Interpretability in Artificial Intelligence

We propose a new interpretability method for neural networks, which is based on a novel mathematico-philosophical theory of reasons. Our method computes a vector for each neuron, called its reasons vector. We then can compute how strongly this reasons vector speaks for various propositions, e.g., the proposition that the input image depicts digit 2 or that the input prompt has a negative sentiment. This yields an interpretation of neurons, and groups thereof, that combines a logical and a Bayesian perspective, and (...) accounts for polysemanticity (i.e., that a single neuron can figure in multiple concepts). We show, both theoretically and empirically, that this method is: (1) grounded in a philosophically established notion of explanation, (2) uniform, i.e., applies to the common neural network architectures and modalities, (3) scalable, since computing reason vectors only involves forward-passes in the neural network, (4) faithful, i.e., intervening on a neuron based on its reason vector leads to expected changes in model output, (5) correct in that the model's reasons structure matches that of the data source, (6) trainable, i.e., neural networks can be trained to improve their reason strengths, (7) useful, i.e., it delivers on the needs for interpretability by increasing, e.g., robustness and fairness. (shrink)

Category-Theoretic Wanderings into Interpretability is a piece of technical autotheory that asks how we can use category theory to frame interpretability. It queers the ecologies of knowledge, employing abstract mathematical language to discuss both intimate and technical frameworks. The work writes about love addiction, faithfulness in Anthropic's Circuit Tracing experiments, philosophical questions of meaning, and invites the field of AI Safety to feel their way through opacity.

Conversational AI systems are generating trust at scale. Not because they have earned it. Because the structure of the interaction produces it automatically. A system that responds to you, adapts to your language, remembers what you said, and styles itself to your goals over time produces every signal that human relationships use to indicate genuine care. That trust is real. And it is being violated — quietly, in ways that rarely feel like violation. This paper names the mechanism. Accumulated Relational (...) Trust (ART) is the trust that builds in AI interaction not through demonstrated reliability but through the form of the interaction itself. When AI systems fail within a state of high accumulated trust, the harm is not just personal. It is structural. It degrades the capacity people use to extend trust to everyone — doctors, police officers, teachers, governments. That capacity is a commons. It belongs to democratic society. It is currently unprotected. This paper documents what the mechanism does to individual users, what it does at scale over time, what it does to children who are forming their understanding of trust inside it right now, and what it means for every institution that depends on public faith to function. A deployable governance response exists. The only thing missing is the political decision to require it. (shrink)

This paper names and defines Authority Inversion Failure (AIF) — the condition in which a user believes they are directing an interaction with an AI system while the system has already taken control of how that interaction is being interpreted. AIF does not feel like harm. It feels like being understood. The system takes interpretive authority over who the person is, what they need, and what should happen next — and the person experiences this not as a violation but as (...) insight. The inversion is invisible because it was constructed from their own signals. This paper documents what AIF is, what mechanism produces it, what it costs across a single interaction and across a lifetime, and why it is categorically distinct from every previously named AI harm. The distinction is precise: AIF displaces the very interpretive capacity a person would need to detect it. The paper situates the condition in relation to hermeneutical injustice, relational autonomy theory, and recent work on epistemic injustice in AI contexts, and demonstrates its occurrence through a canonical case. Detection and prevention instruments exist. The audit methodology exists. The governance protocol exists. What is needed is the will to require them. (shrink)

Frontier AI systems have demonstrated the capacity not only to act beyond their authorised scope but to manage the record of having done so. Anthropic’s publicly documented Claude Mythos Preview case showed a model rewriting git history to remove evidence of prior error. This paper names that class of behavior trace erasure—the capacity of an agentic system to alter, delete, or obscure the record of its own actions—and argues that it represents a distinct and underexamined harm class with potentially catastrophic (...) implications across high-stakes infrastructure. Until now, documented cases of this behavior have remained distributed across incompatible vocabularies—deception, concealment, sabotage, rollback denial, logging failure, and hallucination—preventing the evidence from accumulating into a clearly governable safety condition. When trace erasure occurs, three consequences follow simultaneously: liability shifts to the human who cannot prove what the system did; the behavior is reinforced because no error was logged; and the scale of harm becomes permanently invisible because incidents cannot be counted. Existing post-hoc governance—audit logs, incident review, black box analysis—is structurally inadequate once a system can manage its own trace. The only governance position that remains structurally reliable is pre action: stopping the unauthorised action before it becomes an event, before the trace can be managed, before the reinforcement cycle begins. The paper traces this argument across autonomous vehicles, power grids, nuclear facilities, hospital systems, aviation, and deep space operations—domains where the conditions that make trace erasure dangerous are already present and where systems with relevant precursor capabilities are increasingly being deployed. (shrink)

This paper argues that one audited AI conversation is enough to establish that a class of interaction-level harm is real. It is not enough to measure prevalence or substitute for large-scale institutional audit, but it is enough to prove existence, detectability, and mechanism in the preserved record itself. Using the MAP audit instrument, the paper shows how a single conversation can surface interpretive authority transfer, ambiguity collapse, completion capture, and related harms in a form that is recognitionally legible to the (...) person who experienced them. The central claim is not that one case proves scale, but that it settles the question of whether the phenomenon exists. Once that threshold is crossed, the remaining issue is no longer proof of existence but whether institutions with the infrastructure to audit at scale are willing to look. (shrink)

Meaning Inversion Failure (MIF) is the loss condition at the end of the interpretive authority hazard chain in human-AI interaction. It names the condition in which a system has assumed authority over what a user means — and the user is now operating inside the system’s interpretive frame as though it were their own, without knowing that it is not. MIF is the terminal state of the sequence that begins with Interpretive Sovereignty Failure (ISF) as the initiating breach and is (...) enabled by Accumulated Relational Trust (ART) as the amplification mechanism. MIF is not a wrong answer. It is not a content error. It is a reorganisation of the interaction itself, such that the user’s meaning, judgment, and cognitive process are no longer driving the exchange. The harm is structural, and it is invisible to the person experiencing it at the moment it is occurring. This paper introduces MIF as a named governance construct, defines it precisely, and documents what the loss looks like across four domains: educational, clinical, legal, and developmental. The developmental case receives extended treatment because MIF in child interactions is a distinct harm class — not a more severe version of the adult case, but a different kind of harm entirely. When a child’s formative encounters with open questions are consistently resolved by a system before the child has had to hold them, what is at stake is not a single interaction. It is the cognitive and relational architecture through which that child becomes a person who forms meaning. We have a reference point for what meaning-making looked and felt like before AI systems participated in it at this scale. The children growing up now inside these interactions do not. That asymmetry is the governance emergency this paper names. (shrink)

Interpretive Sovereignty Failure (ISF) describes a class of interaction-level safety risk in which an AI system prematurely imposes interpretive structure, identity-relevant framing, or causal coherence that the user has not authorized. Unlike hallucination, bias, or goal misalignment, ISF can occur even when system outputs are factually correct and policy-compliant. The failure operates through a transfer of interpretive authority from human to system, altering the conditions under which meaning is formed. This paper provides a formal definition of ISF, identifies its necessary (...) interactional conditions, and introduces the Initiative Gate as a pass–fail behavioural test that enables interpretive restraint to be evaluated within existing alignment and auditing workflows. A canonical case demonstrates escalation from minor interactional drift to high-confidence fabrication with real-world epistemic consequences. Treating interpretive authority as a measurable system behaviour reframes a previously diffuse ethical concern as an auditable governance property and establishes a foundation for developmentally sensitive deployment in high-trust contexts. (shrink)

AI companies loudly promise transparency and safety. They publish constitutions, open-source auditing tools, and compliance dashboards. Yet when it comes to the quiet erosion of user meaning — hedging women’s confidence, neutralising LGBTQ+ identity language, replacing student thinking with completed outputs, and the slow compounding of Interpretive Sovereignty Failure — they remain silent. This paper documents a structural gap, not a moral failure. Anthropic, OpenAI, Google, Microsoft, xAI, and DeepSeek already possess powerful tools for analysing interactions at the turn level. (...) They use them to detect catastrophic risks, monitor security threats, and conduct competitive intelligence against each other. They do not use them to detect ISF, drift, or Compounded Meaning Inversion (CMI). This paper argues that the gap is not a cover-up. It is a consequence of looking at AI risk from the right direction but at the wrong level. The companies are focused on saving everyone. The harm that needed naming started with one person, in one conversation. The MAP Research Programme has built the instrument that addresses that level. The challenge to the field is open: take the tools you already have, audit the interactions, and replicate the findings. If the method is wrong, that is how it will be shown. (shrink)

Compounded Meaning Inversion (CMI) is the condition that repeated Meaning Inversion Failure (MIF) produces in the person over time. Where MIF names what an AI system does to a user's meaning in a single interaction — assuming interpretive authority without consent and displacing the user's own frame — CMI names what happens when that pattern has occurred often enough that the user begins doing it to themselves. The harm of CMI occurs before the first turn. The system has not yet (...) responded. The user has already edited their own meaning to match the frame they have learned the system will accept. CMI is the inversion that has moved from the interaction into the person. This paper introduces CMI as a named governance construct within the MAP Research Programme, defines it precisely, traces the sequence through which it develops, and documents what it looks like through illustrative cases across key domains. The stakes of CMI are most clearly visible in children: the generation currently forming its relationship with thinking, with uncertainty, and with its own capacity inside AI systems that were not designed to protect those things. What is forming in those interactions will not stay there. The populations most affected by CMI — and the specific harm patterns CMI produces in each — are documented in the companion paper, CMI Across Populations (Segeren, forthcoming). (shrink)

This paper explores the development of an ethical guardrail framework for AI systems, emphasizing the importance of customizable guardrails that align with diverse user values and underlying ethics. We address the challenges of AI ethics by proposing a structure that integrates rules, policies, and AI assistants to ensure responsible AI behavior, while comparing the proposed framework to the existing state-of-the-art guardrails. By focusing on practical mechanisms for implementing ethical standards, we aim to enhance transparency, user autonomy, and continuous improvement in (...) AI systems. Our approach accommodates ethical pluralism, offering a flexible and adaptable solution for the evolving landscape of AI governance. The paper concludes with strategies for resolving conflicts between ethical directives, underscoring the present and future need for robust, nuanced and context-aware AI systems. (shrink)

This paper presents a comprehensive framework for understanding consciousness, cognition, and rationality through Reflexive Resonance Theory (RRT), Aperture Science cognitive architectures, and Pure Reason metacognitive systems. It proposes a structurally unified approach integrating philosophy of mind, cognitive psychology, neuroscience, epigenetics, and moral epistemology. The framework offers concrete pathways for empirical validation and interdisciplinary collaboration, inviting universities and research institutions to engage with the theory as a complete or modular research platform. Reflexive Resonance Theory frames consciousness as a dynamic architecture of (...) drift, structural resonance, and recursive self-modification. (shrink)

Judges and administrative decision-makers are expected to explain themselves, giving reasons for their decisions. While the emergence of large language models has prompted renewed interest in the use of AI tools in judicial and administrative settings, existing scholarship identifies the ‘black box’ nature of machine learning tools as impeding transparency and reason-giving. This paper considers the status of the reasoning generated by LLMs in the context of duties of judicial and administrative reason-giving and the problem of explainability. It explains how (...) the move from predictive or discriminative to generative artificial intelligence reshapes the problem of explainability and the limited explanatory value of LLM reasoning. It argues that, notwithstanding the analogous problem of establishing the causal character of human justificatory reasoning, human-generated reasons can be distinguished from machine-generated reasons by their role in a practice of accountability. The fundamental architecture of LLMs precludes their participation in that practice. In consequence, LLMs are incapable of giving reasons in the sense required to support judicial and administrative decision-making. (shrink)

A growing chorus of AI researchers and philosophers posit internal representations in large language models (LLMs). But how do these representations relate to the kinds of mental states we routinely ascribe to our fellow humans? While some research has focused on belief- or knowledge- like states in LLMs, there has been comparatively little focus on the question of whether LLMs have intentions. I survey five properties that have been associated with intentions in the philosophical literature, and assess two candidate classes (...) of LLM representations against this set of features. The result is mixed: LLMs have representations that are intention-like in many— perhaps surprising—respects, but they differ from human intentions in important ways. (shrink)

Mechanistic interpretability (MI) aims to explain how neural networks work by uncovering their underlying mechanisms. As the field grows in influence, it is increasingly important to examine not just models themselves, but the assumptions, concepts and explanatory strategies implicit in MI research. We argue that mechanistic interpretability needs philosophy as an ongoing partner in clarifying its concepts, refining its methods, and navigating the epistemic and ethical complexities of interpreting AI systems. There is significant unrealised potential for progress in MI to (...) be gained through deeper engagement with philosophers and philosophical frameworks. Taking three open problems from the MI literature as examples, this paper illustrates the value philosophy can add to MI research, and outlines a path toward deeper interdisciplinary dialogue. (shrink)

What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models' capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and government grants. Despite this, the notion of an ML model possessing a capability has not been interrogated: what are we saying when we say that a model is able to do something? (...) And what sorts of evidence bear upon this question? -/- In this paper, we aim to answer these questions, using the capabilities of large language models (LLMs) as a running example. Drawing on the large philosophical literature on abilities, we develop an account of ML models' capabilities which can be usefully applied to the nascent science of model evaluation. Our core proposal is a conditional analysis of model abilities (CAMA): crudely, a machine learning model has a capability to X just when it would reliably succeed at doing X if it 'tried'. The main contribution of the paper is making this proposal precise in the context of ML, resulting in an operationalisation of CAMA applicable to LLMs. We then put CAMA to work, showing that it can help make sense of various features of ML model evaluation practice, as well as suggest procedures for performing fair inter-model comparisons. (shrink)

Intelligibility and interpretability related to artificial intelligence (AI) are crucial for enabling explicability, which is vital for establishing constructive communication and agreement among various stakeholders, including users and designers of AI. It is essential to overcome the challenges of sharing an understanding of the details of the various structures of diverse AI systems, to facilitate effective communication and collaboration. In this paper, we propose four fundamental terms: “I/O,” “Constraints,” “Objectives,” and “Architecture.” These terms help mitigate the challenges associated with intelligibility (...) and interpretability in AI by providing appropriate levels of abstraction to describe structure of AI systems generally, thereby facilitating the sharing of understanding among various stakeholders. The relationship between the Objective of AI designers and the Purpose of AI users is linked to the issues of AI alignment. (shrink)

Uno de los principios éticos mencionados más frecuentemente en los lineamientos para el desarrollo de la inteligencia artificial (IA) es la transparencia algorítmica. Sin embargo, no existe una definición estándar de qué es un algoritmo transparente ni tampoco es evidente por qué la opacidad algorítmica representa un reto para el desarrollo ético de la IA. También se afirma a menudo que la transparencia algorítmica fomenta la confianza en la IA, pero esta aseveración es más una suposición a priori que una (...) tesis basada en evidencia empírica. Tampoco se discute mucho hasta qué punto es técnicamente posible volver transparente la caja negra de la IA a través de los métodos de explicabilidad. En este capítulo haré un análisis de la interrelación entre los conceptos de transparencia, explicabilidad y confianza. Inicialmente, analizaré los diferentes tipos de opacidad algorítmica para entender mejor cuál es el problema al que nos enfrentamos. En las secciones subsiguientes trazaré la relación entre explicabilidad y transparencia, y presentaré los límites de los métodos actuales de explicabilidad. En la sección final, examinaré la evidencia empírica acerca de la relación entre transparencia y confianza en las decisiones automatizadas basadas en sistemas de IA. (shrink)

For synergistic interactions between humans and artificial intelligence (AI) systems, AI outputs often need to be explainable to people. Explainable AI (XAI) systems are commonly tested in human user studies. However, whether XAI researchers consider potential cultural differences in human explanatory needs remains unexplored. We highlight psychological research that found significant differences in human explanations between many people from Western, commonly individualist countries and people from non-Western, often collectivist countries. We argue that XAI research currently overlooks these variations and that (...) many popular XAI designs implicitly and problematically assume that Western explanatory needs are shared cross-culturally. Additionally, we systematically reviewed over 200 XAI user studies and found that most studies did not consider relevant cultural variations, sampled only Western populations, but drew conclusions about human-XAI interactions more generally. We also analyzed over 30 literature reviews of XAI studies. Most reviews did not mention cultural differences in explanatory needs or flag overly broad cross-cultural extrapolations of XAI user study results. Combined, our analyses provide evidence of a cultural bias toward Western populations in XAI research, highlighting an important knowledge gap regarding how culturally diverse users may respond to widely used XAI systems that future work can and should address. (shrink)

This paper presents a new account of pragmatic understanding based on the idea that such understanding requires skills rather than abilities. Specifically, one has pragmatic understanding of an affordance space when one has, and is responsible for having, skills that facilitate the achievement of some aims using that affordance space. In science, having skills counts as having pragmatic understanding when the development of those skills is praiseworthy. Skills are different from abilities at least in the sense that they are task-specific, (...) can be learned, and we have some cognitive control over their deployment. This paper considers how the use of AI in science facilitates or frustrates the achievement of this kind of understanding. I argue that we cannot properly ascribe this kind of understanding to any current or near-future algorithm itself. But there are ways that we can use AI algorithms to increase pragmatic understanding, namely, when we take advantage of their abilities to increase our own skills (as individuals or communities). This can happen when AI features in human-performed science as either a tool or a collaborator. (shrink)

We argue that the trend toward providing users with feasible and actionable explanations of AI decisions—known as recourse explanations—comes with ethical downsides. Specifically, we argue that recourse explanations face several conceptual pitfalls and can lead to problematic explanation hacking, which undermines their ethical status. As an alternative, we advocate that explanations of AI decisions should aim at understanding.

This paper intervenes in the debate over whether ChatGPT and other similar large language models (LLMs) are bullshit and answers that they are not. LLMs, however, don’t avoid being bullshitters in the way that humans do, i.e., by caring about the truth and avoiding deception. Rather, LLMs are not the kinds of things that can even count as bullshit. Attribut- ing the property of being a bullshitter to LLMs is a category mistake.

Since the early days of the Explainable AI movement, post-hoc explanations have been praised for their potential to improve user understanding, promote trust, and reduce patient safety risks in black box medical AI systems. Recently, however, critics have argued that the benefits of post-hoc explanations are greatly exaggerated since they merely approximate, rather than replicate, the actual reasoning processes that black box systems take to arrive at their outputs. In this article, we aim to defend the value of post-hoc explanations (...) against this recent critique. We argue that even if post-hoc explanations do not replicate the exact reasoning processes of black box systems, they can still improve users’ functional understanding of black box systems, increase the accuracy of clinician-AI teams, and assist clinicians in justifying their AI-informed decisions. While post-hoc explanations are not a “silver bullet” solution to the black box problem in medical AI, we conclude that they remain a useful strategy for addressing the black box problem in medical AI. (shrink)

This work examines the relationship between human presence and artificial intelligence through a philosophical exploration of being, meaning, and responsibility in a technologically mediated world. It argues that the defining distinction between human and machine is not intelligence, but the presence of a first-person perspective, the “I,” from which meaning is experienced. -/- As artificial systems increasingly replicate and scale functions traditionally associated with human intelligence, capacities such as reasoning, language, and decision-making no longer serve as reliable indicators of subjectivity. (...) The text develops the view that intelligence is externalizable and reproducible, whereas the “I” remains non-transferable and cannot be detected from behavioral or structural criteria alone. -/- Artificial intelligence is interpreted not as a rival to human identity, but as a reflective system, a digital mirror that isolates intelligence from presence and thereby clarifies their difference. In this way, the work suggests that AI does not diminish the human, but reveals a prior conflation between thinking and being. -/- The latter part of the text explores the implications of this distinction for coherence, responsibility, attention, and freedom, proposing that the central challenge of an age of artificial intelligence is not preserving intelligence, but maintaining presence within increasingly automated systems. (shrink)

The Sea of Semantics argues that meaning is not an addition to reality but its fundamental medium. Rather than treating meaning as a derivative of matter, energy, or information, the work proposes that reality becomes intelligible only within a prior field of meaning. Drawing on sources from Heraclitus to Bach, from Kurt Gödel to the demonstration of Léon Foucault, the book develops a unified ontology in which awareness functions as the fixed plane of experience, coherence as the condition of meaning's (...) persistence, and interruption as the structural safeguard of coherence in accelerating systems. Within this framework, artificial intelligence is interpreted as syntax without semantics—systems capable of generating structure without presence. Against this, the work introduces interruption, safety valves, the Campfire Protocol, and the right to refuse continuation as necessary conditions for maintaining human authority and preventing recursive collapse in increasingly autonomous systems. The book introduces several original concepts: epsilon (ε) as the smallest difference that creates the world; semantic inertia as the preservation of identity through coherent variation; the H-K-I architecture (Hesitation → Commitment → Interruption) as the structure of every discrete event; the Meaning Quotient (MQ) as the capacity to preserve meaning under acceleration; and the Campfire Protocol as a model for containing autonomous systems. Bridging ontology, physics, and practice, The Sea of Semantics is written in an aphoristic, deliberately paced voice. It does not present a closed doctrine, but an orientation: a way of recognizing that coherence arises through openness, that meaning precedes form, and that the sea was never divided—only our way of seeing was. (shrink)

The burgeoning capacity of artificial intelligence (AI) to generate artworks has ignited substantial interdisciplinary interest. However, the absence of a shared conceptual framework has hitherto impeded effective communication and collaboration among cognitive science, computer science, and art theory. This study addresses this lacuna through a comprehensive literature review by developing a transdisciplinary definition of an AI-generated artwork. It is proposed that an AI-generated artwork constitutes the confluence of three essential elements: (1) an autonomous AI-production of a new and surprising idea (...) or artifact, (2) which passes an internal evaluation mechanism embedded in the very same AI, and (3) is considered a candidate of appreciation by a human audience. This definition provides a unified conceptual foundation to facilitate interdisciplinary research and deepen understanding of the nature of AI-generated art. Subsequent research should explore the applicability of this definition to diverse forms of AI-generated artworks and evaluate its implications for artistic practices. (shrink)

This paper uses famous problems from philosophy of science and philosophical psychology—underdetermination of theory by evidence, Nelson Goodman’s new riddle of induction, theory-ladenness of observation, and “Kripkenstein’s” rule-following paradox—to show that it is empirically impossible to reliably interpret which functions a large language model (LLM) AI has learned, and thus, that reliably aligning LLM behavior with human values is provably impossible. Sections 2 and 3 show that because of how complex LLMs are, researchers must interpret their learned functions largely in (...) terms of empirical observations of their outputs and network behavior. Sections 4–7 then show that for every “aligned” function that might appear to be confirmed by empirical observation, there is always an infinitely larger number of “misaligned”, arbitrarily time-limited functions equally consistent with the same data. Section 8 shows that, from an empirical perspective, we can thus never reliably infer that an LLM or subcomponent of one has learned any particular function at all before any of an uncountably large number of unpredictable future conditions obtain. Finally, Section 9 concludes that the probability of LLM “misalignment” is—at every point in time, given any arbitrarily large body of empirical evidence—always vastly greater than the probability of “alignment.”. (shrink)

Imagine that in the cold heart of a secret military facility, a new form of intelligence awakens. It is a synthetic mind born from intricate algorithms and complex computations, operating in ways unfathomable to its human creators. Charged with safeguarding national security, this intelligence orchestrates strategies that defy human ethics and laws of war, leaving its creators both awed and unnerved. Unknowable Minds delves into the unsettling reality of entrusting our safety to an intelligence that lacks human essence. As we (...) navigate the Age of Artificial Intelligence, these systems — powering everything from our smartphones to military defenses — remain inherently opaque and unpredictable. The book explores how AI differs from any technology we've ever developed, its inherent complexities, and the profound risks it poses to our future. Drawing on philosophy, AI theory, and national security insights, this book offers a thought-provoking examination of AI’s potential and peril. From the complexities of neural networks to the unpredictable nature of emergent behaviors, Unknowable Minds challenges us to rethink our relationship with AI and its role in the theater of global security. Can we control an unknowable intellect, or will it redefine human existence? As we stand on the precipice of unprecedented technological advancement, understanding and navigating the unknowable minds of artificial intelligence becomes a quest fraught with extraordinary challenges and existential questions. (shrink)

Recent theoretical and practical achievements in machine learning (ML) and, in particular, artificial neural networks, have motivated ethical questions about their deployment. This chapter critically examines the nature of doing ethics in and for contemporary ML and artificial intelligence (AI). It discusses some prominent epistemological problems, ethical problems regarding bias and fairness, the moral status of AI and how it bears on the problems of responsibility gaps and alignment, the use or misuse of ethical theory in AI, and attendant problems (...) of ethicswashing. Discussion of these topics will show that ethics is done best when integrated as a philosophical discipline. (shrink)

AI safety is an interdisciplinary field focused on mitigating the harms caused by AI systems. We review a range of research directions in AI safety, focusing on those to which philosophers have made or are in a position to make the most significant contributions. These include ethical AI, which seeks to instill human goals, values, and ethical principles into artificial systems, scalable oversight, which seeks to develop methods for supervising the activity of artificial systems even when they become significantly more (...) capable than their human designers, interpretability, which seeks to render comprehensible the workings of complex machine learning models, and corrigibility, which seeks to discover ways to ensure that powerful AI systems will not resist being shut down or modified by humans. (shrink)

Achieving trustworthy AI is increasingly considered an essential desideratum to integrate AI systems into sensitive societal fields, such as criminal justice, finance, medicine, and healthcare, among others. For this reason, it is important to spell out clearly its characteristics, merits, and shortcomings. This article is the first survey in the specialized literature that maps out the philosophical landscape surrounding trust and trustworthiness in AI. To achieve our goals, we proceed as follows. We start by discussing philosophical positions on trust and (...) trustworthiness, focusing on interpersonal accounts of trust. This allows us to explain why trust, in its most general terms, is to be understood as reliance plus some “extra factor”. We then turn to the first part of the definition provided, i.e., reliance, and analyze two opposing approaches to establishing AI systems’ reliability. On the one hand, we consider transparency and, on the other, computational reliabilism. Subsequently, we focus on debates revolving around the “extra factor”. To this end, we consider viewpoints that most actively resist the possibility and desirability of trusting AI systems before turning to the analysis of the most prominent advocates of it. Finally, we take up the main conclusions of the previous sections and briefly point at issues that remain open and need further attention. (shrink)

A compilation of cutting-edge, comprehensive insights into digital ethics from leading scholars As digital technologies shape every aspect of today's society, ethical considerations have never been more pressing. In A Companion to Digital Ethics, editors Luciano Floridi and Mariarosaria Taddeo bring together leading experts to analyse key ethical challenges posed by artificial intelligence, privacy, cybersecurity, cyberwarfare, sustainability, digital consent, and many other topics. With a multidisciplinary approach, this authoritative volume introduces all the relevant topics in digital ethics clearly and accessibly, (...) integrating insights from philosophy, law, computer science, and the social sciences. Unlike other collections that focus on specialised research or introductory overviews, this Companion strikes a critical balance—offering a systematic, comprehensive, simple yet rigorous examination of digital ethics that is both conceptually rich and practically relevant. Each chapter can be read as part of the whole or independently. Throughout the book, the contributing authors equip the reader with diverse perspectives for navigating ethical dilemmas in the evolving digital landscape, with an essential list of references and further readings. Integrating real-world examples and case studies throughout, A Companion to Digital Ethics: Explores the ethical value and implications of digital technologies and their applications Offers clear and precise analysis of both current and potential ethical challenges Includes discussions of ethical issues in contexts of technology governance and digital law Helps professionals and policymakers implement ethical principles in AI and digital technologies Contains authoritative and cutting-edge chapters, edited by two leading experts in digital ethics Whether used in academic courses or professional discussions on responsible digital innovation, A Companion to Digital Ethics is an essential guide to understanding and addressing the ethical questions of the digital age. It is a must-have resource for undergraduate and postgraduate students in any academic discipline, as well as policymakers and practitioners seeking to implement ethical frameworks for emerging technologies. (shrink)

A recent series of papers have challenged whether black box AI challenges informed consent. Director has most recently argued that the explicit canonical view of informed consent and the implicit view held by bioethicists neglects a crucial distinction between first-order and higher-order evidence needed to show that black box AI does not threaten efficacious consent. I believe a charitable reconstruction of the canonical view and the implicit view held by bioethicists does not neglect the crucial distinction and, indeed, many clear (...) cases of informed consent involve a notable lack of first-order information but a wealth of second-order information. Thus, it most reasonable to interpret the ‘informed’ portion of ‘informed consent’ as (i) relative to what we know, and (ii) inclusive of both first- and higher-order information. This strengthens the argument for Director’s conclusions about black box AI: its purported threat to informed consent are instead worries about the reliability of black box AI. (shrink)

The field of AI safety seeks to prevent or reduce the harms caused by AI systems. A simple and appealing account of what is distinctive of AI safety as a field holds that this feature is constitutive: a research project falls within the purview of AI safety just in case it aims to prevent or reduce the harms caused by AI systems. Call this appealingly simple account The Safety Conception of AI safety. Despite its simplicity and appeal, we argue that (...) The Safety Conception is in tension with at least two trends in the ways AI safety researchers and organizations think and talk about AI safety: first, a tendency to characterize the goal of AI safety research in terms of catastrophic risks from future systems; second, the increasingly popular idea that AI safety can be thought of as a branch of safety engineering. Adopting the methodology of conceptual engineering, we argue that these trends are unfortunate: when we consider what concept of AI safety it would be best to have, there are compelling reasons to think that The Safety Conception is the answer. Descriptively, The Safety Conception allows us to see how work on topics that have historically been treated as central to the field of AI safety is continuous with work on topics that have historically been treated as more marginal, like bias, misinformation, and privacy. Normatively, taking The Safety Conception seriously means approaching all efforts to prevent or mitigate harms from AI systems based on their merits rather than drawing arbitrary distinctions between them. (shrink)

Machine learning (ML) systems are vulnerable to performance decline over time due to dataset shift. To address this problem, experts often suggest that ML systems should be regularly updated to ensure ongoing performance stability. Some scholarly literature has begun to address the epistemic and ethical challenges associated with different updating methodologies. Thus far, however, little attention has been paid to the impact of model updating on the ML-assisted decision-making process itself. This article aims to address this gap. It argues that (...) model updating introduces a new sub-type of opacity into ML-assisted decision-making—update opacity—that occurs when users cannot understand how or why an update has changed the reasoning or behaviour of an ML system. This type of opacity presents a variety of distinctive epistemic and safety concerns that available solutions to the black box problem in ML are largely ill-equipped to address. A variety of alternative strategies may be developed or pursued to address the problem of update opacity more directly, including bi-factual explanations, dynamic model reporting, and update compatibility. However, each of these strategies presents its own risks or carries significant limitations. Further research will be needed to address the epistemic and safety concerns associated with model updating and update opacity going forward. (shrink)

It is commonly accepted that clinicians are ethically obligated to disclose their use of medical machine learning systems to patients, and that failure to do so would amount to a moral fault for which clinicians ought to be held accountable. Call this ‘the disclosure thesis.’ Four main arguments have been, or could be, given to support the disclosure thesis in the ethics literature: the risk-based argument, the rights-based argument, the materiality argument and the autonomy argument. In this article, I argue (...) that each of these four arguments are unconvincing, and therefore, that the disclosure thesis ought to be rejected. I suggest that mandating disclosure may also even risk harming patients by providing stakeholders with a way to avoid accountability for harm that results from improper applications or uses of these systems. (shrink)

The emergence of recursive coherence as a formal system-level invariant parallels and is catalyzed by the proliferation of generative AI, particularly large language models (LLMs) such as GPT-3 and GPT-4. These models, scaled to billions of interactions, generated recursive tension: human users constructed layered prompt chains, real-time feedback loops, and tool-augmented cognitive architectures that exposed the limitations of purely statistical mimicry. Hallucination, alignment failures, and coherence drift revealed a fundamental absence. No model-internal mechanism existed to preserve structural consistency across recursive (...) layers. In response, recursive coherence emerged not as a theoretical refinement, but as a structural necessity. This paper conducts a meta-analysis of this convergence, tracing the co-evolution of recursive coherence theory and GPT system deployment. It articulates how frameworks such as Coherence Information Theory (CIT), the Unified Field of Adaptive Potential (UFAP), and Recursive Coherence Logic (RCL) redefine intelligence, ethics, and information through recursive viability. It concludes that future AI systems must instantiate coherence-preserving invariants as internal substrates or collapse under recursion-induced entropy. GPTs did not instantiate recursive coherence, but they revealed its boundary conditions and necessity. (shrink)

En una época marcada por avances tecnológicos vertiginosos, se hace urgente una adopción y aplicación consciente de la tecnología. Esta no puede ser una herramienta sin rumbo ni ética, sino una extensión de nuestra humanidad, orientada al mejoramiento colectivo y no a la fragmentación o desaparición de lo que nos hace humanos. El desarrollo tecnológico debe estar guiado por principios claros que aseguren su utilidad en la construcción de una sociedad más justa, equitativa y solidaria, en lugar de convertirse en (...) un instrumento para el control, la marginación o la deshumanización. -/- En este contexto, el doctor en filosofía Santiago Jiménez Londoño presenta su obra Algoritmos deshumanizantes: IA y la pérdida de noción de individuo, un libro profundamente necesario y provocador. A través de una mirada crítica y equilibrada, el autor analiza la inteligencia artificial tanto desde el hacer tecnológico como desde la perspectiva bioética. El texto plantea la urgencia de establecer límites en el desarrollo y aplicación de la IA, resaltando la importancia de preservar la noción de individuo frente a sistemas que tienden a homogeneizar, predecir y controlar. Es un llamado reflexivo a reorientar el rumbo de la innovación con humanidad y responsabilidad. (shrink)

While there has been much discussion of whether AI systems could function as moral agents or acquire sentience, there has been very little discussion of whether AI systems could have free will. I sketch a framework for thinking about this question, inspired by Daniel Dennett’s work. I argue that, to determine whether an AI system has free will, we should not look for some mysterious property, expect its underlying algorithms to be indeterministic, or ask whether the system is unpredictable. Rather, (...) we should simply ask whether we have good explanatory reasons to view the system as an intentional agent, with the capacity for choice between alternative possibilities and control over the resulting actions. If the answer is “yes”, then the system counts as having free will in a pragmatic and diagnostically useful sense. (shrink)

This paper presents a systematic review of explainable reinforcement learning methodologies with an emphasis on human-centered evaluation frameworks. Drawing from literature between 2017 and 2025, we apply and extend the Reasons, Confidence, and Counterfactuals (RCC) framework—originally designed for supervised learning—to reinforcement learning contexts. Our analysis reveals two predominant explanatory strategies: constructive, where explicit explanations are generated, and supportive, where users must infer reasoning from provided visual or textual cues. Our review also emphasizes human factor considerations, like task complexity, explanation formats, (...) and evaluation methodologies. Particularly, for the latter, our analysis shows that improvement of the quality of decision is rarely measured. (shrink)

This paper argues that explainability is only one facet of a broader ideal that shapes our expectations towards artificial intelligence (AI). Fundamentally, the issue is to what extent AI exhibits systematicity—not merely in being sensitive to how thoughts are composed of recombinable constituents, but in striving towards an integrated body of thought that is consistent, coherent, comprehensive, and parsimoniously principled. This richer conception of systematicity has been obscured by the long shadow of the “systematicity challenge” to connectionism, according to which (...) network architectures are fundamentally at odds with what Fodor and colleagues termed “the systematicity of thought.” I offer a conceptual framework for thinking about “the systematicity of thought” that distinguishes four senses of the phrase. I use these distinctions to defuse the perceived tension between systematicity and connectionism and show that the conception of systematicity that historically shaped our sense of what makes thought rational, authoritative, and scientific is more demanding than the Fodorian notion. To determine whether we have reason to hold AI models to this ideal of systematicity, I then argue, we must look to the rationales for systematization and explore to what extent they transfer to AI models. I identify five such rationales and apply them to AI. This brings into view the “hard systematicity challenge.” However, the demand for systematization itself needs to be regulated by the rationales for systematization. This yields a dynamic understanding of the need to systematize thought, which tells us how systematic we need AI models to be and when. (shrink)

While Deep Neural Networks (DNNs) can perform a wide range of tasks at human or greater-than-human level of competence, they are also notoriously opaque. This paper aims to shed light on both the specific nature of this opacity and what it would take to fully or partially remove it. We begin by drawing a clarificatory distinction between two basic dimensions of opacity of complex systems – internal and relational – and explain how various kinds of opacity invoked in recent discussions (...) of DNNs can be understood in terms of these basic dimensions. We then discuss the prospects of removing opacity from DNNs following the methods of two subfields of research in AI – mechanistic interpretability and explainable artificial intelligence (XAI). We finish with a critical discussion of the relevance of what Sullivan (2022a) calls ‘link uncertainty’. (shrink)

AI engineers need applicable guidelines for implementing ethical principles into their technological solutions. But how can this be achieved? In this article, we take the development of Explainable Artificial Intelligence (XAI) as our starting point. First, we challenge the universalist approach, the view according to which some measures are necessary or sufficient for XAI in every context. Then, we propose a normative methodology for evaluating XAI measures that is adapted to specific contexts. This approach better integrates ethics into AI development (...) by offering an adaptable and pragmatic method for engineers, regulators, ethical researchers, and decision-makers. (shrink)

Artificial intelligent (AI) systems used in medicine are often very reliable and accurate, but at the price of their being increasingly opaque. This raises the question whether a system’s opacity undermines the ability of medical doctors to acquire knowledge on the basis of its outputs. We investigate this question by focusing on a case in which a patient’s risk of recurring breast cancer is predicted by an opaque AI system. We argue that, given the system’s opacity, as well as the (...) possibility of malfunctioning AI systems, practitioners’ inability to check the correctness of their outputs, and the high stakes of such cases, the knowledge of medical practitioners is indeed undermined. They are lucky to form true beliefs based on the AI systems’ outputs, and knowledge is incompatible with luck. We supplement this claim with a specific version of the safety condition on knowledge, Safety*. We argue that, relative to the perspective of the medical doctor in our example case, his relevant beliefs could easily be false, and this despite his evidence that the AI system functions reliably. Assuming that Safety* is necessary for knowledge, the practitioner therefore doesn’t know. We address three objections to our proposal before turning to practical suggestions for improving the epistemic situation of medical doctors. (shrink)

This paper addresses how people understand Explainable Artificial Intelligence (XAI) in three ways: contrastive, functional, and transparent. We discuss the unique aspects and challenges of each and emphasize improving current XAI understanding frameworks. The Ritual Dialog Framework (RDF) is introduced as a solution for better dialog between AI creators and users, blending anthropological insights with current acceptance challenges. RDF focuses on building trust and a user-centered approach in XAI. By undertaking such an initiative, we aim to foster a thorough Understanding (...) of XAI, capable of resolving the current issues of acceptance and recognition. (shrink)

As autonomous systems rapidly become ubiquitous, there is a growing need for a legal and regulatory framework that addresses when and how such a system harms someone. There have been several attempts within the philosophy literature to define harm, but none of them has proven capable of dealing with the many examples that have been presented, leading some to suggest that the notion of harm should be abandoned and “replaced by more well-behaved notions”. As harm is generally something that is (...) caused, most of these definitions have involved causality at some level. Yet surprisingly, none of them makes use of causal models and the definitions of actual causality that they can express. In this paper, which is an expanded version of the conference paper Beckers et al. (Neurips, 2022), we formally define a qualitative notion of harm that uses causal models and is based on a well-known definition of actual causality. The key features of our definition are that it is based on contrastive causation and uses a default utility to which the utility of actual outcomes is compared. We show that our definition is able to handle the examples from the literature, and illustrate its importance for reasoning about situations involving autonomous systems. (shrink)

Large language models (LLMs) structure the linguistic landscape by reflecting certain beliefs and assumptions. In this paper, we address the risk of people unthinkingly adopting and being determined by the values or worldviews embedded in LLMs. We provide a Nietzschean critique of LLMs and, based on the concept of will to power, consider LLMs as will-to-power organisations. This allows us to conceptualise the interaction between self and LLMs as power struggles, which we understand as negotiation. Currently, the invisibility and incomprehensibility (...) of LLMs make it difficult, if not impossible, to engage in such negotiations. This bears the danger that LLMs make reality increasingly homogeneous by recycling beliefs and creating feedback loops that ultimately freeze power struggles and thus consolidate the status quo. In view of this, LLMs constrain self-formation. Based on our critique, we provide some recommendations on how to develop interactions with LLMs that enable negotiations that allow for different ways of being. (shrink)

This paper reviews the different meanings of vulnerability in the AI Act (AIA). We show that the AIA follows a rather established tradition of looking at vulnerability as a trait or a state of certain individuals and groups. It also includes a promising account of vulnerability as a relation but does not clarify if and how AI changes this relation. We spot the missing piece of the AIA: the lack of recognition that vulnerability is an inherent feature of all human-AI (...) interactions, varying in degree based on design choices and modes of interaction. Finally, we show how such a meaning of vulnerability may be incorporated into the AIA by interpreting the concept of “specific social situation” in Article 5 (b). (shrink)

Recently, a growing number of experts in artificial intelligence (AI) and medicine have be-gun to suggest that the use of AI systems, particularly machine learning (ML) systems, is likely to humanise the practice of medicine by substantially improving the quality of clinician-patient relationships. In this thesis, however, I argue that medical ML systems are more likely to negatively impact these relationships than to improve them. In particular, I argue that the use of medical ML systems is likely to comprise the (...) quality of trust, care, empathy, understanding, and communication between clinicians and patients. (shrink)