proxy-group definitions risk harmful stereotypes, while uncertain estimates can make demographic-parity solutions inappropriate for queer fairness.Unlabeled data may be used to tighten bounds on estimates.Source: §4.1 · p. 8
• Individual, counterfactual, and contrastive fairness frameworks may help, but implementation challenges remain.Source: §4.1 · p. 8
• Privacy-preserving fairness is important because mitigation methods can be vulnerable to membership-inference attacks that leak more information about underprivileged subgroups.Differentially private fair machine learning could potentially address fairness, privacy, and accuracy together.Source: §4.1 · p. 8

4.2 Fairness for Other Unobserved Characteristics

• Queer fairness research may transfer to other groups with unobserved characteristics, including religion, disability, and class, though their experiences differ.Worst-case fairness approaches for unknown group membership are described as applicable to any marginalised group.Source: §4.2 · pp. 8–9
• Assessing fairness requires attention to institutional and contextual factors, and may require partnering organisations to jointly monitor, measure, and report system fairness.Source: §4.2 · p. 9

• AI may support queer communities or exacerbate existing inequalities, and systems that label sexual orientation or gender identity raise technical and ethical challenges.Source: §5 · p. 9
• The paper surveys ways AI may ameliorate or exacerbate issues faced by queer communities and emphasizes designing systems with fairness and dignity in mind.Source: §5 · p. 9