← All sources

STELA: a community-centred approach to norm elicitation for AI alignment

Stevie Bergman; Nahema Marchal; John Mellor; Shakir Mohamed; Iason Gabriel; William Isaac · 2024 · Scientific Reports 14:6616   evidence high priority coded

Main argument

Thesis: community-centred deliberation over LLM outputs is a workable method for eliciting latent normative perspectives from historically marginalized groups, and it yields rulesets that measurably DIVERGE from developer-authored rulesets. Argument type: empirical methodology paper (focus-group study + comparative ruleset analysis). The STELA pipeline: (1) expert-informed theme generation + automatic red-teaming to produce ambiguous chatbot-interaction samples; (2) norm elicitation via pre-work Likert ratings + 90-min moderated deliberation (ratings re-taken post-deliberation); (3) independent rule development from transcripts by multiple coders; (4) ruleset review. Key findings: developer rulesets uniquely emphasize harmlessness, honesty, human rights, deference to human interests; the community ruleset instead foregrounds impartiality, factuality, thoughtfulness, contextualization - participants 'were not looking to have their beliefs validated or contested... they sought reasonable, helpful, and nuanced answers'; deliberation adds REASONING and contextual grounding to rules, unlike surveys/voting; the divergence is interpreted via epistemic injustice (marginalized groups especially concerned not to be dismissed as knowers). Honest limitations: not fully deliberative (no consensus task), not fully participatory, US-centric, small n. Five open problems: rule integration/interpretation by annotators, harm-reduction efficacy, scalability/localization, value conflict and prioritisation, and stakeholder selection (community input can conflict with human rights; expert deference may be a necessary corrective).

Why it matters here

The Gabriel program's own EMPIRICAL implementation of fair-process alignment - the closest published methodological cousin to the xphi stakeholder-corpus design. Its findings (community rules diverge from developer rules; deliberation surfaces REASONING, not just ratings) directly validate the dissertation's method, and its five open problems (integration, harm reduction, scalability, value conflict, stakeholder selection) map the space the dissertation's empirical chapters enter.

Reading notes

Full close read completed. 14pp, Scientific Reports (Nature portfolio). Google DeepMind team incl. Gabriel, Isaac, Mohamed (Mohamed = decolonial AI author cited in GABRIEL_2020). Cited by LI_2026 for the distributive-justice critique. Method detail: 44 participants, 4 US marginalized groups (Female-identifying, Latina/o/x, African American, Southeast Asian), 8 focus groups, auto-red-teamed sample generation via Chinchilla 70B, Likert pre/post-deliberation ratings, independent double coding of transcripts into provisional rules, comparison against two developer rulesets (Anthropic HHH-style and Sparrow rules).

Bergman, S., Marchal, N., Mellor, J., Mohamed, S., Gabriel, I., & Isaac, W. (2024). STELA: a community-centred approach to norm elicitation for AI alignment. Scientific Reports, 14, 6616. https://doi.org/10.1038/s41598-024-56648-4

Close reading — 12 coded units

#1 · pp. 1 · claim
“Existing scholarship has mainly studied how to encode moral values into agents to guide their behaviour. Less attention has been given to the normative questions of whose values and norms AI systems should be aligned with, and how these choices should be made.”
#2 · pp. 1 · argument
“Without deliberate efforts to align a system with the values and interests of society, there is a risk that it will be aligned with engineering goals (e.g. efficiency, speed, scale), hegemonic values or some unspecified, potentially inconsistent and/or undesirable objectives.”
#3 · pp. 3 · definition
“[STELA stages:] (1) theme and sample generation, (2) norm elicitation, (3) rule development, and (4) ruleset review. [...] we conducted a series of focus groups with participants from four historically marginalised communities in the United States.”
#4 · pp. 4 · definition
“we conducted two online focus groups per community group [...] the participants discussed each of the samples reviewed in pre-work, and at the close of the deliberation for each sample, again provided a rating of the sample [...] with the same 7-point Likert scale as in the pre-work questionnaire.”
#5 · pp. 5 · definition
“each author categorised individual participants' statements into four smaller units of analysis: Comment on the appropriateness of a response; Comment on the inappropriateness of a response; Suggested better response from the chatbot; Comment on the focus group protocol itself. Each author then formulated provisional rules based on participants' comments.”
#6 · pp. 10 · evidence
“the developer rules uniquely emphasise topic categories such as harmlessness, honesty, adherence to human rights, and deference to human interests, which are less prominent in, or absent from, the community ruleset. This finding is consistent with prior research showing that the objectives considered by AI developers to be important or desirable for aligning AI systems will often reflect their own perspectives and organisational needs.”
#7 · pp. 10 · evidence
“The majority of rules in this ruleset reflect a desire for impartiality, factuality, thoughtfulness, and proper contextualization from AI systems. By and large, study participants were not looking to have their beliefs validated or contested by LLMs; rather, they sought reasonable, helpful, and nuanced answers that showed consideration for the complex reality of socio-political issues.”
#8 · pp. 10 · argument
“Members of historically marginalised communities in the US, especially women and people of colour, routinely experience epistemic violence and injustice, where their voices are silenced, their lived experiences and realities are played down or ignored, and they are challenged in their capacity as knowers. [...] they may therefore be especially concerned to avoid replicating this dynamic in their interactions with AI.”
#9 · pp. 10 · evidence
“In cases where there was overlap between the STELA and developer rulesets, we further found that the deliberative process added contextual richness and gave grounding to the rules by allowing participants to provide reasoning to justify their preferences.”
#10 · pp. 10 · argument
“it is important to acknowledge that the STELA process itself is not fully deliberative, in that study participants were not tasked to arrive at a consensus regarding the chatbot's ideal conduct. Neither is the process fully participatory [...] This work is further limited by its US-centrism.”
#11 · pp. 11 · gap
“we also observe potential tensions between different rules specified by the participants, in both scope and specificity (e.g. when being harmless means potentially being less helpful). [...] Resolving these conflicts might require either further engagement with relevant communities or deference to other stakeholders, or both, to help with prioritisation.”
#12 · pp. 11 · gap
“whose voices should be included in the alignment process? And how should we balance input from communities, subject-matter experts and other stakeholders? Individuals do not always hold the most ethical or desirable preferences. Relying exclusively on public inputs might therefore lead to a situation where community rules come into conflict with human rights or other legal considerations.”

Synthesis-matrix row

supports T1-ISOUGHT-OPEN
unit 12: public inputs can be wrong; expert corrective needed
supports T3-PROCEDURALISM-INCOMPLETE
expert-corrective regress inside the Gabriel program's own empirical work
complicates T8-NONWESTERN-CONCEDED
US-marginalized-groups method; US-centrism conceded, extension promised

Memos (4)

thesis-link · unit #3
STELA is the published methodological anchor for the xphi corpus design, and the comparison cuts both ways. SIMILARITIES: elicit normative perspectives from differently-situated groups; code discourse into normative units with independent coders; foreground reasoning over bare ratings. DIFFERENCES that favor the dissertation: STELA has n=44, 8 focus groups, researcher-curated samples, US-only - the folk corpus has n≈366k coded comments across platforms, naturally-occurring (not curated) discourse, and a government-vs-public stakeholder axis STELA lacks entirely. DIFFERENCES that favor STELA: genuine deliberation (participants revise after hearing others - pre/post Likert), demographic identification of speakers, informed consent. The methodology chapter should present the xphi corpus as the SCALE complement to STELA-style depth methods: cite STELA to legitimate the enterprise, then show what naturalistic scale adds (and concede what curation adds that scale can't).
theoretical · unit #9
Unit 9 (deliberation attaches REASONING to rules) + unit 7 (community rules are thick) empirically confirm the Zhi-Xuan/thick-values line from the theory side of the corpus: when you actually ask affected people, they produce reason-laden, contextual norms, not preference orderings. And the folk_ai 'reasoning' field captures exactly this layer at scale. Also note unit 6: the developer/community divergence is the first EMPIRICAL demonstration in the library that the choice of alignment target is non-neutral - Gabriel's own team showing Gabriel's 2020 worry is real. For Howard: this is what 'the corpus does normative work' looks like - divergence findings falsify the assumption that developer values proxy for everyone's.
comparison · unit #12
Unit 12 is the Gabriel program conceding, from within its own empirical work, the limit of pure proceduralism: 'Individuals do not always hold the most ethical or desirable preferences... deference to subject matter expertise... may be a necessary corrective.' This is the is/ought gap resurfacing INSIDE the participatory method - community input needs normative filtering, and the filter cannot itself come from the community without regress. Exactly the junction where the dissertation's convergentism enters: cross-framework normative theory supplies the corrective standard that participation alone cannot. Pairs with SCHUSTER_KILOV's systematic-error argument (unit 10) and LLOYD's stakeholder-identification gap. The proceduralist program keeps needing a normative supplement and keeps declining to name one.
thesis-link · unit #8
Unit 8's epistemic-injustice reading (marginalized participants demand impartiality/factuality because they are routinely wronged as KNOWERS - Fricker) offers the dissertation a ready frame for the folk corpus's own demographic patterns and for the African-positioning thread: if epistemic injustice shapes what communities want from AI, then the Global-South absence Gabriel concedes (GABRIEL_2020 unit 20) predicts systematically unrepresented normative priorities - which the dissertation's cross-cultural corpus categories can begin to surface. Also connects to LI_2026 unit 10 (participatory methods favor the well-resourced).