Research on model collapse has shown that generative AI models degrade when recursively trained on their own outputs, losing the tails of their distributions. That work addresses one pathway, machines degrading machines, and leaves the human side unexamined. This paper distinguishes two components of the human-side pathway: one internal to the interaction between a single user and a single system, and one environment-mediated, theorised here as stigmergic convergence: the process by which LLM outputs accumulate on the web and alter the knowledge base of non-users as well as users, through search results and ambient environments. Where the human-side and machine-side pathways proceed together, spiral detection failure would arise: the capacity to detect such change is itself degraded. Proposed countermeasures (real-data mixing, strategic human selection, provenance tracking) presuppose the integrity of that capacity and the non-contamination of reference data, neither of which can be structurally guaranteed. A four-layer closed-loop model is developed in which Layer 0 (RLHF preference convergence) already operates, while Layers 1, 2 and 2ʹ (cognitive drift, already-shaped human data entering training corpora, and detection tools and evaluators forming within that environment) are advanced as a hypothesised ordering, not an established result. Within this structure, the independence of any reference point for judging interventions must be confirmed. Set alongside the accountability vacuum described by Medsker and Virmani, this identifies a prior condition. Accountability operates on harms already detected and classified as harms, and it is that prior capacity the loop erodes. Beneath the accountability vacuum lies a detection vacuum.
Artificial Intelligence (AI) and the proliferation of conversational agents have brought transformative changes in education but pose challenges in ensuring the academic integrity of assessments. This article proposes ‘affirmation vivas’ as a potent response to reasonable doubts about the authorship of assessed student work. It draws on a case-study of the first affirmation viva conducted at the University of Glasgow, conducted after ChatGPT “confessed” to generating part of a student dissertation. During the viva, while acknowledging experimenting with AI chatbots, the student successfully demonstrated understanding and ownership of their work. The article reviews practical and ethical issues involved in responding to suspected AI misuse and develops a principled framework for affirmation vivas. We conceptualise these vivas as a means of affirming student understanding rather than detecting AI misuse or presuming misconduct. We derive five principles for their fair and proportionate implementation, including a high threshold for convening them, a presumption of innocence throughout the process and for marking to take place only after the viva, avoiding biases that derive from assuming a student’s guilt.
Being told that an algorithm decided is not the same as being able to change what it decided. This article argues that people subject to consequential automated decisions are owed an effective opportunity to change them, on grounds of justification, non-domination and procedural justice, and that this opportunity exists only when seven guarantees are present for the individual decision: notice, reasons, access to the data used, correction, appeal, human review and remedy authority. It measures this chain in 28 legal instruments and 117,460 privacy policies with PROVE, a transparent provision extractor that separates disclosure of automated decision-making (ADM) from denials and rights recitals (held-out micro-F1 against human coding 0.65; keyword matching 0.23). Of 4303 policies that mention ADM, 259 disclose using it. Among these, 51
The rise of agentic AI and so-called AI companions introduces a shift in the modes of interactions between humans and machines. The exact contours of this shift remain unclear, but early signs of social, affective, and even romantic engagements demonstrate areas of potential benefit and potent harms to individuals and to the broader social fabric. A full appreciation of these benefits and harms demands a multi-perspectival account not only of technological innovation and emerging modalities of human-machine interaction, but also of the historical, socio-cultural, psychological, and ethical aspects of companionship. This essay provides such as account through a pragmatist perspective that considers the material, institutional, and psycho-social conditions behind the rise of AI. From this perspective, the essay broadens the notion of “alignment” as one between means and ends rather than the commonly accepted formulation between design goals and outputs. We argue that AI companionship blurs the boundaries between tools and social actors—a dynamic that emerges from the interplay of technical architectures, intentional design practices, user psychology, and political-economic forces. Given that interplay, a set of interrelated questions arise, which we address in three clusters (supply, demand, mediation), highlighting the emerging social and ethical dilemmas, and suggesting pragmatic interventions to inform public discourse and policymaking. In so doing, we hope to curtail the tremendous fallouts that have become evident in other domains such as social media only in retrospect, and to contribute to the creation of an environment conducive to social cohesion and individual wellbeing.
The rapid adoption of Generative AI (GenAI) has exposed a critical gap in existing AI governance literature: the absence of an integrated model explaining how leaders and organizations build the decision-making capacities needed to enact ethical AI governance in practice. While frameworks such as the NIST AI Risk Management Framework and OECD AI Principles provide essential governance infrastructure, they do not address how leaders develop the judgment, adaptive capacity, and organizational learning required to act on that infrastructure. This paper addresses this gap by proposing a framework that integrates leadership capability, ethical reasoning, and organizational learning for AI governance across sectors. This research employs a conceptual synthesis approach to develop CALLOSSE (Complex Adaptive Leadership, Learning Organization, Sociotechnical Systems, and Ethics), drawing on Sociotechnical Systems Theory, Adaptive Leadership extended through DeRue’s complex adaptive process model, and Organizational Learning, with Ethics positioned as the foundational boundary condition governing all framework decisions. The study also adapts the Hourglass Model for AI governance [28] as the institutional architecture within which CALLOSSE operates, adding a Feedback and Learning layer, establishing a dual-pathway mechanism for routine operations and risk-triggered governance responses. CALLOSSE makes four specific contributions. First, it positions ethics as an active governance decision tool with a prioritized conflict-resolution logic rather than a compliance backdrop. Second, it applies Sociotechnical Systems Theory to diagnose how AI reshapes human-technology relationships, enabling leaders to understand consequences rather than only assign accountability. Third, it uses complex adaptive leadership to navigate the value conflicts and adaptive challenges that policy alone cannot resolve. Fourth, it embeds double-loop organizational learning as an endogenous governance mechanism, enabling organizations to revise the assumptions behind their AI policies over time. The four lenses interact through constraint, design translation, mediation, and institutionalization across four stages: Assessment, Design, Implementation, and Learning and Evaluation. CALLOSSE bridges governance architecture and leadership capability, supporting the development of calibrated trust—leaders’ capacity to assess when AI recommendations warrant acceptance and when human judgment should take precedence. This research contributes to ongoing dialogue on algorithmic accountability and offers a cross-sector, empirically testable approach for future research into values-based AI governance and leadership education.
