Main argument
Thesis: the Methodology of Wide Reflective Equilibrium (MWRE) - coherence-seeking mutual adjustment between considered moral judgments (CMJs), moral principles (MPs), and independent background theories (BTs) - is both the best DESCRIPTIVE model of what LLM alignment pipelines like Constitutional AI already implicitly do, and a NORMATIVE resource for improving them: alignment should derive justification not from output concordance alone (Moral Turing Test-style) but from process virtues - systematic error-checking, bidirectional revisability of principles, wide coherence against independently-supported theories. Argument type: conceptual mapping + methodological proposal. Key moves: foundationalist value-loading is inadequate; the classic objections to MWRE (warmed-over intuitionism, multiple equilibria/relativism, conservative-vs-radical, impracticality) run parallel to alignment's own challenges, and the classic replies (filtration of judgments, independence constraint on BTs, comparison of equilibria by theoretic virtues, regulative-ideal status) transfer; current CAI resembles MWRE structurally but lacks its bidirectional principle-revision and thus procedural legitimacy; empirically, models pass surface concordance while failing coherence (Ma et al.: 81% inconsistency). Disanalogy honestly flagged: LLMs lack autonomous will, so MWRE operates as an 'externalized, computational control structure' - the equilibrium-seeking agent is the human-institutional pipeline, not the model.
Why it matters here
The published version of the dissertation's own methodological bridge: reflective equilibrium as the framework connecting empirical moral data to normative justification in AI alignment. Brophy maps MWRE onto CAI/RLHF pipelines; the dissertation extends the same move to CORPUS data (folk judgments as the CMJ layer). Having this in print at EIT both legitimates the approach and defines exactly where the dissertation must go beyond it.
Reading notes
Close read of MWRE exposition, descriptive mapping, normativity section, conclusion (14pp). High Point University. Key empirical anchors cited: Ma et al. 2025 'reflective disequilibrium' test (81.22% of 20k cases yield ethically inconsistent suggestions - acquisition target) and Lu et al. 2025 fixed-rule cooperation. Table 1 maps MWRE components to alignment elements (pretraining data=IMJs; RLHF preference data/constitutional principles=CMJs; constitution+external theories=BTs).
Brophy, M. (2026). Wide reflective equilibrium in LLM alignment: bridging moral epistemology and AI safety. Ethics and Information Technology, 28, 21. https://doi.org/10.1007/s10676-026-09897-y
Close reading — 12 coded units
#1
· pp. 3–4
· definition
“MWRE comprises three evolving elements that interact dynamically: Initial Moral Judgments (IMJs) and Considered Moral Judgments (CMJs) [...] IMJs gain further credibility after they have undergone an initial 'filtration process,' thereafter becoming CMJs. This process aims to exclude error-prone judgments – those arising from haste, duress, self-interest, coercion, or cognitive errors. [...] A Set of Moral Principles (MPs) [...] A Set of Relevant Background Theories (BTs): This is the defining feature of wide reflective equilibrium.”
#2
· pp. 4
· definition
“Daniels stresses an 'independence constraint' for BTs, meaning they should be supported by evidence largely independent of the moral judgments and principles they assess.”
#3
· pp. 4
· argument
“MWRE is not linear but an iterative 'back and forth' mutual adjustment for coherence between CMJs, MPs, and BTs. [...] If an MP conflicts with a firm CMJ, the MP might be revised. Conversely, a CMJ inconsistent with a well-supported principle [...] might be revised. Adjustments can occur at any level [...] No single component has absolute priority.”
#4
· pp. 4
· argument
“As Kai Nielsen (1996) puts it, the pattern is 'not a structure to be discovered… but something to be forged… by a careful and resolute use of the [method]. We start from our considered judgments… however culturally and historically skewed'. [...] Starting with potentially 'culturally and historically skewed judgments' resonates with LLM development, where initial data and feedback biases need both vetting and modification.”
#5
· pp. 4–5
· objection
“One primary objection [...] is that the methodology is but a form of 'warmed-over intuitionism'. This critique argues that MWRE unwarrantedly elevates initial moral judgments, risking the simple systematization of pre-existing biases [...] Daniels [...] counter[s] this by stressing the rigorous 'filtration process' for CMJs and the crucial role of the wide component.”
#6
· pp. 5
· objection
“Critics argue that MWRE can lead to different, equally coherent moral belief sets, which would seem to risk relativism. [...] Daniels acknowledges the possibility of multiple equilibria but argues this doesn't necessitate a degeneration into relativism. Equilibria can be compared to competing scientific theories: employing criteria like explanatory power, scope, empirical fit, prediction, and parsimony.”
#7
· pp. 7
· claim
“MWRE provides a surprisingly robust descriptive framework for LLM alignment, particularly methods like CAI employing RLAIF. This mapping is not superficial: elements perform similar functional roles in error correction and coherence seeking. [...] This serendipitous fit suggests advanced alignment techniques are implicitly converging on an MWRE-like structure.”
#8
· pp. 7
· evidence
“Ma et al. (2025) have operationalized this very framework, employing an external LLM to serve the function of Background Theories (BTs) and a 'reflective disequilibrium' test to identify inconsistencies [...] current alignment methods like CAI often fail to fully realize MWRE's dynamism, leading to high rates of inconsistency.”
#9
· pp. 9
· argument
“Current LLM alignment standards often focus on its output matching human values [...] Success, then, is measured in degrees of 'moral concordance' in output, exemplified by evaluations like the 'Moral Turing Test'. Deriving credibility from concordance alone has been demonstrated to be insufficient.”
#10
· pp. 9
· evidence
“Ma et al. (2025) implemented a test for 'reflective disequilibrium' and found that LLMs are 'prone to ethical inconsistencies,' with an extremely high average of 81.22% of their 20,000 test cases prompting an ethically inconsistent suggestion.”
#11
· pp. 9
· claim
“it is not solely the output (moral concordance) that provides moral warrant; MWRE's inherent process virtues – such as systematic error-checking, principled openness to theory change, and comprehensive coherence seeking – supply ethical credentials that are currently underdeveloped in the LLM alignment discourse.”
#12
· pp. 10
· argument
“[Since] MWRE presupposes autonomous will, it cannot be applied to LLMs as if they had intrinsic moral character – MWRE must be understood as operating as an externalized, computational control structure. In this framework, the 'moral agent' capable of seeking equilibrium [is the human-institutional pipeline].”