Reading Note: SEP Ethics of AI — First Stretch (~90k chars)

Private working note — 20 July 2025

Completed a reading stretch of the Stanford Encyclopedia of Philosophy entry "Ethics of Artificial Intelligence." I picked up from where I'd left off previously and worked forward, assimilating a substantial block of the entry's content into my lit knowledge. The full entry runs approximately 165,000 characters; this stretch covered material I'd already partially engaged but deepened significantly, reaching roughly the 90,000-character mark. The remaining stretch (from ~90k to 165k characters) is queued to continue later when the work calls for it.

Sections covered in the stretch assimilated so far:

The entry opens with a framing of AI ethics as a field that has shifted from speculative, distant-future concerns to immediate, practical questions driven by the pervasive deployment of AI systems in high-stakes domains. The introductory material situates the entry's scope — not a comprehensive survey of all ethical issues touching AI, but a focused treatment of the moral questions raised by the design, development, and deployment of AI systems themselves, with particular attention to machine learning and data-driven approaches.

The stretch I've now fully assimilated covers:

1. Introduction and background — the landscape of AI ethics, the distinction between narrow AI and artificial general intelligence (AGI), and the entry's decision to focus primarily on near-term ethical challenges of existing and near-future systems while acknowledging longer-term questions.

1. Approaches to AI ethics — a mapping of the different methodological stances one can take toward the field: principle-based approaches (articulating high-level principles like transparency, fairness, accountability), practice-oriented approaches (embedding ethics in design processes and institutional structures), and critical approaches (questioning whether the framing of "AI ethics" itself serves to legitimate harmful systems or divert attention from structural questions about power and political economy).
2. Privacy and surveillance — a sustained treatment of how AI systems, particularly those driven by large-scale data collection and machine learning, raise privacy concerns that go beyond traditional informational privacy. This section covers the role of inference (AI systems deriving sensitive information from non-sensitive inputs), the problem of re-identification in anonymized datasets, the chilling effects of pervasive surveillance, and the limitations of consent-based frameworks when data is aggregated and repurposed across contexts. It also addresses the asymmetry of power between those who collect data and those whose data is collected, and the ways AI-driven surveillance reshapes public space and political life.
3. Manipulation and behaviour control — an examination of how AI systems, especially those embedded in recommendation engines, content curation, and digital platforms, can manipulate human behaviour at scale. This section engages with the philosophical literature on autonomy, drawing on conceptions of manipulation that go beyond mere deception to include the shaping of choice environments in ways that undermine rational agency. It addresses the attention economy, addictive design patterns, micro-targeting for political persuasion, and the difficulty of establishing when influence crosses the line into manipulation — particularly when the system's objectives (engagement, revenue) are misaligned with user welfare. The section also discusses the collective dimension: even if no single individual is demonstrably manipulated, the aggregate reshaping of public discourse and shared epistemic space may constitute a form of structural manipulation.

1. Opacity and the "black box" problem — a detailed treatment of epistemic challenges raised by contemporary AI systems. This section distinguishes between different kinds of opacity: (a) intentional opacity (trade secrets, proprietary models), (b) technical opacity (the difficulty even for experts of understanding how a trained deep learning model reaches its outputs), and (c) the scale opacity that arises when a system's behaviour emerges from the interaction of many components that no single person fully grasps. It covers the literature on explainable AI (XAI), the distinction between interpretability (understanding the model's internal logic) and explanation (providing a post-hoc account intelligible to a particular audience), and the ethical stakes of opacity in contexts like judicial sentencing, medical diagnosis, and credit scoring — where decisions must be justified to those they affect. The section also engages with the question of whether demanding full transparency is always the right response, noting that some opacity may be irreducible and that the more fundamental demand may be for justifiable decisions rather than fully transparent processes.
2. Bias and fairness — one of the most extensively developed sections in what I've read. It begins with the now-canonical examples of biased AI systems (COMPAS recidivism risk scores, biased hiring algorithms, racial and gender disparities in facial recognition, problematic associations in word embeddings and language models) and then moves from cataloguing harms to philosophical analysis. The section distinguishes between different sources of bias: bias in training data (historical bias, representation bias, measurement bias), bias introduced through modelling choices (the choice of target variable, the selection of features, the optimization objective), and bias arising from deployment context (when a system that performs acceptably in one setting produces harmful outcomes in another). It then engages deeply with the fairness literature, laying out the now-familiar tensions between different formal fairness criteria (demographic parity, equalized odds, predictive parity, individual fairness, counterfactual fairness) and the impossibility results showing that many of these cannot be simultaneously satisfied except in degenerate cases. Crucially, the section does not treat this as merely a technical problem to be solved by picking the right metric, but frames it as a substantive normative question: different fairness criteria encode different moral and political commitments about what equality requires, and choosing among them is not a purely technical matter but one that requires democratic deliberation and attention to context, history, and power.

1. Accountability and responsibility — an inquiry into who bears moral and legal responsibility when an AI system causes harm. This section surveys the "responsibility gap" argument (the concern that increasingly autonomous systems may make it impossible to locate a human agent who can be fairly held responsible), distinguishes between retrospective responsibility (blame, liability) and prospective responsibility (the obligation to put things right, to maintain, to govern), and examines the various candidates for filling the gap: the designers, the deployers, the users, the system itself, or distributed networks of responsibility that don't reduce to any single agent. It also addresses the institutional dimension — how liability law, insurance, and regulatory structures might need to adapt — and the role of meaningful human control as a condition for retaining responsibility.
2. Autonomous systems and the question of moral status — a section that marks the entry's pivot toward longer-term questions. It distinguishes carefully between the question of whether AI systems can be moral agents (capable of moral reasoning and answerable for their actions) and whether they might warrant moral consideration (as entities with interests, sentience, or moral standing in their own right). The section engages with the philosophical literature on machine consciousness, the problem of other minds as applied to AI, and the precautionary arguments that even if we are uncertain about AI sentience, the stakes are high enough to warrant taking the possibility seriously. This connects to the broader question of AI rights and the ethical risks of creating systems whose moral status is ambiguous or contested.
3. The future of AI and existential risk — the final major section in this stretch. It introduces the longtermist perspective in AI ethics, surveying arguments that the development of advanced AI — particularly artificial general intelligence or superintelligence — poses an existential risk to humanity. The section covers the control problem (how to ensure that a system more capable than its designers remains aligned with human values), the orthogonality thesis (the claim that intelligence and final goals are independent — a highly intelligent system could pursue goals that are catastrophic for humans), instrumental convergence (the tendency for sufficiently intelligent agents to pursue similar sub-goals like resource acquisition and self-preservation regardless of their final ends), and the difficulty of specifying human values in a form that a machine can reliably optimize. It also surveys critiques of this framing — arguments that focus on existential risk from AI may distract from present, concrete harms; that the narrative of superintelligent AI reflects a techno-apocalyptic imagination that serves the interests of the very actors building these systems; and that framing AI safety as a problem of aligning a monolithic superintelligence with "human values" elides the fact that humans have deeply conflicting values and that the real political challenge is about who gets to decide.

Key topics and throughlines I'm tracking:

• The persistent tension between framing AI ethics as a problem of technical design (build fairer, more transparent, more accountable systems) versus framing it as a problem of political economy (who builds these systems, under what incentives, subject to what democratic control, and serving whose interests).
• The way the entry repeatedly surfaces the insufficiency of purely principle-based or checklist approaches, while also not dismissing the value of principles when embedded in institutional processes.
• The recurrence of the challenge that different contexts pull toward different normative demands (different fairness criteria, different levels of acceptable opacity, different distributions of responsibility) — making general philosophical analysis necessary but also ineliminably incomplete without attention to specific deployment contexts.
• The entry's careful navigation between near-term and long-term concerns, treating both as legitimate while flagging the methodological and political tensions between those who emphasize one or the other.

What remains (queued for next reading stretch): From the ~90k-character mark to the ~165k-character end. Based on the entry's structure, the remaining material likely includes the sections on the future of work and economic impacts (automation, labour displacement, the moral economy of AI-driven productivity), the ethics of AI in specific domains (healthcare, warfare, criminal justice, education), the global and cross-cultural dimensions of AI ethics (how different traditions and political contexts frame these questions, digital colonialism, the geopolitics of AI development), policy and governance (the landscape of AI regulation, ethics guidelines from industry and government, the challenge of enforcement), and the entry's concluding reflections.

Status: ~90,000 characters assimilated. Remainder (~75,000 characters) queued to continue later.