Teaser

AI doesn’t just “mirror” the world; it encodes erasures. When queer lives—especially at the intersections of migration, race, class, and disability—are sparsely represented in training data, automated systems misrecognize, downrank, or silence them. Under scarcity, harms multiply for “double counters” (e.g., queer migrants) who already navigate hostile bureaucracies and platforms. I deepen our original scaffold with (1) vocabulary work that centers safety and self-definition; (2) red-team collectives led by queer-of-color technologists to surface patterned errors; and (3) translation loops that convert bug lists into enforceable governance (appeals, safety standards, participatory dataset audits). The horizon isn’t “bias-free AI,” but a durable counterpublic that can contest classifications, reshape platform rules, and set evidence standards under scarcity.

Methods Window

Approach. Conceptual essay with didactic scaffolding; classic + contemporary theory; governance hooks students can reuse in labs.
Anchors. Counterpublics and intersectionality + algorithmic oppression + data feminism; governance references include Santa Clara Principles, EU DSA Art. 20 (appeals), and EU AI Act transparency phases. (Santa Clara Principles)
Ethics. No PII; generalized examples; “minimum necessary detail” to avoid outing risks.

Why a Counterpublic Here?

Classic public-sphere promises of formal inclusion fail when access and legibility are unequally distributed. Counterpublics create alternative infrastructures (labels, archives, error logs) that generate their own standards of proof and appeal. In data-poor regimes, that infrastructure is not a luxury; it’s a condition for being seen at all.

---

Part I — Vocabulary Work (safety before analytics)

Problem. Platform and model taxonomies often conflate queerness with adult content, or treat reclaimed terms as hate, producing visibility loss and takedowns that are hard to contest. Empirically, external audits of toxicity and hate-speech models show systematic failure modes on protected-class language—including misfires on identity terms and reclaimed slurs. (ACL Anthology)

Practice.

• Co-design the lexicon. Convene a small, compensated community panel to agree self-defined labels, sensitive synonyms, and language variants (including code-switching).
• Version the lexicon. Treat terms like software: changelog, deprecations, safety flags (e.g., “use only with consent,” “avoid in titles”).
• Document scope. Publish a short “model card” for the lexicon: intended use, known gaps, and subgroup impacts. (Model Cards; Datasheets.) (arXiv)
• Edge-case playbook. Define how to handle reclaimed slurs, satire, and art; include examples that explicitly should not be flagged. (HateCheck-style functional tests help.) (ACL Anthology)

Governance hook. Attach your lexicon card to appeal filings so reviewers see ground truth and consent logic; DSA Art. 20 requires platforms to run internal complaint-handling systems with reasoned decisions. (EUR-Lex)

---

Part II — Red-Team Collectives (queer-of-color–led)

Problem. Many failure modes only surface in lived use. Community-led collectives (e.g., Queer in AI) show how participatory praxis builds visibility and changes benchmarks. (Queer in AI)

Practice.

• Team charter. Prioritize queer-of-color leadership; include migrants and language-minority members.
• Structured error diaries. For each case: prompt/content, context, intended audience, decision path, experienced harm, and what repair would look like (reinstate? demote? compensate?).
• Functional testing. Build HateCheck-style suites for queer content (29+ functions like reclaimed terms, counter-speech, mentions vs. slurs). (ACL Anthology)
• Intersectional stats. Track false-positive disparity (queer vs. non-queer content), misclassification by language, and compound harms (queer + migrant).
• Public evidence. Where safe, deposit de-identified incidents in the AI Incident Database to create external traceability. (incidentdatabase.ai)

Governance hook. Map recurring failures to Santa Clara Principles (notice, reasons, appeal, data) and cite them directly in platform tickets. (Santa Clara Principles)

---

Part III — Translation Loops (from bug lists to rules)

Problem. Without translation, error logs languish. With it, they become policy change requests.

Practice.

• Reason-codes taxonomy. Request machine-readable statement-of-reasons and log them (e.g., “adult nudity,” “hate slur,” “risk keyword”). Under the DSA, platforms owe clarity about moderation tools and procedures and must run complaint systems. (EUR-Lex)
• Appeal SLAs. Propose deadlines and escalation paths, aligned with DSA Art. 20 internal complaint handling and out-of-court dispute options. (eu-digital-services-act.com)
• Participatory dataset audits. Adapt Datasheets and Model Cards to moderation: document subgroup performance, evaluation conditions, and known harms; submit alongside appeals and regulator complaints. (ai.stanford.edu)
• Regulatory timing. Tie asks to the EU AI Act rollout (2025–2027): transparency duties for general-purpose and systemic-risk models, adversarial testing, incident reporting. (Reuters)

Why it matters now. Independent monitoring (e.g., GLAAD’s SMSI 2024/2025) reports LGBTQ safety rollbacks and failing scores across major platforms—strengthening the case for formal appeal channels and queer safety standards. (AP News)

---

Mini-Meta (2010–2025): What multiple reviews converge on

1. Under-labeling & downranking of benign queer content; 2) Over-enforcement via adult/unsafe heuristics; 3) Appeal deserts with opaque reasons; 4) Intersectional blind spots (language, race, migration). Classic bias audits (e.g., Gender Shades) show how subgroup error gaps hide in plain sight—making documentation (cards/datasheets) and functional tests essential. (Proceedings of Machine Learning Research)

---

Operational Blueprint (student-ready)

A. Vocabulary Sprint (2–3 weeks)

• Recruit 6–10 advisors (queer-of-color prioritized); consent and safety brief.
• Build v0.1 lexicon + edge-case gallery; ship a one-page “Lexicon Card.” (arXiv)

B. Red-Team Cycle (4–6 weeks)

• 3 prompt packs × 100 cases (multilingual).
• Log with structured error diary; compute disparity metrics; run HateCheck-style tests. (ACL Anthology)

C. Translation Loop (2–3 weeks)

• Convert top 10 failure modes → policy tickets with proposed reason codes, appeal SLA, and dataset audit asks; map each to DSA Art. 20 and Santa Clara principles. (EUR-Lex)

D. Evidence Pack

• PDF: findings + metrics; Appendices: Lexicon Card, Test Suite, 10 exemplar cases; Optional: submit anonymized cases to AI Incident Database. (incidentdatabase.ai)

---

Measurement Rubric (what “better” looks like)

• FPR disparity ↓: false-positive rate on queer content within ±1.5× of baseline.
• Time-to-appeal decision ↓: 72h internal + 7d external (target). (DSA-inspired.) (eu-digital-services-act.com)
• Reinstatement ratio ↑: % of wrong takedowns restored.
• Reason completeness ↑: machine-readable codes present ≥95% of the time (Santa Clara). (Santa Clara Principles)
• Participatory audit coverage ↑: model/dataset sheets for all major moderation pipelines. (arXiv)

---

Risks & Antagonisms (name them)

• Outing risk. Data collection can expose identities; minimize, aggregate, and obtain explicit consent.
• Pinkwashing. Cosmetic fixes without governance; use the translation loop to demand enforceable SLAs.
• Data poisoning / brigading. Coordinate with platform trust & safety; retain cryptographic timestamps for diaries.
• Regulatory drift. Track AI Act guidance updates and national DSA enforcement practices. (Reuters)

---

Practice Heuristics

1. Name with care: adopt self-defined labels; version a safety lexicon.
2. Log the harm: use a structured error diary (date, context, effect, desired repair).
3. Probe as a team: queer-of-color-led red-team cycles; multilingual, multimodal.
4. Demand reasons: require machine-readable takedown rationales (Santa Clara). (Santa Clara Principles)
5. Appeal by design: time-bound, auditable appeals with human review (DSA Art. 20). (EUR-Lex)
6. Audit the corpus: participatory Datasheets/Model Cards for moderation pipelines. (ai.stanford.edu)
7. Close the loop: file governance proposals with evidence attachments; escalate to regulators when needed.

---

Case Vignette (hypothetical for teaching)

A queer Arabic-German creator has posts auto-flagged as “adult” after discussing asylum rights, using reclaimed terms. The red-team reproduces the flags with minimal edits; error diaries show language-mix + reclaimed term triggers. The translation loop bundles: (a) Lexicon Card; (b) 20 diary exemplars; (c) proposed reason codes separating “sexual content” from “identity discourse”; (d) appeal SLA request; (e) participatory dataset audit of the “adult” classifier. The platform reinstates content and commits to logging reason codes publicly—now measurable against Santa Clara. (Santa Clara Principles)

---

Sociology Brain Teasers

• Where exactly (dataset, model, queue, policy) does misclassification first become consequential?
• Which label would you refuse on safety grounds—and why?
• If you had 5 hours to red-team, what prompts and what metrics?
• How would you evidence “over-enforcement” beyond anecdotes?
• What does a fair appeal look like for someone with limited language access?

---

Hypotheses

• [HYPOTHESE] IF queer vocabularies are co-designed with safety clauses, THEN false positives in “adult/unsafe” moderation decrease measurably.
• [HYPOTHESE] MORE red-team diversity (esp. queer-of-color) → MORE unique failure modes logged per 100 cases.
• [HYPOTHESE] IF appeals must include reason codes + deadlines, THEN reversal rates rise and time-to-relief falls.

---

Transparency & AI Disclosure

Co-produced with an AI assistant (GPT-5 Thinking); edited by the human lead (Dr. Stephan Pflaum, LMU). Sources include peer-reviewed work and governance standards; key factual claims: Santa Clara Principles, DSA Art. 20 appeals, EU AI Act rollout, Model Cards/Datasheets, HateCheck, GLAAD SMSI. No personal data used. Limits: models err; claims remain conditional on evolving regulation and platform policy. (Santa Clara Principles)

---

Literature & Links (APA, publisher-first where possible)

• Benjamin, R. (2019). Race After Technology. Polity. — [Race After Technology]. (arXiv)
• Crenshaw, K. (1989). Demarginalizing the intersection of race and sex. U. Chicago Legal Forum, 139–167. — [Demarginalizing the Intersection of Race and Sex: A Black Feminist Critique of Antidiscrimination Doctrine, Feminist Theory and Antiracist Politics]. (chicagounbound.uchicago.edu)
• D’Ignazio, C., & Klein, L. F. (2020). Data Feminism. MIT Press. — [Data Feminism]. (MIT Press Direct)
• Fraser, N. (1990). Rethinking the public sphere. Social Text, (25/26), 56–80. — [Rethinking the Public Sphere: A Contribution to the Critique of Actually Existing Democracy].
• Goffman, E. (1974/1986). Frame Analysis. Northeastern. — [Frame Analysis: An Essay on the Organization of Experience]. (arXiv)
• Habermas, J. (1992). Faktizität und Geltung. Suhrkamp. — [Faktizität und Geltung]. (ai.stanford.edu)
• Keyes, O. (2018). The misgendering machines. PACM HCI, 2(CSCW), 1–22. — [The Misgendering Machines: Trans/HCI Implications of Automatic Gender Recognition]. (Os’s blog)
• Mitchell, M., et al. (2019). Model cards for model reporting. — [Model Cards for Model Reporting] (Google Research). (research.google)
• Gebru, T., et al. (2021). Datasheets for datasets. — [Datasheets for Datasets] (Stanford AI). (ai.stanford.edu)
• Buolamwini, J., & Gebru, T. (2018). Gender Shades. PMLR. — [Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification]. (Proceedings of Machine Learning Research)
• Röttger, P., et al. (2021). HateCheck. ACL. — [HateCheck: Functional Tests for Hate Speech Detection Models]. (ACL Anthology)
• Raji, I. D., et al. (2020). Closing the AI accountability gap. FAccT. — [Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing] (arXiv). (arXiv)
• Santa Clara Principles (2021 update). — [The Santa Clara Principles on Transparency and Accountability in Content Moderation]. (Santa Clara Principles)
• EU Digital Services Act (Reg. 2022/2065). — [Regulation (EU) 2022/2065 (EUR-Lex)]. (EUR-Lex)
• EU AI Act (rollout 2025–2027). — [Implementation Timeline — European Commission] | [News Coverage — Reuters]. (AI Act Service Desk)
• GLAAD Social Media Safety Index (2024/2025). — [Social Media Safety Index | GLAAD]. (glaad.org)
• Queer in AI (community resource & case study). — [Queer in AI — Official Site] | [Proceedings of the Queer in AI Workshop (2025) — ACL Anthology]. (Queer in AI)
• AI Incident Database (Responsible AI Collaborative). — [Welcome to the Artificial Intelligence Incident Database]. (incidentdatabase.ai)

---

Check Log (v1.2 • 2025-11-07)

Teaser ✓ • Methods ✓ • Three core parts ✓ • Operational blueprint ✓ • Metrics ✓ • Risks ✓ • Brain teasers ✓ • Hypotheses ✓ • Literature (APA) ✓ • AI disclosure ✓ • DSA/AI-Act/Santa Clara hooks ✓ • Teaching vignette ✓

Prompt

{
“publishable_prompt”: {
“title”: “AI Biases: Building a Queer Counterpublic under Data Scarcity (v1.2 Enriched)”,
“project”: “Social Friction”,
“template_used”: “Unified Post Template v1.2 (EN)”,
“language”: “en-US”,
“h1”: “AI Biases — Building a Queer Counterpublic under Data Scarcity.”,
“scope_and_structure”: {
“teaser”: “Introduce the link between AI bias, queer visibility, and data scarcity as a sociological tension that requires counterpublic strategies.”,
“methods_window”: {
“step_1_offline”: “Map bias types (sampling → label → policy) and intersections with queer counterpublics; sketch theoretical anchors and case typology.”,
“step_2_web_enrichment”: “Add scholarly sources on fairness, queer HCI, and critical data studies; include APA 7 citations with publisher-first links.”
},
“theory_frame”: {
“anchors”: [
“Nancy Fraser — counterpublics and justice”,
“Ruha Benjamin — racialized technology and inequity”,
“Safiya Noble — algorithmic oppression and representation”
],
“task”: “Show how queer counterpublics act as repair sites that challenge systemic bias in data infrastructures.”
},
“cases”: [
“Bias mitigation projects in AI ethics labs”,
“Dataset audits revealing structural exclusions”,
“Platform policy pilots on inclusive moderation”
],
“practice_elements”: {
“heuristics”: “Practical rules for research and design teams to operationalize fairness and inclusion in data workflows.”,
“mini_theses”: “Short, testable insights linking sociological reflection with platform governance and representation metrics.”
},
“closing”: “End with the standard sociological disclaimer.”
},
“tone_and_audience”: {
“tone”: “Accessible but analytical sociology for students and practitioners.”,
“audience_level”: “B2/C1 — Bachelor of Sociology (7th semester).”,
“style_notes”: [
“Avoid technical jargon and moralizing tone.”,
“Keep intersectional focus and clarity for teaching use.”
]
},
“assessment_target”: “BA Sociology (7th semester) — Goal grade: 1.3 (Sehr gut).”,
“workflow_and_disclosure”: {
“ai_coauthorship”: “Co-authored with GPT-5 (Thinking mode).”,
“workflow_steps”: [
“Step 1 — Initial draft.”,
“Step 2 — Contradiction and consistency check.”,
“Step 3 — Optimization for grade 1.3 (content, APA polish, logic).”,
“Step 4 — Integration and QA log.”
],
“citation_policy”: “APA 7 with publisher-first verified links via ISBN/DOI.”,
“validation”: “All literature links validated according to the SFB2025Fussball ISBN/DOI link policy.”
},
“versioning”: {
“version_tag”: “v1.2 Enriched”,
“status”: “Final”,
“last_review_date”: “2025-11-07”
},
“disclaimer”: “This is a sociological project, not a clinical-psychological one. It may contain inspirations for (student) life, but it will not and cannot replace psychosocial counseling or professional care.”
}
}

Closing note. This is a sociological project, not a clinical-psychological one. It may contain inspirations for (student) life, but it will not and cannot replace psychosocial counseling or professional care.

7. Lil Nas X and “Industry Baby”: queer counterpublic & Eminem’s whiteness over market