Themes
Each theme is a claim about the literature, with its dissertation link and evidential profile. Full grid on the Matrix page.
T1-ISOUGHT-OPEN — The field states the is/ought gap but never closes it
At least seven sources independently assert that descriptive data (preferences, folk judgments, legal records, benchmark concordance) cannot by itself ground normative alignment claims - and no source supplies the metaethical bridge; the gap is the field's most-repeated self-criticism and remains open.
supports 7 complicates 3 contradicts 2 silent 44
T2-PREFERENTISM-BROKEN — Preferentism has collapsed under convergent multi-front attack
Preference-based alignment fails as psychology (preferences are constructed from reasons and values), as decision theory (incommensurability, incompleteness, formal Goodhart results), as training practice (sycophancy is structurally entailed; RLHF alters surface tokens), and as social epistemology (endogeneity: systems shape the preferences that score them) - the critiques converge from independent disciplines.
supports 11 complicates 4 contradicts 0 silent 41
T3-PROCEDURALISM-INCOMPLETE — The procedural turn cannot discharge the normative question
The dominant successor to preferentism - fair-process selection of principles (Gabriel program, deliberative, bargaining, democratic variants) - fails at the encoding step (legitimacy does not survive principles-to-policy transformation), cannot scope its own stakeholders (all-affected gaps), and repeatedly re-imports an unaccounted normative corrective (expert deference, human-rights limits), leaving P4 of the proceduralist syllogism undischarged.
supports 7 complicates 7 contradicts 0 silent 42
T4-ROSSIAN-DEMAND — The literature demands Rossian weighing machinery without naming it
Across safety (normative-conflict jailbreaks), design (moral juries, meta-autonomy), review findings (dynamic multi-logic alignment), and formal work (multi-objective incompleteness), the operative need is a capacity to weigh plural defeasible considerations into contextual all-things-considered verdicts - Ross's machinery, explicitly identified only once (Millière) and supplied as a worked account by no one.
supports 10 complicates 4 contradicts 0 silent 42
T5-AGENCY-DENIED-EVALUABILITY-KEPT — Moral agency is denied while normative evaluability is retained, on assumption
Every tradition applied to current AI (Kantian, analytic-behavioral, continental, virtue-ethical, enactivist) denies moral agency while keeping systems normatively evaluable (constrained-not-accountable; media of moralisation; X-ethicality) - but the load-bearing empirical presupposition (no stable practical standpoint) is stipulated or footnoted, not argued; six empirical instability studies exist yet are never integrated into the philosophical position.
supports 14 complicates 1 contradicts 2 silent 39
T6-RESPONSIBILITY-UNALLOCATED — Responsibility is affirmed as distributed and allocated by no one
The alignment literature systematically brackets responsibility attribution (explicit deferrals in Gabriel-Keeling, Kästner, Lundgren); where responsibility is affirmed (Noller's shared responsibility, Dignum-adjacent gestures, IEAI), no allocation mechanism is given; the allocation resources exist scattered (difference-making causation, epistemic/access conditions, principal-agent structure, coercion analysis) but no source assembles them.
supports 8 complicates 9 contradicts 0 silent 39
T7-AGENTIC-BREAKS-FRAMES — Agentic AI breaks the dyadic pre-agentic frame of every major framework
The canonical frameworks presuppose a live principal anchoring alignment at decision time; multi-step delegated systems break this (context-shaping, preference-endogeneity, collective effects, enlarged attack surfaces, the agency-misuse tradeoff) - and the flagship treatments respond by deferral (collective effects, genAI, responsibility) or containment proposals that reintroduce maximal misusability.
supports 9 complicates 5 contradicts 0 silent 42
T8-NONWESTERN-CONCEDED — The non-Western gap is conceded by the field itself
Gabriel (geographic parochialism), Schuster-Kilov (Ubuntu/Confucian limitation), and McKinlay's SLR (Western-canon confession) explicitly concede that alignment theory speaks from and for a Western frame; the acknowledged gap is a standing invitation for African-philosophy contributions, with only scattered responses (Metz citations, Basotho governance chapter, Japanese relational autonomy) so far.
supports 3 complicates 4 contradicts 0 silent 49