Democratizing value alignment: from authoritarian to democratic AI ethics

Linus Ta-Lun Huang; Gleb Papyshev; James K. Wong · 2024 · AI and Ethics 5:11-18 interlocutor medium priority coded

Main argument

Thesis: current alignment (RLHF, constitutional AI) is 'authoritarian' - power-asymmetric, opaque, prioritizing developer values - and should be replaced by Dynamic Value Alignment (DVA): moral reasoning modeled as parallel constraint satisfaction, implemented by judgment-aggregation over a JURY of moral modules, each operationalizing a distinct normative source (an ethical theory like Kantian ethics, a value dimension like fairness, social-norm rules of thumb, or context-specific best practice e.g. medical), each scoring candidate responses, with USER-CONTROLLED weights aggregating the scores; high-scoring options are displayed WITH explanations of which principles support/undermine them, so the user makes the final informed choice. Argument type: conceptual proposal + five-step implementation sketch. Explicit benefits claimed: context-sensitivity emerges naturally; extreme single-value maximization is avoided; transparency of value influence; and satisfying the 'knowledge conditions of responsibility' - users can be responsible for decisions because they know which values shaped the options. Distinct normative modules prevent 'unintended interference' between value sources.

Why it matters here

The most concrete architectural proposal in the democratizing literature: modular 'moral jury' systems where distinct normative-principle modules score responses and users control the weights. Matters for the dissertation as (a) an implemented-style vision of PLURALIST alignment (modules = plural theories, aggregation = weighing) and (b) an explicit invocation of the knowledge condition of responsibility as a design goal - a rare responsibility-aware alignment proposal.

Reading notes

Close read of abstract, sec 4 (8pp; HKUST team). Cited by Steingrüber & Baum and Zhi-Xuan as the 'democratizing' pole. 'Authoritarian AI ethics' as their name for RLHF/CAI's power asymmetry. Parallel constraint satisfaction (connectionist moral cognition) as the theoretical base.

Huang, L. T.-L., Papyshev, G., & Wong, J. K. (2024). Democratizing value alignment: from authoritarian to democratic AI ethics. AI and Ethics, 5, 11-18. https://doi.org/10.1007/s43681-024-00624-1

Close reading — 5 coded units

#1 · pp. 11 · claim

“Existing approaches, such as reinforcement learning with human feedback and constitutional AI, however, exhibit power asymmetries and lack transparency. These 'authoritarian' approaches fail to adequately accommodate a broad array of human opinions, raising concerns about whose values are being prioritized.”

RL-DEV

#2 · pp. 15 · definition

“this approach will treat an AI system as a jury comprising many moral modules, each responsible for evaluating response options based on evidence and its own system of normative principles. [...] Different moral modules will be given different weights, which should reflect the user's own value priorities. The weighted numerical scores from different modules are then aggregated to form a collective evaluation.”

NF-PLURAL-OTHER

#3 · pp. 15 · definition

“The first step involves embedding a wide range of diverse values into individual modules respectively. This can include operationalizing principles of an ethical theory (e.g., Kantian ethics), an individual dimension of a value theory (such as fairness from moral foundation theory), or rule of thumbs for social norms. [...] The additional benefits [...] are to represent the diverse sources of values, to prevent unintended interference between them, and to enable flexible combination of moral modules in moral reasoning.”

NF-ROSS-PF

#4 · pp. 16 · argument

“we can avoid a response that maximizes a particular value to the extreme, which typically leads to undesirable behaviors. Most importantly, this algorithm will allow for flexibility in assigning weights to different modules, adapting to diverse user value priorities and contexts.”

VC-INTRA-VALUE

#5 · pp. 16 · argument

“This interface will display the high-scoring options to the users, along with explanations regarding some of the most relevant moral principles supporting or undermining them [...] Importantly, this process contributes to the necessary 'knowledge conditions' of responsibility for their decisions.”

RL-EPISTEMIC

Synthesis-matrix row

complicates T3-PROCEDURALISM-INCOMPLETE

democratic architecture that relocates weighing to users

supports T4-ROSSIAN-DEMAND

moral-jury weighted aggregation = engineering pluralism

complicates T6-RESPONSIBILITY-UNALLOCATED

knowledge-condition design - responsibility-aware but launderable

Memos (2)

theoretical · unit #2

The DVA moral-jury architecture (units 2-3) is the closest thing in the library to an ENGINEERING blueprint for Rossian pluralism: distinct normative modules (prima facie perspectives), context-sensitive weighted aggregation (the all-things-considered judgment), anti-extremization as a design property (unit 4 - answering Lundgren's local-optimum loops exactly as the dissertation's memo predicted pluralism would). Two dissertation-relevant critiques to develop: (a) user-controlled weights RELOCATE rather than solve the normative problem - the user becomes the weigher, inheriting the intra-value fragmentation problem (FISCHLI memo: meta-autonomy presupposes the user can do the weighing the theory couldn't); (b) numerical score aggregation re-imports the cardinal-comparability assumption Lloyd showed to be undefined across theories - a convergence-first architecture (flag verdict agreement before aggregating scores) would be more defensible. Both critiques are constructive extensions, publishable as engagement.

thesis-link · unit #5

Unit 5 is the only place in the coded set where an alignment DESIGN explicitly targets a responsibility condition: explanations exist so users satisfy the knowledge condition and can bear responsibility for choices. This inverts KAESTNER's scapegoat analysis constructively (there: opacity defeats responsibility; here: explanation infrastructure manufactures it) - but it also creates a new consent-laundering risk the authors don't see: if the interface's explanations are unfaithful (MILLIERE's thought-injection; the wise-judge problem), the 'knowledge condition' is merely simulated, and responsibility is transferred to users on false pretenses. The dissertation's faithfulness-testing methodology is the missing audit for responsibility-aware designs like DVA. Direct Work-chapter relevance: an AI Interviewer built on DVA would make the HIRING MANAGER 'responsible' via explanations - genuine or laundered?