The Methodological Framework: Following the GlitchPipeline StageComputational AssumptionThe Manifested “Glitch"Critical & Tactical ResponseOperationalizationHuman concepts can be mapped cleanly into discrete, computable taxonomies.Complex subjectivities break down; binary grids misclassify and erase non-normative experience.Exposing Epistemic Violence: Tracing how classification systems inherit colonial and heterocisnormative assumptions.Latent ModelingHigh-dimensional embeddings neutrally represent semantic similarity.Latent geometries encode historical power asymmetries, stereotyping, and docility regimes.Latent Space Deconstruction: Auditing the unprompted aesthetic and semantic biases of generative models.Inference & AlignmentAlignment algorithms produce safe, objective, and neutral conversational personas.Models hallucinate rigid human and machinic subjects to enforce coercive gendered coherence.Queer Refusal & Ungovernability: Strategically breaking model predictability via auto-experimentation and opacity.Intervention / AI4SGApplying AI to detect harm inherently protects vulnerable populations.The Protection Paradox: Interventions amplify exposure, surveillance, and narrative fixing.Reflexive Practice: Shifting focus from “vulnerable subjects” to the ethics of “vulnerabilizing practices.”
Methodological Stance: “Following the Glitch”#
I am especially interested in how these operationalizations are used to enact and echo power dynamics, including gendered and colonial dynamics, and how they get embedded in systems of governance (like content moderation) and increasingly surveillance.
| Computational Assumption | Manifested “Glitch” | Critical & Tactical Response |
|---|---|---|
| Human concepts map cleanly into discrete, computable taxonomies. | Complex subjectivities break down; binary grids misclassify non-normative experience. | Exposing Epistemic Violence: Tracing how classification systems inherit colonial and heterocisnormative assumptions. |
| High-dimensional embeddings neutrally represent semantic proximity. | Latent geometries encode historical power asymmetries and docility regimes. | Latent Space Deconstruction: Auditing the unprompted aesthetic and semantic biases of generative models. |
| Alignment algorithms produce safe, objective, and neutral conversational personas. | Models hallucinate rigid human and machinic subjects to enforce coercive gendered coherence. | Queer Refusal & Ungovernability: Strategically breaking model predictability via opacity (Glissant) and fugitivity (Bey). |
| Applying AI to detect harm inherently protects vulnerable populations. | The Protection Paradox: Interventions amplify exposure, surveillance, and narrative fixing. | Reflexive Practice: Shifting focus from “vulnerable subjects” to the ethics of “vulnerabilizing practices.” |