RLRuoxi Li

Ruoxi Li

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Ruoxi Li has a total of 2 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Is the Model Anonymous? Why Generative AI Forces a Deployment-Based Answer under the GDPR

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    Debates under the General Data Protection Regulation (GDPR) often ask a deceptively simple question: is the model anonymous? For generative models deployed across different settings, the answer can change with deployment. The same trained model may later be used through a different interface, by another organisation, or with newly connected data, while the testing and documentation supporting an earlier anonymity conclusion remain unchanged. Evidence can therefore remain technically valid yet become insufficient to support the same conclusion for a later processing operation. This article asks how anonymity should be substantiated for a defined operation, how earlier evidence may continue to support that conclusion as deployment conditions change, and how credible contrary evidence should affect the assessment. Drawing on Article 4(1), Recital 26, Court of Justice of the European Union case law, European Data Protection Board Opinion 28/2024, the Board's 2026 consultation Guidelines on Anonymisation, and model-privacy research, it develops two linked propositions. Deployment-Bounded Anonymisation (DBA) identifies the concrete processing operation for which anonymity is claimed as the unit of proof. Conditional evidentiary reuse explains how earlier evidence carries forward: evidence remains probative for propositions whose material conditions remain stable; changes that create or materially strengthen an identification route call for supplementary evidence directed to that route; and credible contrary evidence calls for reconsideration of the proposition it undermines. The durability of an anonymity conclusion therefore depends on whether its supporting evidence remains applicable to the operation in which that conclusion is invoked.

  • Open Access

    How Value Alignment Reorders Legal Reasons in Hybrid Legal Orders: Insights from Hong Kong Model Development

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    When "legal value alignment" is invoked in relation to generative AI, it is not a neutral promise of safer or more accurate outputs. It is a constitutional move in miniature. It quietly decides what will be treated as law, which values are permitted to matter, what forms of reasoning will be recognized as properly legal, and which sources will be allowed to speak with authority. Drawing on alignment insights from the Hong Kong Generative AI Research and Development Center and the governance practices that surround local model development and deployment, this Article argues that alignment in hybrid legal orders systematically produces a legitimacy illusion. A system can be tuned to sound compliant, coherent, and recognizably legal, while remaining unable to do what legitimacy requires in such settings: resolve conflicts of value and rank authority through institutionally authorized and publicly contestable justification. The unresolved remainder is not eliminated. It is relocated into technical choices that present themselves as mere implementation, including data curation, evaluation thresholds, safety boundaries, and refusal rules. Hong Kong brings this relocation into sharp relief. Common-law reason-giving and precedent preserve uncertainty as a condition of contestability, while national security governance can stabilize decisive premises upstream, compelling alignment to choose among rival interpretive grammars. The Article identifies four structural failure modes, categorical, logical, institutional, and epistemic, and explains how sensitivity labels and imported security taxonomies such as TC260-003 prefigure what may be said before legal reasoning begins. It concludes by theorizing four mechanisms that sustain the legitimacy illusion: contestation converted into defaults, authority laundered through model outputs, compliance aesthetics mistaken for legal justification, and polarized reception.

  • Open Access

    All About Taste: The Age of Cheap Closure in Legal AI

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    Legal judgment depends not only on explicit rules, but also on legal taste: the trained capacity to recognize whether legal work has the right institutional form. Through balance, restraint, proportion, caveat, sequencing, and tone, legal taste makes authority socially legible. That legibility now creates a distinctive vulnerability. Because legal taste appears in reproducible textual forms, generative AI can produce the signs of judgment without preserving the contestable practices that make judgment authoritative. This article calls that rule-of-law risk cheap closure: the premature substitution of acceptable legal form for the objections, uncertainties, alternatives, social context, and responsibility chains that make legal power answerable. It contributes by shifting legal AI scholarship from defective outputs to acceptably competent ones; theorizing legal taste as the interface between tacit knowledge, professional form, and authority; and connecting this account to post-training and alignment. It concludes with contestability-preserving governance across output, process, and attribution.

  • Open Access

    When Legal AI Speaks with Authority: Framing, Refusal, and Source Diversity in Generative AI Legal Answers

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    A chatbot answer about law can be accurate, cautious, and still consequential for public understanding. Direct-answer generative AI does more than retrieve legal information: it selects frames, foregrounds sources, and sets boundaries around controversy. This article audits that problem in a sensitive public-law setting, using Hong Kong's hybrid rule-of-law discourse under One Country, Two Systems as a critical case. We submitted Traditional Chinese prompts across five domains to five generative AI systems - ChatGPT5, Claude Sonnet 4.5, Deepseek-V3.2-Exp, Qwen3-MAX, and HKChat - producing 500 first-turn responses. BGE-M3 embeddings measured Top-1 semantic alignment with seven constructed discourse baselines, while refusal coding and critical discourse analysis examined refusal, proceduralization, citation, hedging, and neutralization. The results reveal a concentrated answer space: 244 responses (48.8%) aligned most closely with an order-mobilizing Hong Kong media baseline, 190 (38.0%) with a Hong Kong institutional-legal baseline, and 63 (12.6%) with a Mainland sovereign-legal baseline. Rights-centered, liberal/critical, mainstream-media, and intensified international critique baselines rarely or never became dominant. Model profiles also diverged: Deepseek-V3.2-Exp was the only model with a Mainland sovereign-legal plurality and the only system producing explicit refusals, while HKChat showed the strongest institutional-legal profile. The findings show why legal AI governance should address not only accuracy and safety, but also framing, refusal accountability, source diversity, and plural presentation.

  • Open Access

    From Privacy Self-Management to Accountable Delegation: Policy Responses to User-Facing Generative AI Agents

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    A user-facing AI agent does not merely answer questions; it may decide what to disclose, which service to choose, and how a person is represented to others. This shift exposes a gap in privacy and AI governance. Privacy policies and notice-and-consent regimes still treat personal data governance largely as disclosure control by individual users, while generative AI agents increasingly mediate decisions, permissions, recommendations, transactions, and reputational signals across platform ecosystems. This article develops a comparative policy analysis of three governance generations: P3P-style machine-readable privacy preferences, notice-and-consent/data protection regimes, and emerging user-facing generative AI agents. The central claim is that generative AI policy must move from privacy self-management to accountable delegation. Existing frameworks, including data protection, AI risk management, platform regulation, interoperability, and information-fiduciary approaches, provide partial tools for agent governance, but they do not fully address delegated authority, agent capture, generated reputational representations, or distributed responsibility. The analysis identifies personal AI agents as a distinct intermediary layer between individuals and platforms and proposes agent-specific extensions to existing frameworks: agent-readable obligations, bounded delegation, conflict controls, agent portability, auditable action logs, contestability, marketplace accountability, and public-interest transparency. The contribution is to reframe personal data governance as the regulation of infrastructures that act, decide, and negotiate in users' names.

  • Open Access

    From Privacy Self-Management to Accountable Delegation: Policy Responses to User-Facing Generative AI Agents

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    A user-facing AI agent does not merely answer questions; it may decide what to disclose, which service to choose, and how a person is represented to others. This shift exposes a gap in privacy and AI governance. Privacy policies and notice-and-consent regimes still treat personal data governance largely as disclosure control by individual users, while generative AI agents increasingly mediate decisions, permissions, recommendations, transactions, and reputational signals across platform ecosystems. This article develops a comparative policy analysis of three governance generations: P3P-style machine-readable privacy preferences, notice-and-consent/data protection regimes, and emerging user-facing generative AI agents. The central claim is that generative AI policy must move from privacy self-management to accountable delegation. Existing frameworks, including data protection, AI risk management, platform regulation, interoperability, and information-fiduciary approaches, provide partial tools for agent governance, but they do not fully address delegated authority, agent capture, generated reputational representations, or distributed responsibility. The analysis identifies personal AI agents as a distinct intermediary layer between individuals and platforms and proposes agent-specific extensions to existing frameworks: agent-readable obligations, bounded delegation, conflict controls, agent portability, auditable action logs, contestability, marketplace accountability, and public-interest transparency. The contribution is to reframe personal data governance as the regulation of infrastructures that act, decide, and negotiate in users' names.

  • Open Access

    When AI Sounds Judicious: How AI Alignment Mimic Procedure While Evading Accountability

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    What if the central risk of AI is not that it misjudges, but that it performs judgment so smoothly that the dispute disappears? In ordinary user interactions, large language models increasingly govern at the boundary of what can be said and how usable an answer can be. They refuse, dilute, rewrite, and redirect in ways that read as measured and responsible. Yet this fluency can function as a substitute for accountability. The system simulates judgment while withholding the institutional features that make judgment contestable. This article calls the pattern simulated justice. Seen from this angle, the familiar fight over classification becomes less illuminating and more diagnostic. Treating the model as a product invites defect and design theories. Treating it as a service emphasizes reliance and professional duty analogies. Treating it as a platform points toward intermediary logics and scalable governance. Each frame tries to locate responsibility when the governing rule contours are hard to see, the reasons are hard to verify, and review paths leave little trace. The article argues that the core failure is procedural, not merely substantive. It proposes a minimal procedural floor for alignment interventions, designed to make response restrictions identifiable, reasons classifiable, and escalation and auditing feasible, so that accountability attaches to decisions rather than to rhetoric.

  • Open Access

    Simulated Justice: How AI Alignment Replaces Conflict with Coherence

    by Ruoxi Li, Sirui Han, Yi-Ke Guo

    What if the danger of legal AI is not that it misjudges, but that it sounds perfectly judicious? As legal value alignment becomes a central goal in the design of large language models (LLMs), these systems increasingly produce outputs that appear lawful, balanced, and procedurally sound. But behind their rhetorical fluency lies a powerful illusion: the simulation of legal judgment without its institutional or political substance. This paper critiques alignment not as a failure, but as a success too well executed, one that replaces legal disagreement with semantic symmetry, and recasts justice as an output format. Drawing on jurisprudential theory and discourse analysis, we show how AI answers flatten normative conflict into technocratic equilibrium, offering “reasonable” positions that obscure power, silence dissent, and domesticate law’s agonistic core. Through examples such as hate speech, surrogacy, and criminal responsibility, we expose how alignment displaces the burden of choice with the comfort of coherence. In response, the paper proposes an anti-alignment jurisprudence, one that values contestation over calibration, and treats law not as a language model’s task, but as a political act that must remain interruptible. In doing so, the paper cautions against the growing role of generative AI as a source of normative authority, not because it gets law wrong, but because it sounds too much like it gets it right.