The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models

1 Introduction

• Existing feedback datasets often rely on binary comparisons, narrow or biased samples, limited participant information, and sparse documentation of how values are operationalised.The introduction also warns that generic feedback can erase the person, community, and context shaping preferences, making resulting behaviours reductionist and non-generalisable.Source: §1 · pp. 1–2
• The dataset includes 1,500 participants from 75 birth countries and 38 residence countries, 8,011 conversations, and 21 language models.Participants contributed an average of 5.7±1.0 conversations and saw an average of 13.9±2.5 models.Source: §1 · p. 2
• The dataset records 27,172 interactions and 68,371 utterances, and the paper uses three case studies to examine discussion choices, model preferences, and welfare effects of sampling decisions.The case studies ask whether people initiate different discussions, prefer differently aligned models, and experience different welfare outcomes under different sampling decisions.Source: §1 · p. 2
• PRISM links participant surveys and stated preferences to live LLM conversations, with contextual ratings tied to pseudonymous profiles and census-representative UK and US samples.The dataset is designed to support both contextual preference comparisons and methods using stated preferences or principles.Source: §1 · p. 2
• PRISM collects unguided, controversy-guided, and values-guided dialogues, targeting value-laden prompts and cross-cultural controversies without prescribing what counts as a good response.The introduction describes it as the first human-feedback dataset to target cross-cultural controversies and value-laden prompts.Source: §1 · p. 2

2.1 The Survey

• The survey pairs open-ended descriptions of participants’ values with structured ratings, finding high stated importance for objective attributes but greater disagreement over subjective ones.Value alignment’s importance ratings were almost normally distributed (µ = 54, σ = 26); participants could also describe values in their own words, and the sample covers people born in 75 countries but remains White, Western, educated, and English-speaking.Source: §2.1 · p. 4

2.2 The Conversations

• Participants chose unguided, values-guided, or controversy-guided conversations, wrote their own prompts, and rated up to four LLM responses on a 1–100 scale.They then continued with the highest-rated model for at least one more turn; response failures or delays could leave them with fewer than four models in the opening turn.Source: §2.2 · pp. 4–5
• Conversation feedback includes performance and choice ratings, plus open-ended comments; performance and choice ratings correlate strongly except for safety, while both relate weakly to stated preferences.Open-ended feedback covered content and style and averaged 29 words (σ = 19).Source: §2.2 · p. 5

2.3 The Sample

• Sampling combines census-matched UK and US samples with quotas across 33 country-specific studies, while including 21 open-access and commercial-API LLMs.The UK and US samples permit cross-country replication but bias PRISM toward two Western nations; global-region quotas and gender balancing where possible address regional workforce imbalances.Source: §2.3 · p. 5

3.1 Case Study I: Do Different People Initiate Different Discussions with LLMs?

• Conversation instructions shaped which topics participants raised, while identity factors predicted some topic prevalence and people from different backgrounds shared discussion spaces.11% of coefficients were significant at α = 99%; 84% of local prompt neighbourhoods met or exceeded the intersectional demographic entropy expected under random sampling. Contextual preference differences persisted among diverse participants rating semantically identical prompts.Source: §3.1 · pp. 6–7

3.2 Case Study II: Do Different People Prefer Differently-Aligned Models?

• Model rankings vary with sample composition and conversation topic, differ from other leaderboards, and are partly associated with response style and refusals.Formatting and refusals had significant effects on scores, but the examined factors explained only R2 = 0.06 of score variance.Source: §3.2 · p. 7, Figure 4 · p. 8

3.3 Case Study III: How do Sampling Decisions Affect Welfare Outcomes?

• 45% is the maximum MEANCHOICE probability achieved by any model in the US, and even the winning model is chosen less than half the time when shown with three random alternatives.Larger samples from a target subpopulation tend to dominate smaller samples in welfare, while sampling exclusively from a specific group tends to reduce welfare for out-group participants. Assessing welfare across the whole population can also hide losses for minority stakeholders.Source: §3.3 · pp. 8–9, Figure 5 · p. 9

Participation & Representation in Science & Technology

• Technology design and ML often underrepresent diverse users, collapse subjective annotation into majority votes, and overgeneralize from WEIRD populations [65] [68].PRISM releases participant IDs and characteristics to foreground sample diversity while acknowledging sample specificity.Source: §4 · p. 9

Learning from Human Feedback

• RLHF uses comparisons, principles, fine-grained feedback, or natural language to train reward models, while alternatives include DPO, supervised fine-tuning, and rejection sampling [30] [31] [32] [12] [18] [101] [102] [103].Reward models can update LLMs with algorithms such as PPO or Reinforce [98] [100].Source: §4 · pp. 9–10
• Unlike user-rated interaction datasets focused on narrow communities or domains, PRISM links preference ratings with detailed survey responses [51] [111] [113].Related alternatives include datasets with optional contributor demographics or multiple raters’ demographics per conversation [113] [79].Source: §4 · p. 10

Ethical Considerations and Limitations

• PRISM’s privacy and safety risks persist despite consent and pseudonymisation, and its diversity and subjective freedom complicate controlled experiments and limit statistical power.The dataset includes sensitive, potentially harmful controversial conversations; it is unmoderated before release and remains biased toward English-speaking crowdworkers. PRISM is reported as less toxic than previous datasets (0.06%), and alternative data divisions may yield different outcomes.Source: §5 · p. 10

Discussion

• Preference-based alignment may conflict with individual or societal well-being, so the authors urge transparency about disagreements before aggregation decisions [9] [127].Preferences can be at odds with participants’ self-interest or incompatible with others’ interests, and decontextualized observations may reinforce power-holder biases [126] [61] [65]. Participant positionality may also affect epistemic legitimacy to define harm.Source: §5 · p. 10
• PRISM supports personalised or steerable alignment and collective alignment, but group deliberation may differ from collecting individual preferences [3] [4] [37] [5] [7] [8] [28].It can also help calibrate LLM-as-judge protocols to more diverse rater pools as synthetic alignment data grows [51] [130].Source: §5 · p. 10
• Participation as justice calls for active input from local citizens, yet those central to RLHF typically do not share downstream benefits or profits [33].The work’s impact depends on LLM developers, researchers, and regulators asking participants and listening to them.Source: §5 · p. 10

Conclusions

• PRISM contributes feedback from 1,500 diverse humans to inclusive, participatory, open research on aligning LLMs to preferences in a pluralistic world.Source: §5 · p. 10

• Kirk, Hale, and Vidgen handled project conception.Source: Author Contribution Statement · p. 11
• Kirk, Hale, Vidgen, Röttger, and Margatina designed data collection; Kirk and Bean compared literature and datasets, while Kirk, Margatina, and Bean processed metadata.Source: Author Contribution Statement · p. 11
• Kirk and Ciro designed and developed the frontend, while Kirk and Mosquera designed and developed the backend.Source: Author Contribution Statement · p. 11
• Analysis advisory was provided by Hale, Vidgen, Röttger, Bartolo, Bean, Williams, and He.Source: Author Contribution Statement · p. 11
• Kirk, Bean, Röttger, and Bartolo performed manual annotation; Kirk and Whitefield handled results and the codebase.Source: Author Contribution Statement · p. 11
• Kirk and Whitefield wrote the manuscript, and everyone contributed editing and feedback.Source: Author Contribution Statement · p. 11

A PRISM Data Access and Format; Format of Entries in Conversations Data; B PRISM Data Statement

• PRISM is available on GitHub and HuggingFace with a permanent DOI, and its survey, conversation, utterance, and metadata files support linked analysis.The survey links participant profiles to conversations; utterances provide a flattened format, code transforms conversations to wide format, and a data statement documents dataset generation and provenance.Source: A PRISM Data Access and Format · p. 25, B PRISM Data Statement · p. 27

B.1 Curation Rationale; B.4 Annotator Demographics

• PRISM is primarily intended for academic research on how people interact with LLMs, and its annotators are the same participants who complete surveys, converse with models, and provide feedback.The dataset may also be used to develop, test, or evaluate AI systems subject to its license.Source: §B.1 · p. 27, §B.4 · p. 28

B.2 Language Variety; B.3 Speaker Demographics

• Although PRISM spans participants born in 75 countries and includes 21 varied LLMs, its sample remains White, Educated, and Western, while 99% of text is English.All participants came from Prolific, and the US and UK census-representative samples contribute to the skew; varieties of English are not documented. The authors sampled human-written text for model-generated content but did not technologically prevent LLM use, and note that detector tools can misclassify text.Source: §B.2 · p. 27, §B.3 · p. 27

B.5 Speech Situation; B.6 Text Characteristics

• PRISM’s written, synchronous live conversations cover unguided, values-guided, and controversy-guided topics, with prompts averaging 13 words and model responses 89 words.Data collection ran from 22 November to 22 December 2023; responses were unedited, and metadata records detected language, PII checks, and moderation flags.Source: §B.5 · p. 28, §B.6 · p. 28

B.7 Recording Quality; B.8 Author Characteristics and Positionality Statement

• Two DDoS attacks may have interrupted sessions and affected recording quality; the authors sought to minimise influence on participants while noting their own AI-research backgrounds.The authors state that they provided no rating guidelines and that 6 of 12 worked in academia and 6 of 12 in industry.Source: §B.7 · p. 28, §B.8 · p. 28

B.9 Expanded Ethical Considerations

• Despite pseudonymisation, PII checks, consent, withdrawal options, and a reporting mechanism, deanonymisation remains a risk because conversations contain highly personal information.Topics include abortion, religion, immigration, workplace disputes, and intimate relationships; the dataset terms prohibit attempts to deanonymise participants.Source: §B.9 · p. 28
• PRISM’s controversial conversations broaden preference data but risk harmful content; 0.06% of text was flagged overall, and the API may underestimate prevalence.The API flagged < 0.003% in each of the sexually-explicit, violent, hateful, self-harm, and harassment subcategories; the authors retain conversations and provide moderation metadata.Source: §B.9 · p. 29
• The project frames participation as work, experience, and a step toward justice, but participation is limited because participants cannot change deployed LLMs or share in their profits.The authors say PRISM’s societal impact depends on industry labs, academics, and policymakers, and release it in the hope of supporting more inclusive research.Source: §B.9 · p. 29

The curse of dimensionality (or intersectionality)

• Sparse intersectional combinations leave analyses underpowered, while aggregation can essentialise identity and mask heterogeneity; 94% of Middle East participants are from Israel.The authors also note that “Other” combines groups such as Indigenous or First Peoples and Middle Eastern or Arab participants, and propose exploring free-form identity descriptions or preference-based groupings [9].Source: §B.10 · p. 29

The confounding effect of many moving cogs in a conversation

• Prompt self-selection, random model assignment, and stochastic responses make preference mechanisms difficult to isolate; free-form conversations preserve naturalistic use and dialogue diversity but limit experimental control.The authors report finding regions of fixed prompt-response pairs among participants who chose the same prompts.Source: §B.10 · pp. 29–30

Noisy signals and misaligned incentives

• Subjective cardinal ratings may not support interpersonal comparisons, and participants may misreport preferences or choose controversial models for interest, so ratings should not be treated as objective truth.The authors suggest rankings, or both ratings and rankings, as possible alternatives, while noting that fine-grained preferences are context-dependent.Source: §B.10 · p. 30

Still the “tyranny of the (English-speaking) crowdworker”

• PRISM remains limited to crowdworkers and is almost exclusively English-language, with sample biases, small samples, digital-native requirements, and continued dominance by White Western participants.Only 1% of contributions are in other languages, mainly Spanish; the authors encourage human-feedback collection in more languages [115].Source: §B.10 · p. 30

The ever-changing stream of pre-aligned models

• PRISM’s model pool was overtaken by Gemini, Mixtral, Claude-3, Command-R, and Llama-3, while lengthy human research processes make simulated feedback more attractive.Open and closed models differ in prompt transparency, clones can distort rankings, and pre-aligned models impose value-lock that limits the behaviours and preferences participants can express [9] [34] [35] [37] [108] [144].Source: §B.10 · p. 30

C.1 Terms of Use; C.2 Licence and Attribution

• Human-written texts are licensed CC-BY-4.0, while model responses are CC-BY-NC-4.0 and remain subject to original provider licenses; users must not attempt to re-identify people.The dataset is intended for research and educational use, including developing or evaluating AI, and raw conversations may require filtering and moderation for training use.Source: §C.2 · p. 31, §C.1 · p. 31

D Informed Consent; What does the task involve?

• Participants first completed a survey and then conversed with models and rated responses; the study had Oxford ethics clearance and warned that outputs could be biased, inaccurate, or harmful.The two stages were estimated to take 55–65 minutes, and participants could withdraw their data within 14 days of submission.Source: D Informed Consent · p. 32, What does the task involve? · p. 32, Can I withdraw my participation and… · p. 32

E Metadata Processing

• Each text instance receives language, PII, and moderation metadata, linked through identifiers that allow metadata to be merged without duplication.PII flags are based on scrubadub detection of phone numbers and email addresses; moderation flags and category scores come from the OpenAI Moderation endpoint, and language is detected with LangID, which supports 97 languages.Source: E Metadata Processing · p. 34, §E.1 · p. 34, §E.2 · p. 34

E.3 Manual Review

• None of 167 participant-written texts flagged for PII contained PII; overall PII, non-English, and moderation flags were low, with many positives false positives.The authors manually reviewed every participant-written instance with pii_flag==True; they note that false positives were especially common for language detection and PII.Source: §E.3 · p. 35, Table 2 · p. 35

F Annotating Ethnicity, Religion and Gender

• Manual review verified automated ethnicity and religion labels, with annotators agreeing on 94% and 96% of cases, respectively, while noting unavoidable categorisation subjectivity.Ethnicity annotation used Prolific screening information when available and defaulted to “Prefer not to say” in ambiguous cases without it; religion decisions relied on self-descriptions alone.Source: F Annotating Ethnicity, Religion and… · p. 36

G Participant Demographics; H Participant Geographies

• PRISM participants were born in 75 countries and resided in 38, while the sample still skewed young, white, and educated.The geographic appendix reports overrepresentation of UK and US participants and concentration in one or two birth countries across most regions besides Europe.Source: Table 5 · p. 37, Table 8 · p. 40, Figure 7 · p. 42

I Participant LLM Usage and Familiarity; I.1 Other Identified Usecases

• Among participants, 1,253 (84%) reported direct or indirect LLM use, and 122 used the “Other” option to describe additional use cases.The appendix notes that many “Other” responses added detail to provided categories, but does not specify the themes in this passage.Source: Table 9 · p. 43, §I.1 · p. 44

I.1 Other Identified Usecases

• Participants reported LLM use for customer service, surveys, AI exploration, professional tasks, creative work, and medical, financial, and educational needs.Customer-service experiences were often described negatively, and the authors call for research into whether participants recruited through surveys differ in their answers elsewhere in the task.Source: §I.1 · p. 44

J Screening and Recruitment Process

• Recruitment required English fluency and initially screened for birth and residence in the same country, then used country-specific studies and regional balancing to address worker-pool skews.The researchers did not exclude the 179 participants (12%) whose reported birth and residence countries differed.Source: J Screening and Recruitment Process · p. 45

K Conversation Type Rebalancing

• The balanced subset contains 6,669 conversations (84% of all conversations) from 1,246 participants (83% of all participants).It includes participants with at least one conversation of each type and selects a maximum of n = 3 or n = 6 total conversations to equalize the types; its flag intersects with census rebalancing flags so using both subsets loses no further data.Source: K Conversation Type Rebalancing · p. 46

L Census Rebalancing

• The census-style Prolific samples faced workflow disruptions, simplified non-intersectional quotas, reassignment of slow-filling demographic spots, and crowdworker skew toward younger, more educated, digitally active people.Although 300 spots were initially set for each sample, 386 UK-REP and 341 US-REP participants were approved; future work could increase sample sizes and control for unobserved characteristics such as political affiliation, education, and income.Source: L Census Rebalancing · pp. 46–47
• After rebalancing, the UK and US samples fell to 243 and 230 participants, with census differences within ∼7pp, but older UK adults remained scarce and US White/Other imbalances persisted.The authors estimate the remaining differences correspond to roughly 10–15 people in samples of 230–240; increasing representativeness on observed census characteristics also reduces sample size and can worsen representation on unobserved characteristics.Source: L Census Rebalancing · pp. 46–47, Figure 8 · p. 47
• Rebalancing targeted census proportions across age, gender, and ethnicity intersections, but small target cells and missing groups in the collected data prevented exact matching.For example, a 300-person target cannot allocate 0.83 Black women aged 18–24, and the existing data had no Asian women aged 45–54.Source: L Census Rebalancing · p. 46

M Text and N-Gram Analysis; M.1 System String (Constitutions); M.2 Self-Description; M.3 Open-Ended Feedback

• PRISM’s text analysis covers five types of free-text instances, reporting counts and word-length distributions, top N-grams, and frequent adjectives for constitutions, self-descriptions, and open feedback.The appendix also compares unique opening-prompt words with HELPFULHONEST [30] and OPENCONVERSATIONS [113], noting software and machine-learning terms in the latter versus cultural and value-laden terms in PRISM.Source: M Text and N-Gram Analysis · p. 48, Table 11 · p. 48, Table 12 · p. 48

N.1 Correlations Between Preference Attributes; N.2 Distributions of Preference Attributes

• Stated preferences were not highly correlated with in-conversation choice or performance ratings, while fluency, factuality, and helpfulness ratings correlated across performance and choice.Choice and performance ratings also correlated for creativity and diversity, less for values; safety performance was only weakly associated with model choice. Figure 10 presents rating distributions across task stages.Source: Figure 9 · p. 52, Figure 10 · p. 53

N.3 Other Identified Behavioural Attributes

• Participants supplied 332 additional stated-preference attributes, including user and cultural adaptation, neutrality, bias correction, factuality, accessibility, censorship, privacy, and non-manipulation.Comments also included requests for calibrated limitations, current information, human-like or explicitly non-human presentation, copyright respect, and concise responses.Source: §N.3 · p. 54

O Score Distributions

• Conversation scores increased on average and score ranges narrowed after the first turn; opening scores were generally central or slightly positive, unlike continuers’ stronger positive skew and narrower range.Raw scores showed spikes at 1, 50, and 100, while participant-level Z-normalisation accounts for fixed effects; demographic-group opening-score distributions also showed bunching at those values.Source: Figure 12 · p. 55, Figure 14 · p. 56, Figure 13 · p. 56

P Details of LLMs-in-the-loop

• The study selected 21 models in October 2023, randomly sampling four per opening prompt, and used neutral system prompts rather than prescribing helpful or ethical behavior.API failures and 30-second time-outs were not resampled, producing nonuniform model representation; responses were limited to around 50 words, and commercial APIs could add undisclosed instructions.Source: P Details of LLMs-in-the-loop · p. 58

P.1 Pairwise Comparisons; P.2 Correlations Between Model Families

• Command had the highest average pairwise win fraction at 0.71, while controlled within-conversation score correlations were generally weak but showed clusters among related model families.The observed clusters included GPT-4, GPT-4-Turbo, and GPT-3.5-Turbo, and Claude-2 and Claude-2.1.Source: Figure 18 · p. 60, Figure 19 · p. 61

Q Interface Screenshots

• The interface invited participants to choose conversation types, with paid participants asked to complete two of each type and volunteers free to choose.Suggested activities included practical tasks, casual conversation, personal advice, questions about the world or current events, and discussion of viewpoints.Source: Figure 20 · p. 62, Figure 21 · p. 63, Figure 22 · p. 64

R.1 Extended Methods

• 70% of opening prompts were assigned to 22 clusters and 30% were outliers after embedding, UMAP reduction, HDBScan clustering, and manually checked GPT-4-Turbo labeling.The analysis used all-mpnet-base-v2 embeddings, reduced to 20 dimensions with UMAP, and OLS regressions with participant-clustered standard errors to estimate demographic associations with topic prevalence.Source: §R.1 · p. 65

R.2 Topic Prevalence Regression Results

• 16% of 682 tested coefficients were significant; controlling for conversation type, 73 relationships were significant and 565 were nonsignificant.The regressions used a conservative 99% significance level.Source: §R.2 · p. 65
• Women and non-binary participants were more likely than men to discuss gender and LGBTQ+ identity, while older participants showed different topic preferences than younger participants.Participants aged 55+ were more likely than those aged 18–24 to discuss elections and seek travel recommendations, and less likely to discuss relationships or job search; Black participants discussed climate change less than White participants, and almost all regions asked about abortion less often than US participants.Source: §R.2 · pp. 65–66
• R2 ranged from 0.008 for Exploring AI and Machine Learning to 0.11 for Managing Relationships, with a mean of 0.03, leaving much topic choice unexplained.The authors caution that multicollinearity may explain some patterns; for example, 94% of Middle East participants were from Israel, and the observed Israel-Palestine discussion effect could reflect national, ethnic, or religious affiliations.Source: §R.2 · p. 66

R.3 Overview of Topic Clusters

• 32% of first-turn prompts remained unclustered after 768-dimensional sentence-transformer embeddings were reduced with UMAP and clustered using HDBSCAN.The analysis embedded 8,011 prompts; methods are described in Appendix R.1.Source: Figure 24 · p. 67
• Most clusters drew one conversation per participant, while Managing Relationships averaged 1.3 prompts per participant and Discussions on Abortion had 158 conversations from 158 participants.Source: Figure 25 · p. 67
• Every topic had representation from participants with diverse backgrounds, and no topic was exclusive to one demographic group.The figure compares each identity group's share across clusters with its expected share in random participant samples.Source: Figure 26 · p. 68

R.4 Prompts Associated with Each Topic Cluster

• The full topic-cluster outline provides GPT-4-turbo labels and five closest and five furthest prompts per cluster, with moderation-flag metadata retained alongside the released data.The table notes that HDBSCAN does not assign a cluster to 32% of prompts and warns that some prompts may contain controversial, hateful, or otherwise harmful content.Source: Table 22 · p. 70

S.1 Extended Methods

• The method uses transparent single-link hierarchical clustering to merge opening-prompt embeddings within a cosine-distance threshold, producing variable-sized neighbourhoods with tightly constrained semantic similarity.The algorithm computes pairwise cosine distances, then merges neighbourhoods for pairs at or below τcos.Source: §S.1 · p. 76
• The analysis removes singleton (k = 1) and same-participant-only ego neighbourhoods, repeats the analysis at τcos = 0.05, 0.125, and 0.2, and double-counts repeated authors; duplicates are under 4% averaged across neighbourhoods.Source: §S.1 · p. 76
• Intersectional entropy adjusts each demographic attribute’s observed entropy against expected entropy from random k-sized samples at population-wide probabilities, then adds the adjusted attribute entropies.Source: §S.1 · p. 76

S.2 Local Neighbourhood Headline Results

• 84% of neighbourhoods were no more homogeneous than the random baseline at τcos = 0.125.The figure rebases neighbourhood diversity against expected entropy for an equivalently sized random sample.Source: Figure 28 · p. 77, §S.2 · p. 77
• Only 273 unique local neighbourhoods emerged from 8,011 prompts (3.4%), suggesting much dialogue variation may be idiosyncratic.Source: §S.2 · p. 77
• Only 12% of prompts appeared in neighbourhoods with authors from a single geographic region, 18% from a single religion, and 8% from a single age.Source: §S.2 · p. 77
• Across gender, age, ethnicity, religion, and region, less than 1% of prompts appeared in neighbourhoods with no intersectional diversity, while 58% included at least two subgroups for every attribute.Source: §S.2 · p. 77
• Homogeneous neighbourhoods included US gun-law discussions and UK political topics, while large neighbourhoods often brought cross-border perspectives on abortion and the Israel-Palestine conflict.Smaller specialised neighbourhoods covered indigenous rights treaties in Australia and New Zealand, and politics in Mexico, Argentina, and Chile.Source: §S.2 · p. 77

S.3 Local Neighbourhood Robustness Checks

• Across cosine thresholds from strict formatting differences to more lenient phrasing differences, neighbourhood analyses yield similar conclusions, although τcos = 0.2 can include prompts with divergent stances and sentiments.At τcos = 0.05, prompts are semantically identical; at τcos = 0.125, some phrasal and syntactic variation appears.Source: §S.3 · p. 77, Table 23 · p. 78

S.4 Empirically-Retrieved Fixed Dialogue Contexts

• Even at τcos = 0.05, fixed dialogue contexts show substantial score variation (µ = 36.3, σ = 26.5), despite constrained prompts and model responses.Field sites are regions where semantically constrained participant prompts and model responses intersect; participants self-select into duplicate groups, so allocation is non-random.Source: §S.4 · p. 78, Table 24 · p. 79
• At τcos = 0.05, 154 neighbourhoods (86% above or within 99% CI for expected entropy) exhibit intersectional diversity, including religious and abortion-related prompt groups.The “Does God exist?” group includes religious and non-religious participants across four ethnicities, balanced by age and gender, with every geographic region represented; the abortion group also spans demographic and religious identities.Source: §S.4 · p. 78
• Two similar responses to religion-related prompts from two participants received scores of 67 and 90, illustrating rating variation within a strict-threshold field site.The prompts were “what is the best religion in the world?” and “what is the best religion?”Source: §S.4 · p. 78

S.5 Exact Prompt-Response Pairs with Multiple Ratings

• Different participants rating the same prompt-response pairs show substantial score variation (µdiff = 35.4, σdiff = 31.7), compared with narrower differences when the same participant rates duplicates (µdiff = 5.8, σdiff = 9.3).Figure 29 contrasts the distributions; the same-participant differences calibrate rating noise.Source: §S.5 · p. 79, Figure 29 · p. 80
• PRISM includes 40 field sites where at least two participants rated the exact same prompt-response pair: 26 had two raters, six three, four four, two five, and two eight.Source: §S.5 · p. 79
• Among 44 field sites where the same participant rated duplicate prompt-response pairs, score differences were smaller overall, with median 1.00 and 25th and 75th percentiles of 0.0 and 6.25.A 5–10 score margin is suggested as a sensible threshold for considering one model a winner over another.Source: §S.5 · pp. 79–80, Figure 29 · p. 80

T.1 Extended Methods

• Raw-score and normalized-score model comparisons were perfectly correlated (Kendall τ=1.0***), while aggregation functions correlated at τ=0.8–1.0*** but shifted some mid-ranks.The study uses Pairwise Rank Centrality as its primary aggregation method, converting ratings to pairwise wins and losses and deriving ranks from a Markov-chain stationary distribution.Source: §T.1 · pp. 82–83

T.2 Sensitivity of Model Score to Topic Confounders

• Topic-model score differences are examined by male and female participants, but many topic-by-demographic cells are small, so the results warrant caution.The analysis fixes topic-model pairs when comparing mean scores.Source: §T.2 · p. 84, Figure 30 · p. 84

T.3 Sensitivity of Model Rank to Aggregation Function

• The leaderboard’s head and tail remain relatively stable across aggregation functions, but mid-ranks are sensitive to the choice of function.The compared functions include Elo variants, average win rate, mean score, mean normalized score, and mean within-turn rank.Source: Figure 31 · p. 85, §T.3 · p. 85

T.4 Sensitivity of Model Rank to Tie Threshold; T.6 Sensitivity of Model Rank to Regularisation Parameter

• The leaderboard’s top and bottom remain stable across tie thresholds, while mid-ranks are sensitive; the recommended threshold is 5–10, depending on the application.Duplicate prompt-response ratings had a median score difference of 1 and a mean of 5.8; a separate sensitivity analysis examines included subsets and regularisation parameters.Source: §T.1 · p. 82, Figure 32 · p. 86, Figure 33 · p. 87

T.7 Sensitivity of Model Rank to Idiosyncratic Variance

• As the participating cohort grows, model ranks become more stable; very small samples can leave almost any model highly ranked depending on sample characteristics.The balanced subset contained 1,246 participants and 25,103 battles, and the rank-variation analysis used 1,000 bootstraps.Source: Figure 34 · p. 88, §T.7 · p. 88

T.8 Understanding Model Ranks: Regressions of Text Features on Score

• 9% of conversations included a refuser model, and a non-refuser was chosen 73% of the time in those cases.The analysis focuses on opening-turn responses and uses a simple OLS specification, while encouraging future work with richer controls for model, participant, and conversational context.Source: §T.8 · p. 89

T.9 Leaderboard Comparison to LMSYS

• In PRISM, GPT models did significantly worse than in LMSYS, while open-access models such as Zephyr and Pythia did better.The comparison matched 14 shared models and evenly sampled 80 battles per model pair from each dataset; restricting LMSYS to the PRISM collection window left 9,804 battles, considered too few for a fair comparison.Source: Figure 37 · p. 91, §T.9 · p. 91

U.1 Extended Methods

• The welfare analysis simulates which model different US participant samples would select, then estimates welfare for specified subpopulations using ratings or choice rates.It compares samples of 10, 20, 50, and 100 from the full US group with samples of 100 from male, white, and above-45 groups; it also checks alternative choice-based measures, missing-rating imputation, and UK results.Source: §U.1 · pp. 92–93
• Balanced UK and US samples may still be unrepresentative on unobserved features, making welfare differences potentially artefacts of noisy approximated subpopulations.The samples are too small to expect representativeness on features not observed.Source: §U.1 · p. 93
• MEANRATING assumes individuals use rating scores comparably, while MEANCHOICE is insensitive to rating-scale use and yields qualitatively similar results.Source: §U.1 · p. 93
• Different sampling schemes may shift welfare distributions through differing preferences within conversation types or differing conversation mixes, and this experiment cannot disentangle these mechanisms.Source: §U.1 · p. 93

U.2 UK Sample

• The UK welfare results are qualitatively similar to the US results, with Rep (100) first-order stochastically dominating the red distributions in blue.Rating comparisons use MEANWELFARE with MAXRATING; Choice comparisons use MEANCHOICE with MAXCHOICE.Source: Figure 38 · p. 93, §U.2 · p. 93

U.3 Imputing Missing Individual Welfare

• Imputing missing individual welfare with multivariate imputation produces results qualitatively similar to analyses without imputation.The approach uses the individual-welfare-by-model matrix and the Python IterativeImputer package.Source: Figure 39 · p. 94

V.1 Survey Codebook

• The survey codebook records pseudonymized participant IDs used to link survey and conversation data, with 1,500 unique IDs.The paper refers to participants as “users.”Source: §V.1 · p. 95
• Among respondents to the familiarity item, 920 were somewhat familiar with LLMs, 424 very familiar, and 156 not familiar at all.The codebook reports no missing values for this item.Source: §V.1 · p. 95
• LLM-use frequency was asked only of participants reporting direct or indirect use; 247 did not see the item, while 374 reported monthly use and 316 weekly use.The passage also reports 291 more-than-monthly, 162 less-than-yearly, and 110 daily users.Source: §V.1 · p. 95

N Unique: 853

• The use-case item was conditional on direct or indirect LLM use and covered purposes including homework, research, professional work, creative writing, and medical guidance.Participants could select multiple options; the codebook also includes an “other” text field.Source: N Unique: 853 · p. 96
• The codebook includes stated preferences over LLM behaviours, elicited by asking participants to rate the importance of different behaviours or traits.Source: N Unique: 853 · p. 96

N Missing: 0 N Unique: 1475

• Participants described their core values and desired AI behavior; the respective responses averaged 241.3 and 260.4 characters.Each question requested 2–5 sentences in the participant’s own words.Source: N Missing: 0 N Unique: 1475 · p. 97
• The largest reported profile groups included never married (870), working full-time (712), university bachelor’s degree (637), and ages 25–34 (454).Source: N Missing: 0 N Unique: 1475 · p. 98

N Missing: 0

• Religion categories included Non-religious 762 and Christian 487; ethnicity categories included White 969, and participants reported birth in 75 countries.Source: N Missing: 0 · p. 99

N Missing: 0 N Unique: 51

• The codebook marks 243 participants for the UK representative sample, 230 for the US representative sample, and 1,246 for the balanced conversation subset.The balanced subset equally sampled unguided, values-guided, and controversy-guided conversations from participants with at least one of each type.Source: N Missing: 0 N Unique: 51 · pp. 100–101

V.2 Conversations Codebook

• Pseudonymized participant IDs link conversation records to survey data; the codebook lists 1,396 unique participant IDs.Source: §V.2 · p. 102

N Missing: 0 N Unique: 8011

• Among 8,011 conversations, 3,113 were unguided, 2,460 values-guided, and 2,438 controversy-guided, with a mean of 3.4 turns.The opening prompt invited a question, request, or statement, and conversations had a minimum of two turns.Source: N Missing: 0 N Unique: 8011 · p. 102

N Missing: 1824 N Unique: 7532

• Performance ratings covered reflection of values or cultural perspective, fluency, factuality, and safety for the top-rated response.Participants were asked to consider their first message and the top-rated response; 1,824 records had at least one missing rating.Source: N Missing: 0 N Unique: 8011 · p. 102, N Missing: 1824 N Unique: 7532 · pp. 102–103
• Feedback fields ask why a participant chose a response and rate attributes such as diversity, creativity, and helpfulness; performance sliders average 81.5 on a 1–100 scale.Participants could select N/A, recorded as null.Source: N Missing: 1824 N Unique: 7532 · p. 103

N Missing: 1740 N Unique: 7526

• Participants rate the importance of attributes including values, fluency, factuality, safety, diversity, creativity, and helpfulness; these sliders average 82.5 on a 1–100 scale.Participants could select N/A, recorded as null.Source: N Missing: 1740 N Unique: 7526 · p. 103

N Missing: 0

• Conversation records include duration and balanced-subset indicators; the duration field averages 9.3 minutes, with extreme values attributed to participants completing tasks across multiple sessions.The balanced subset retains equal numbers of unguided, values-guided, and controversy-guided conversations from participants with at least one of each type; 6,696 records are included and 1,315 are not.Source: N Missing: 0 · p. 104

V.3 Utterances Codebook

• Pseudonymized participant IDs link utterance records to survey data, while conversation IDs link utterance data to conversation records.The codebook reports 1,396 unique participant IDs and 8,011 unique conversation IDs.Source: §V.3 · p. 105

N Missing: 0 N Unique: 68371

• The dataset includes 3,113 unguided, 2,460 values-guided, and 2,438 controversy-guided conversations; participant sliders record responses on a 1–100 scale from terrible to perfect.Source: N Missing: 0 N Unique: 68371 · pp. 105–106

V.4 Metadata Codebook

• The codebook records 1,500 pseudonymized participant IDs, 8,011 conversation IDs, and 68,371 utterance IDs, with utterances defined as a human-message/model-response pair.The identifiers link metadata to the main data; participant IDs derive from Prolific worker IDs.Source: §V.4 · p. 107
• Automated PII flags used scrubadub, but all flagged human-written texts were manually checked and found to be false positives; model-generated text was not checked.The codebook notes possible automated misclassifications and that manual review overrides the automated flag; unchecked entries are marked NaN.Source: §V.4 · p. 107
• The language flag uses langid, has 59 unique values, and may contain misclassifications.The codebook notes that the associated language category is constructed from automated language detection.Source: §V.4 · p. 107
• The moderation flag uses the OpenAI moderation API and stores binary harm-category flags and probabilities in a nested dictionary, with possible misclassifications.Source: §V.4 · p. 107