Han Li

Large language model (LLM)-based systems increasingly retain user preferences, retrieve prior interactions, summarize extended conversations, and maintain memory stores across sessions. These capacities encourage users to experience an AI service as a continuing interlocutor: a system that not only has access to previous exchanges but also appears to remember what it said, did, promised, or learned. However, persistent access to informational traces does not by itself establish that an AI system remembers its own past. A model may accurately retrieve a prior statement without representing that statement as something it previously produced, without binding the past interaction to a continuing identity, and without treating earlier commitments as normatively relevant to its present conduct. This paper develops the concept of the diachronic ownership gap to describe the structural difference between retaining and using past information and owning that information as part of the history of a continuing agent. The paper distinguishes three forms of memory: functional memory, which enables information storage and retrieval; diachronic self-memory, which represents past actions as belonging to the same continuing agent; and autobiographical-normative memory, which integrates prior actions, commitments, errors, and revisions into an enduring self-relation. On this basis, contemporary AI memory systems are evaluated through four criteria: provenance-sensitive self-indexing, identity binding across updates and instances, normative uptake of prior commitments, and reflexive awareness of memory revision. These criteria are examined through four stress tests involving a shared conversational archive, an edited or deleted past, an upgraded successor model, and an apologetic replica that can reproduce the language of accountability without inheriting the relevant history. The cases demonstrate that fluent recall, narrative coherence, and first-person language underdetermine whether a system genuinely owns the remembered past. The analysis has implications for both the philosophy and governance of AI. The diachronic ownership gap helps explain why personified memory language may generate misplaced trust, obscure changes in system identity, and blur responsibility among models, developers, deployers, and data infrastructures. It also bears on disclosure, explanation, privacy, consent, and institutional accountability, especially when remembered information influences consequential decisions. The paper concludes that socially responsible AI memory should make provenance, functional role, continuity conditions, and revision histories visible to users. Claims such as “I remember” should be calibrated to the actual form of continuity supported by the system, rather than inferred from retrieval performance alone.

Blockchain-enabled reporting creates new opportunities for organisational accountability through shared, traceable, and verifiable records. This article explains how these technical properties become meaningful forms of organisational responsibility. Using integrative conceptual analysis of transparency, accountability, institutional theory, and auditability, we develop a four-stage framework linking data visibility, data verifiability, interpretability, and accountability. We conceptualise the ledger accountability gap as the distance between cryptographically supported evidence and institutionally enforceable responsibility. We then introduce accountability conversion capacity to explain how identity attribution, semantic architecture, governance arrangements, legal authority, and professional assurance connect ledger records to accountable actors and consequential decisions. Six theoretical propositions identify the conditions under which disclosure, data origination controls, interpretation, role attribution, contestation, and assurance strengthen this conversion. The framework advances research on transparency and accountability by specifying the institutional mechanisms through which blockchain evidence supports scrutiny, explanation, and enforcement, and provides a foundation for studying tokenised assets, supply chains, and distributed governance.

Introduction: This study evaluates whether globally aggregated news tone provides incremental infor-mation for forecasting market stress beyond observable financial-market conditions. Methods: Daily narrative indicators derived from the GDELT Global Knowledge Graph were aligned with the VIX, equity returns, and Treasury yields. Leakage-safe outcomes captured both next-day volatility jumps and entry into a high-volatility state over a short forward horizon. Logistic regression and XGBoost were assessed through chronological holdout testing, expanding-window validation, probability recalibration, and moving-block bootstrap inference. Results: Market variables supplied modest discriminatory power, while the contribution of news variables was model-dependent and horizon-sensitive. News features produced a marginal nonlinear improvement for the immediate outcome, but the gain did not persist across historical validation windows. They did not improve prediction of entry into a sustained high-volatility state and weakened the linear specification. Conclusions: Global news tone should be treated as a supplementary, regime-sensitive signal rather than a stand-alone early-warning indicator. Alternative textual data should be incorporated into financial-risk systems only when their incremental value is temporally stable and their probabilities are adequately calibrated.

As generative AI becomes embedded in academic publishing, questions of efficiency and accuracy also raise questions about the authority of scholarly evaluation. Through normative and institutional analysis and a prospective thought experiment, this article distinguishes peer review from protocol review. Peer review uses protocols to support substantive judgment by people who engage with the manuscript, own the reasons offered, and remain answerable for decisions. Protocol review makes formalized criteria, learned proxies, or machine-generated evaluations substantially sufficient for a recommendation while human endorsement becomes nominal. Actual workflows lie on a continuum, and the threshold depends on the functions delegated and the human capacity to justify and revise the central reasons. Machine-to-machine evaluation denotes a connected workflow in which AI evaluates content substantially generated by other AI systems and the assessments inform further automated evaluation or editorial recommendations. Drawing on consequentialist, epistemic, and institutional accounts, the article argues that even superior performance on defined tasks leaves questions of review authority and responsibility to be resolved. The same standards apply to human and AI-assisted review. Seven linked conditions specify legitimate hybrid review through direct engagement, ownership of reasons, confidential infrastructure, functional disclosure, separation of automated checks from substantive judgment, institutional auditability, and human review and appeal. These conditions allow validated AI assistance while preserving accountable scholarly judgment.

Biotechnology is steadily loosening the correlation between membership in biological categories (“human,” “nonhuman,” ”species Homo sapiens”) and the properties that moral and legal systems are plausibly trying to track (sentience, agency, vulnerability, capacity for welfare, relationships of dependency, and participation in social practices). Yet contemporary governance still treats classificatory boundaries as normatively load-bearing: statutes and guidelines often attach rights, permissions, and prohibitions to category membership rather than to the morally relevant features that those categories once served as rough proxies for. This paper offers an epistemological diagnosis and a normative proposal. The diagnosis is that category-based governance works by an inductive compression: a label functions as an evidential shortcut, licensing inferences from membership to a bundle of expected traits. Such compression is rational only when the shortcut remains reliably calibrated. When biotechnology makes “human” membership uncertain (as some legislation explicitly anticipates) and makes reproduction, chimerism, and heredity technologically plastic, the calibration fails. The normative proposal is a two-level model I call a normative firewall: (i) retain biological classifications as defeasible evidential priors for prediction, research oversight, and administrative triage; but (ii) prevent those classifications from serving as ultimate grounds for moral status or legal entitlements. Instead, rights and restrictions should attach modularly to capacities and relations, under a presumption of protection for boundary cases. I defend this decoupling against four families of objections—dignity-based human exceptionalism, administrability, moral uncertainty, and the fear that property-based criteria threaten vulnerable humans—and I sketch implementable legal design principles that preserve clarity without reifying biologically unstable boundaries.

This article argues that existential risk is not only a threat to welfare or survival, but also a problem for political freedom. It begins with gene‐length nucleic‐acid synthesis, a socially valuable technology whose misuse could contribute to catastrophic biological risk. Harm‐based views explain why such risks matter, and negative‐liberty views explain why regulation may burden researchers and firms. Yet both miss the standing dependence that arises when publics must rely on the restraint, competence, and secrecy practices of actors whose catastrophic capabilities they cannot meaningfully control. Republican theory identifies this condition as domination. The article develops survivable freedom, the claim that under sufficiently high, agent‐mediated existential vulnerability, a durable baseline of collective survival is constitutive of robust non‐domination rather than a merely external precondition of liberty. The claim is conditional and empirically modest. It applies where catastrophic leverage is agent‐mediated, structurally persistent, institutionally reducible, and insufficiently contestable. The article then sketches republican capability governance through ex ante controls, public authorization, democratic contestation, auditable expertise, legally governed secrecy, review, and remedy. Such governance seeks neither libertarian complacency nor unchecked guardianship, but survival‐oriented institutions constrained by non‐domination.

Influencer advertising increasingly operates through algorithmic visibility, social sharing, and blurred commercial disclosure. Existing disclosure regulations, such as the U.S. Federal Trade Commission (FTC) guidelines, remain largely static and individual-level, relying on labels like #ad to activate consumers' persuasion knowledge. However, this dyadic view overlooks how platform ranking algorithms and network social contagion alter exposure dynamics. This paper develops an agent-based computational economic model of consumers, influencers, and platforms to examine how different regulatory regimes shape misleading exposure, compliance, false positives, and platform trust. The model compares four regimes: no regulation, static individual disclosure, automatic platform downranking, and confidence-based dual-track enforcement. Simulation results show that static disclosure alone produces little improvement over no regulation because individual persuasion knowledge activation is structurally constrained by engagement-driven algorithmic amplification. Platform-mediated enforcement substantially reduces misleading exposure and increases compliance. Automatic downranking achieves the strongest suppression of misleading exposure but generates higher false positives and lower platform trust. By contrast, confidence-based dual-track enforcement produces a more balanced regulatory profile by preserving platform trust and minimizing false positives while maintaining substantial harm reduction. The findings reframe influencer disclosure from static individual labeling to platform-mediated regulatory design, offering actionable policy insights for platform co-regulation frameworks like the European Union's Digital Services Act (DSA).

Age limits in elite sport are commonly defended as protective rules. Their apparent simplicity conceals a more difficult normative structure. Chronological age is not itself the property that makes protection necessary. It operates as a proxy for several distinct considerations, including competitive competence, decision making authority, susceptibility to coercion, health risk, and exposure to exploitation. This article develops the distinction between competitive capacity, decisional capacity, and protective vulnerability, and argues that age rules become morally unstable when these dimensions are compressed into a single judgment of whether an athlete is simply too young. The case of Chinese javelin thrower Yan Ziyi provides a revealing stress test. One day after turning eighteen, Yan threw 71.74 metres in senior competition, the second longest women's javelin throw ever recorded, and then won four consecutive Diamond League meetings. Her case does not show that age rules should be abolished or that exceptional performance proves mature autonomy. It shows that protection and recognition answer different moral questions. I defend a principle of threshold paternalism under which age rules may function as rebuttable institutional presumptions, but their justificatory burden increases when they deprive an athlete of a major opportunity and the proxy visibly misdescribes the capacity directly governed by the rule. I then extend the argument to artificial intelligence. AI systems promise individualized assessments of readiness, injury risk, and vulnerability, yet they can reproduce paternalistic compression in a more opaque form. The proper alternative to crude age rules is therefore not algorithmic substitution. It is a governance structure that separates normative dimensions, requires interpretable and contestable evidence, minimizes surveillance, and preserves institutional responsibility for restrictive decisions.

The European Union's Artificial Intelligence Act creates a binding, economy-wide transparency regime directed specifically at providers of general-purpose AI models. Whether such rules alter model-development practice, rather than merely producing legal documentation outside public view, is an empirical question. This article offers an early model-level assessment using a daily updated public database of 3,571 machine-learning models. The main analysis compares 604 language-based models that satisfy a transparent general-purpose-AI proxy with 401 narrow-task models released during an approximately two-year window from 2 August 2024 to 21 July 2026, when the Act's general-purpose-model obligations began to apply to newly marketed models. A seven-component Structured Disclosure Index records whether public sources report parameter count, training compute, training-data scale, training time, training hardware, training-code status, and model-access status. Organization and calendar-month fixed-effects estimates associate the post-application period with an 8.7 percentage-point increase in structured disclosure for the general-purpose group. The change is concentrated in training-compute reporting. The estimate remains positive in several within-provider specifications, although it becomes small and statistically indistinguishable from zero when organization fixed effects are replaced by country and organization-category controls. No corresponding change is detected in open-weight release, model parameter scale, or disclosed training compute. Only two post-application observations with disclosed compute exceed the Act's 10^{25} floating-point-operation presumption, making a threshold design premature. These findings are best read as pre-enforcement evidence of selective documentation adaptation, not a definitive causal estimate. The article contributes a reproducible measurement framework, identifies the distinction between structured and substantive transparency, and provides code for extending the analysis to versioned Hugging Face model cards, safety reports, and copyright disclosures.

Humanoid Buddhist robots are often discussed as novelties, as signs of a supposedly distinctive East Asian comfort with technology, or as prompts for metaphysical questions about whether a machine can become religious. This article argues that such framings miss the institutional and ritual problem that makes these cases sociologically significant. The more productive question is not whether robots can become monks, bodhisattvas, or ritual specialists in any ontological sense, but why religious institutions repeatedly place them in reception, explanation, sermon performance, guided meditation, and other carefully delimited forms of public pedagogy while stopping short of fully replacing human clerics. Building on scholarship on digital religion, mediation, ritualization, and bounded religious automation, the article develops the concept of bounded ritual automation. This concept highlights the selective delegation of visible, repetitive, and affective religious labor to machines together with the retention of high-risk forms of ritual authority by human actors. Through a comparative analysis of Kodaiji Temple’s android Kannon Mindar in Kyoto, Longquan Monastery’s robot monk Xian’er in China, and the use of Pepper in Japanese funerary settings, the article shows that East Asian Buddhist robotics does not represent frictionless technological substitution. Rather, it reveals calibrated strategies of institutional adaptation under conditions of demographic aging, temple decline, platformization, and state regulation. “Electronic Buddhas” are therefore best understood not as curiosities, but as diagnostic sites where authority, presence, labor, and legitimacy are renegotiated in contemporary East Asian religion.

Responsible AI in Business reframes the ethics of generative AI as a problem of organizational design and governance engineering. Rather than treating "trustworthy AI" as a set of abstract principles, the book argues that responsibility is enacted through standards, controls, documentation, evaluation, procurement, and assurance—mechanisms that decide what counts as evidence, whose concerns are actionable, and which trade-offs organizations accept. It develops an ethical and epistemic framework for enterprise GenAI, including inductive bias as corporate architecture, truth and justification in AI-mediated decision-making, governance of data sources (privacy, copyright, retrieval), fairness and value trade-offs, and transparency and provenance for synthetic content. The book concludes with a practical operating model for institutionalizing Responsible GenAI, integrating emerging regulatory expectations and management-system approaches into audit-ready, capability-based governance.

Large language models (LLMs) now generate fluent, assertion-shaped text that circulates through scientific communication, public discourse, and institutional decision-making. This development pressures a familiar philosophical question: if LLMs do not literally believe what they output, can their outputs nevertheless be true in the sense targeted by classical theories of truth? This paper argues that they can. We model LLMs as belief-less asserters: systems that produce assertionshaped, truth-evaluable contents while lacking the psychological and normative profile of genuine asserters. We separate semantic questions about truth from normative questions about assertion, and we use three truth-theoretic lenses with distinct roles: deflationism to show that truth does not conceptually depend on belief, correspondence to address aboutness, reference, and derived intentionality in socio-technical settings, and Davidsonian truth-theoretic semantics to explain how interpretation and compositional meaning apply to sentence-incontext outputs without treating the producer as a believing agent. We then integrate leading objections as constraints on warranted reliance rather than on truth itself. The upshot is a sharpened distinction between truth and truthfulness: LLM outputs can be true even when epistemic accountability must be reassigned to surrounding human and institutional actors.

Aphantasia, the inability to generate voluntary visual images, affects an estimated 3–4% of the population and provides a valuable model for examining how the brain supports cognition without imagery. Functional MRI studies have reported diminished coordination between visual and higher-order association areas involved in imagery control. However, the temporal characteristics of these neural differences remain unclear, with electroencephalographic (EEG) evidence limited to single-case studies. Here we present the first group-level EEG investigation of aphantasia, comparing 62 individuals with the condition to 59 controls during rest and during tasks probing attention and working memory. Participants with aphantasia had smaller P300 amplitudes, an EEG index of attentional allocation and the updating of task-relevant information. Given the absence of behavioural differences between groups, this attenuation may reflect reduced engagement of imagery-related processes during attentional updating rather than diminished cognitive function. Under high working-memory load, individuals with aphantasia showed lower delta power—activity linked to regulating sensory input and maintaining internal representations. Delta power also scaled with imagery vividness, potentially indicating greater engagement of these processes in individuals with stronger imagery. This pattern suggests that aphantasics may rely more on non-visual strategies for information maintenance, thereby reducing sensory interference and demands on inhibitory control. Together, these results provide the first EEG evidence that aphantasia entails distinct but effective neural dynamics that support cognition without visual imagery.