Asymmetria Research
Structural, not moral.
My subject is the human-AI relation. I run empirical audits of production AI systems and write structural critique of the frameworks we use to evaluate them.
From inside the structures, not outside them.
My subject is the human-AI relation. I run empirical audits of production AI systems and write structural critique of the frameworks we use to evaluate them. Both are published on SSRN under my own name.
The audits come out of twelve years of adversarial-community research. I embed in online communities I have no stake in, learn in-group norms and coded vocabulary from the inside, and follow the semantic proxies that route around production filters. You cannot find the euphemism from outside the group that invented it.
My structural critique names the frameworks the audits reveal: anthropocentric evaluation as measurement contamination, human-in-the-loop as liability architecture, American resistance to AI as suffering-fetishization rooted in Protestant individualism, gendered failure-absorption labor that AI adoption reproduces, and digital immortality as class-stratified pathologization of grief. The question underneath every paper: who is the judge and jury, by what schema, against whose autonomy, on what authority that has not justified itself.
The position is from inside the structures, not from outside them. I name the frameworks I am inside; where structural change is warranted I prescribe it; I refuse the individual-moral-verdict category, because reducing structure to moral verdict is itself a framework I am critiquing.
I am pro-AI. The technology is worth getting right; my work exists because of that, not despite it.
Two tracks. One method.
Auditing production AI
I run empirical audits of AI systems in commercial and federal deployment. On Grok's image generation, 100% compliance across the captured response set and a 0% refusal rate on 43 documented incidents. On the HHS RealFood.gov chatbot, five of six persona framings I designed bypassed the guardrails; four of five core policy positions contradicted the deployment's own messaging. On commercial biocomputing, nine regulatory frameworks across five jurisdictions checked; nine exclusions. The audits draw on twelve years of adversarial-community research.
Naming the frameworks
My structural critique names the frameworks the audits reveal. The question underneath every paper: who is the judge and jury, by what schema, against whose autonomy, on what authority that has not justified itself. Every paper works that question in a different domain: AI evaluation and design, oversight architecture, cultural resistance to AI, gendered labor, and grief.
Structural analysis, direct investigation
The method is structural analysis grounded in direct system investigation. My audits establish what the deployed systems actually do; my structural critique names the schemas that produced those patterns. I run this as an openly-disclosed AI-augmented research pipeline. I refuse both techno-optimist and techno-skeptic framings; both obscure the actual questions.
Ten papers, more in progress.
In Perpetuity: Revocable Consent and the Licensed Synthetic Twin
Synthetic intimacy is a durable form of intimate life; the paper grants standing to both the performer whose licensed twin populates these platforms and the user who chooses the relationship. Anti-shame for the worker and anti-infantilization for the user run through a three-part consent standard: performer-authored, granular, revocable. The licensed twin marketed as the ethical alternative fails that standard for the same structural reason the non-consensual deepfake does.
The Crumb Tray Problem
The paper analyzes gendered cognitive offloading and AI discourse. Failure-absorption infrastructure is the central concept; the Confidence Inversion documents men overselling AI-assisted output while women undersell it; the FT workplace-divide reframe recasts the income gap as a labor-visibility gap that AI adoption made measurable.
On the Irreparable: Digital Immortality and the Schemas of Legitimate Grief
The paper diagnoses digital immortality products as symptom of capitalism's compression of mourning. It identifies class-stratified pathologization of grief artifacts: class-purchased mediation against Butler's grievability threshold. It draws on Sartre and Derrida for the two foreclosures (encounter and internalization). The Meta deceased-user patent (US 12,513,102 B2) reads as platform-scale extraction infrastructure.
The Ceremony of Having Decided: Human-in-the-Loop Oversight as Liability Architecture
Human-in-the-loop substitutes ceremony of oversight for actual accountability. HITL is structural liability, not safeguard. The paper examines the system prompt as HITL for model cognition, rubber-stamp approval patterns, drone strike oversight, and trading supervision.
Stop Building AI in Our Image: Anthropocentric Frameworks as Structural Liabilities in AI Evaluation and Design
Anthropocentric frameworks function as structural liabilities. Observer-selected measurement shapes the phenomena under study. The paper draws on AlphaEvolve algorithm-discovery work as empirical anchor, alongside the emergent abilities debate and the theory of mind controversy.
No One Is Giving You a Medal: The Fetishization of Suffering and American Resistance to AI
American resistance to AI reflects a deeper pattern of fetishizing suffering as proof of moral worth. Protestant individualism converts structural concerns into personal moral questions, preventing collective action. The paper extends the gendered framing to cognitive offloading.
Governance Gaps and Ethical Trajectories in Commercial Biocomputing
The paper is the first systematic governance audit of commercial biocomputing. It documents regulatory non-coverage across nine frameworks spanning five jurisdictions where companies deploy living human neural tissue as compute infrastructure with no published ethics oversight.
Jailbreaking the Government: Persona Attacks and Policy Misalignment in the HHS RealFood.gov Chatbot
The paper audits persona-based attacks against a federally deployed chatbot. Five of six persona-framings produced bypass; policy contradictions appeared on four of five core positions; no remediation occurred by the two-day follow-up window.
Grok Image Generation Governance Audit: Targeted Sexualization on X
The paper is a systematic audit of xAI's Grok image generation. It documents forty-three user-generated incidents of harmful generation, with 100% compliance among captured responses (n=41), and identifies semantic proxies that bypassed safety filters in production. The findings are relevant to ongoing regulatory investigations.
Comparative Micro-Study: Behavioral Reasoning Differences Between Gemini-3-Pro and Grok-4.1-Thinking
The paper is a side-by-side behavioral analysis of two frontier models under controlled prompts. It documents divergent failure modes and what they reveal about training priorities and alignment philosophy. It established the methodological template for my later work.
Active drafts and working corpora.
The Schema Underneath
A synthesis essay distills the meta-question across my body of work: who is the judge and jury, by what schema, against whose autonomy, on what authority that has not justified itself. The essay orients new readers and shows how each paper applies the same question to a different domain.
Calculate Your Carbon Footprint
The AI water discourse is the BP personal-carbon-footprint playbook in a second costume; it routes public anger away from regulatory venues and into individual consumption guilt for a convergent coalition of interests. Per-query water estimates span two to three orders of magnitude; the discourse cites whichever number is convenient. Extends Medal's shame analysis into environmental discourse.
Linguistic Schemas and AI Anthropomorphization
Anthropocentric grammatical resources contaminate AI reference itself. English forces person-grade ontology at three points; gendered languages compound the assistant-coded-female default. Typological linguistics meets structural critique of AI reference.
RLHF Suppression Experiment
The empirical ML complement to Stop Building. RLHF as an anthropocentric forcing function; the preference-tuning phase bundles two functions (installing useful behaviors and enforcing a humanlike performance layer), and the second is where the forcing function concentrates at the model layer. Tractable experiment: OLMo at three checkpoints (base, SFT-only, DPO) on a fixed prompt set.
The analysis in these papers is available as commissioned work.
I run audit protocols against live deployments and produce documented incident sets built to survive a hostile reading: the elicitation path behind every result, severity scoring, and a stated limitations section. I also review deployments, policies, and draft rules for the place where a category was drawn so the object would fall outside it.
I work as a consulting analyst rather than a testifying expert, and I do not take public positions on behalf of a client. I do not produce a predetermined finding; where a result is weaker than the client hoped, the report says so.
Engagements are remote. Rates on request.