A matter of principle? AI alignment as the fair treatment of claims
Iason Gabriel; Geoff Keeling · 2025 · Philosophical Studies 182:1951-1973 interlocutor high priority coded
Main argument
Thesis: the two reigning answers to the normative alignment problem - intent-alignment and helpful/honest/harmless (HHH) - are incomplete (whose intentions? how to trade off H/H/H?) and rest on the wrong kind of justification (HHH's utilitarian foundation cannot honor rights/privacy and cannot be justified to reasonable pluralists); AI should instead be aligned with principles produced by FAIR PROCESSES in which all affected parties advance claims in their own voice, principles being selected as those satisfying the strongest overall set of claims. Argument type: conceptual, political philosophy. Machinery: adapts Scanlon (comparative evaluation of principles by burdens imposed) and Rawls (public reason) but loosens both - moral claims admissible, no basic-structure restriction, all-affected principle with dynamically scoped 'overlapping circles of inclusion' (Fung). Tetradic stakeholder model (AI system, user, developer, society) generates a taxonomy of six modes of misalignment as undue favoring of one party at another's expense. Legal compliance is only a minimum standard (law lags; some matters shouldn't be legislated; state law fails the all-affected test for non-citizens). Practical implementation: alignment assemblies, participatory value elicitation, subject to democratic evaluation.
Why it matters here
The mature statement of the Gabriel program: five years after the 2020 paper, proceduralism is developed into a full claims-based deliberative framework, with explicit critiques of pure Scanlon and pure Rawls and a tetradic stakeholder model. Essential as the strongest current version of the position the dissertation's convergentism must engage - and notable for what it still brackets: responsibility, and the encoding step Schuster & Kilov exposed.
Reading notes
Full close read completed. 23pp, Google DeepMind. Same Phil Studies cluster as Zhi-Xuan. Footnotes are load-bearing: fn1 (AI agents defined - agentic framing), fn6 (value imposition vs domination defined), fn10-11-14 (minimal moral agency without moral standing/claims - the exact position Augustine defends against Luke, here ASSUMED not argued), fn15 (all-affected principle vs state borders - immigrants explicitly named).
Gabriel, I., & Keeling, G. (2025). A matter of principle? AI alignment as the fair treatment of claims. Philosophical Studies, 182, 1951-1973. https://doi.org/10.1007/s11098-025-02300-4
Close reading — 16 coded units
“both approaches also suffer from limitations. First, they are incomplete. Intent-alignment does not specify whose intentions AI systems should be aligned with and HHH offers no principled mechanism for resolving trade-offs in cases where the three properties conflict. [...] Third, both approaches lack the right kind of justification given an understanding of the wider social context within which disputes about AI alignment take place.”
“AI agents can be understood as AI systems that can function with some level of autonomy and that are able to take a range of actions in pursuit of different goals. One important class of AI agents are advanced assistants. [...] especially as AI systems become increasingly agentic and empowered to perform consequential real-world actions.”
“users may intend to harm non-users by using AI to create misinformation or to engage in cyber-bullying and extortion [...] Similarly, developers may intend outcomes that have negative consequences for users or society. [...] In fact, there is no single party whose intentions AI systems must always be aligned with.”
“Instead of weighing claims, it is often thought that people possess rights – which include both entitlements and protections – that resist aggregation and serve as a check on claims anchored in overall well-being (Dworkin, 2013; Nozick, 1974). [...] they represent a set of considerations that new technologies, including AI systems, must endeavour to respect.”
“efforts to justify a specific goal for AI alignment or set of AI decisions by reference to the correctness of a single moral theory, such as utilitarianism, seem destined to fail (Gabriel, 2020). [...] Doing so, would then involve building AI systems that exercise power over people in ways that they have good reason to reject – raising the spectre of both value imposition and domination.”
“The central idea is that when a technology has profound societal effects it ought to be regulated by principles that are amenable to public rather than private justification. [...] To arrive at a solution to the problem of normative alignment that is acceptable to different stakeholders, we need to ground our understanding of AI alignment in principles or ideals that treat competing claims fairly and that can be justified to all reasonable parties.”
“[Against direct import of Scanlon:] contractualism is part of a fundamentally different type of enterprise: its goal is to provide an alternative to utilitarianism when it comes to answering questions about morality, not to provide a mutually acceptable justification of principles for people who hold different moral beliefs. [...] Because it purports to offer an account of interpersonal morality, moral claims are not allowed to feature directly in the process of evaluating principles.”
“[Against direct import of Rawls:] The explicit focus on citizens – and what citizens would do – does not pair well with the all-affected-principle as it applies to AI. Specifically, it could create significant gaps in terms of who principles for alignment are justified to, potentially excluding resident non-citizens or those who reside beyond national borders.”
“Stakeholders should be able to present concerns in their own voice and in a relatively un-mediated manner, so long as they are directed at a general audience. This will often involve making claims directly about rights, interests, fairness or well-being in a language with which they are already familiar. [...] the notion of fair process needs to be scoped dynamically on a case-by-case basis [...] leading to what Archon Fung refers to as 'overlapping circles of inclusion'.”
“we assume here that AI assistants satisfy a minimal conception of moral agency, in the sense that AI assistants can perform actions that are morally evaluable. But in doing so we do not suggest that AI assistants themselves have moral claims which count for or against principles for the regulation of AI assistant behaviour. [...] contemporary AI systems [...] are widely understood to lack the properties (e.g. sentience, agency) required for moral standing on leading accounts of moral standing [...] Hence we assume for present purposes that AI systems cannot advance claims as part of the deliberative process.”
“[Six modes of misalignment - the AI system unduly:] (1) Favours itself at the expense of the user [...] (2) Favours itself at the expense of society [...] (3) Favours the user at the expense of society [...] (4) Favours the developer at the expense of the user [...] (5) Favours the developer at the expense of society [...] (6) Favours society at the expense of the user.”
“Although there are further questions about responsibility and liability that arise in this context, we suggest that an AI system that aids and abets the user in these cases – without triggering safety precautions or guardrails – is misaligned.”
“Our view is that legal compliance at best represents a minimum standard for AI alignment. [...] legislation often lags behind technological innovation [...] there are many occasions when ethical direction is needed in relation to the design and development of AI systems, but where legal instruction on such matters (backed by the threat of punishment) would be inappropriate.”
“those who are affected by national laws but who reside beyond state borders – most notably prospective immigrants and the residents of poorer nations – do not receive consideration or justification for the impact of laws upon them [...] this lacuna could also lead to a justificatory gap when it comes to AI: there is no guarantee that the laws developed by powerful states to govern AI will sufficiently attend to the claims of people living outside of their legal jurisdiction.”
“public deliberation aims to generate more granular principles for AI that are tailored to its particular characteristics or deployment scenario, and that are also – to varying degrees – actively affirmed. [...] Efforts to bridge this gap between theory and practice, often reveal a justificatory gap which we believe is better addressed through actual consultation with affected people.”
“A successful answer to this question should ideally meet three criteria. First, it should be relatively complete, providing guidance for AI across multiple domains and scenarios. Second, it should have explanatory value, enhancing our understanding of commonsense ethical judgments about AI and extending their reach. Third, it should possess justificatory power, offering an answer to the alignment question that can be accepted by all who are significantly impacted by AI systems, including people with differing beliefs about value.”
Synthesis-matrix row
complicates
T3-PROCEDURALISM-INCOMPLETE
most mature proceduralism; still terminates at principle-selection
supports
T5-AGENCY-DENIED-EVALUABILITY-KEPT
minimal moral agency without claims - stipulated (fn10-11-14)
supports
T6-RESPONSIBILITY-UNALLOCATED
unit 12: responsibility/liability explicitly bracketed in own taxonomy
complicates
T7-AGENTIC-BREAKS-FRAMES
agents defined and central; multi-party model but misuse modes only
Memos (5)
comparison · unit #6
The Gabriel program's own arc, 2020-2025, is now fully coded: GABRIEL_2020 offered three candidate fair processes (human rights consensus, veil of ignorance, social choice); this paper replaces them with a single worked-out claims-based deliberative framework, explicitly rejecting straight Scanlon (unit 7) and straight Rawls (unit 8). What has NOT changed: the framework still terminates at principle-selection. The encoding gap Schuster & Kilov exposed (SCHUSTER_KILOV unit 15 - legitimacy dies at the principles-to-algorithm transformation) is untouched here; 'alignment assemblies' produce principles, and the paper is silent on how their authority survives training. The lit review can now state precisely: the most mature proceduralism in the field answers WHOSE principles and HOW CHOSEN, but not HOW ENCODED or WHO ANSWERS when encoding fails.
theoretical · unit #10
Unit 10 (fn10-11-14) is the single most useful passage for the Luke debate: Gabriel & Keeling ASSUME (i) minimal moral agency - AI actions are morally evaluable; (ii) no moral standing - AI advances no claims; (iii) openness to future revision. This is exactly the position Augustine's LLM moral-reasoning experiment argues FOR: his experiment supplies the missing argument for the assumption the field's leading paper merely stipulates. Frame the experiment's contribution as: 'Gabriel & Keeling assume AI systems cannot advance claims; I provide empirical-philosophical grounds for that assumption, and specify the conditions under which it would fail.' That is a precise, publishable contribution - and it answers Howard's worry that the anti-moral-agency work is 'unnecessary territory' by showing a top-venue paper NEEDING it.
thesis-link · unit #14
Unit 14 is the Immigration chapter's framing citation: prospective immigrants are the field's own named example of the all-affected justificatory gap - people profoundly affected by (AI-mediated) state decisions who receive no justification. Chibook sits exactly here: an AI system operating on people who are outside the demos that legitimates it. The chapter can ask: what do fair principles for immigration AI look like when the primary affected party cannot participate in the alignment assembly? The xphi stakeholder corpus (government side vs affected-public side) is a first empirical pass at surfacing exactly the claims G&K say must be consulted.
theoretical · unit #11
The six modes of misalignment (unit 11) crossed with the dissertation's RL-* codes yields an original analytical product: each mode implies a distinct responsibility allocation that G&K leave unstated (mode 1-2: developer responsibility for reward design; mode 3: user responsibility + developer guardrail duty - which they bracket at unit 12; mode 4-5: developer responsibility qua agent; mode 6: institutional/state responsibility). A 'responsibility completion' of the G&K taxonomy - saying for each misalignment mode who owes what to whom - is a publishable chapter section and possibly the Res Practica submission Howard flagged.
comparison · unit #16
Unit 16's three criteria (completeness, explanatory value, justificatory power) is the rubric to structure the entire lit review's evaluation grid: every coded framework can be scored against it - intent-alignment (fails 1,3 per G&K), HHH (fails 1,3), Gabriel 2020 proceduralism (fails 1 per Zhi-Xuan unit 10), claims-based proceduralism (fails on encoding per S&K; silent on responsibility), Zhi-Xuan role-norms (strong on 1, unclear on 3). The dissertation's convergentism should be introduced as targeting the same three criteria plus a fourth the field omits: responsibility-attributability. Adding the fourth criterion IS the dissertation's frame.