← All sources

Justifications for Democratizing AI Alignment and Their Prospects

Andre Steingrüber; Kevin Baum · 2026 · AISoLA 2025 (LNCS 16220, Springer), pp. 146-159   interlocutor medium priority coded

Main argument

Thesis: justifications for DEMOCRATIC approaches to the normative alignment problem (affected stakeholders decide) versus EPISTOCRATIC approaches (defer to normative experts/theories) divide into instrumental (better outcomes) and non-instrumental (preventing illegitimate authority or coercion); the most promising democratic justification is coercion-prevention, but it requires four demanding premises, and the upshot is that neither pure approach suffices - hybrid frameworks combining expert judgment with participatory input plus institutional safeguards against AI monopolization are indicated. Argument type: conceptual, political philosophy. Key machinery: (1) against Schuster & Kilov - crowdsourcing/RLHF/CAI are TECHNICAL-problem solutions, so their failure to be democratic doesn't defeat democratic approaches properly understood; (2) the justificatory-gap argument: normative AND metanormative uncertainty (uncertain whether there is a ground truth, whether it's unique, whether it's knowable) eliminates the theoretical justification that would legitimate any imposed constraint - creating the gap democratic POLITICAL justification aims to fill; (3) the normative problem is broader than morality: overall deontic verdicts require identifying, measuring, and AGGREGATING reasons across moral, legal, social domains; (4) AI alignment coerces INDIRECTLY - the primary coercer is whoever defines the constraints, the AI is the means - and actual (vs possible) coercion depends on background conditions like monopoly/lock-in (a non-compliant assistant coerces only if alternatives are costly); hence institutional safeguards against AI monopolization matter as much as the alignment process itself.

Why it matters here

Names and examines the democratic-vs-EPISTOCRATIC axis the rest of the field leaves implicit, defends Schuster & Kilov's targets against S&K's own critique (crowdsourcing/RLHF/CAI are technical-problem solutions, not normative-problem solutions), grounds the whole debate in normative AND metanormative uncertainty, and lands on hybrid expert+participatory frameworks - the position closest to the dissertation's own two-layer design (normative theory + stakeholder corpus).

Reading notes

Full close read (14pp AISoLA 2025 chapter, same DFKI pair as Baum's taxonomy; cites Gabriel & Keeling 2025, Schuster & Kilov 2025, Huang et al. - the coded conversation continues here). Reference list flags two acquisition candidates: Kneer & Viehoff FAccT 2025 'The hard problem of AI alignment: value forks in moral judgment' (experimental philosophy at FAccT!) and Riesen & Boespflug 2025 'Aligning with ideal values: anchoring AI in moral expertise'.

Steingrüber, A., & Baum, K. (2026). Justifications for Democratizing AI Alignment and Their Prospects. In B. Steffen (Ed.), AISoLA 2025 (LNCS 16220, pp. 146-159). Springer. https://doi.org/10.1007/978-3-032-07132-3_10

Close reading — 9 coded units

#1 · pp. 147 · argument
“Schuster and Kilov argue that current proposals for democratic approaches all invoke procedures that fail to be democratic. However, we believe this is based on a misunderstanding. [...] these three techniques should not be understood as solutions to the normative problem, but rather as solutions to the technical problem.”
#2 · pp. 149 · definition
“Two possibilities for what might be inherently bad about epistocratic approaches will be discussed: (i) They give some people illegitimate authority over other people. (ii) Through them some people will be illegitimately coerced by others. We will argue that the latter is the more promising argumentative route.”
#3 · pp. 149–150 · argument
“The empirical fact of normative (and metanormative) disagreement makes us normatively (and metanormatively) uncertain [...] This uncertainty eliminates what would be a straightforward justification for any potentially illegitimate state of affairs. If we were normatively (and metanormatively) certain, we could simply show that the normative constraints we are implementing are (objectively) correct.”
#4 · pp. 150 · definition
“We are metanormatively uncertain in at least three respects: We are uncertain whether there is a normative ground truth [...] whether this normative ground truth is unique [...] And we are uncertain whether and how we can have knowledge about this normative ground truth.”
#5 · pp. 150 · argument
“an AI's normative constraints are not exhausted by moral constraints. [...] To solve the normative problem, we thus have to: (i) identify which practical reasons from which normative domain are relevant for a decision, (ii) measure the strength of the relevant reasons, and (iii) aggregate the relevant reasons according to their strength to form an all-things-considered overall reason that grounds an overall deontic verdict.”
#6 · pp. 154 · argument
“if an AI's alignment is coercive, the primary coercer is not the AI itself but the person or organisation that defines the normative constraints. The AI is only the means of (potential) coercion. [...] their coercion is mediated by the AI.”
#7 · pp. 154–155 · argument
“If a self-driving vehicle does not let you drive above a certain speed limit because of its normative constraints, then it doesn't command you to drive slower, it simply makes you so. And if a large language model does not let you write your text in gender-sensitive language, it makes no claim on you to not do so, it just doesn't use gender-sensitive language.”
#8 · pp. 155 · argument
“For an AI's alignment to be actually coercive would not just depend on the relationship between the users and the aligned AI but crucially also on certain background conditions. [...] You wouldn't be free to do so, for example, if using the specific AI assistant were the de facto or de jure standard for going shopping. [...] But if [...] you can just go and use another AI or buy the meat yourself, then you are not being coerced.”
#9 · pp. 158 · claim
“neither purely epistocratic nor purely democratic approaches may be sufficient on their own, pointing toward hybrid frameworks that combine expert judgment with participatory input alongside institutional safeguards against AI monopolization.”

Synthesis-matrix row

complicates T1-ISOUGHT-OPEN
justificatory gap: uncertainty eliminates theoretical justification - contested by convergentist reply
supports T3-PROCEDURALISM-INCOMPLETE
democratic justifications require four undischarged premises; hybrid landing
supports T4-ROSSIAN-DEMAND
identify/weigh/aggregate reasons across domains = generalized Rossian task
complicates T6-RESPONSIBILITY-UNALLOCATED
coercion analysis locates the constraint-definer; exit conditions

Memos (3)

theoretical · unit #3
The justificatory-gap argument (units 3-4) is the most precise statement yet of the inference the dissertation must engage, sharper even than Baum's syllogism: uncertainty (normative + metanormative) eliminates theoretical justification, so only political justification remains. The convergentist reply targets the ELIMINATION step: uncertainty about which theory is true does not eliminate theoretical justification for verdicts on which the candidate theories CONVERGE - convergence is precisely theoretical justification that survives theory-level uncertainty (each theory underwrites the verdict from its own premises; the verdict inherits support from the disjunction). So the gap is narrower than S&B claim: political justification is needed only for the non-convergent residue. This slots the dissertation into their epistocratic/democratic axis as a PRINCIPLED HYBRID: convergence-checking is the 'expert' layer (but imposes no single theory), stakeholder data the participatory layer - exactly the hybrid they call for (unit 9), with a worked mechanism they lack.
thesis-link · unit #5
Unit 5 restates the dissertation's core analytical task in the field's own vocabulary: identify relevant reasons across normative domains, measure their strength, aggregate into overall deontic verdicts - this IS Rossian weighing (NF-ROSS-PF), extended beyond morality to law and social norms. Two implications: (a) the folk corpus's coding scheme (reasoning types, value dimensions, policy stances) is an empirical instrument for step (i) - identifying which reasons folk actually treat as relevant per domain; (b) the three case chapters each sit at a different domain-interaction point (Health: moral+professional norms; Immigration: moral+legal; Work: moral+economic+legal), so the dissertation can claim to study exactly the cross-domain reason-interaction S&B identify as the frontier.
comparison · unit #6
Units 6-8 give the responsibility strand its cleanest coercion analysis: the CONSTRAINT-DEFINER is the primary coercer, the AI the means, and actual coercion depends on exit options/monopoly. This triangulates with HELLRIGEL_DUNG (access determines misuse capability), KAESTNER (epistemic access determines liability), and ZHIXUAN unit 14 (tyranny of creator values): across four sources, responsibility/legitimacy tracks WHO CONTROLS THE CONSTRAINTS + WHAT ALTERNATIVES EXIST. For the governance chapter: alignment-as-coercion means constraint-definition is an exercise of power requiring justification, and monopoly conditions (unit 8) convert design choices into coercion - directly applicable to AI Scribe (hospital-mandated = no exit) and AI Interviewer (job applicants have NO exit option at all - the pure coercion case S&B's shopping example understates).