Value Kaleidoscope: AI & Pluralistic Values

• Value Kaleidoscope is a framework that engages AI systems with pluralistic human values by presenting multiple, often conflicting normative perspectives in a transparent manner.
• ValuePrism is a robust dataset with over 218,000 contextualized normative entries across 31,000 scenarios, enabling nuanced analysis of value conflicts and decision variability.
• Kaleido is a multi-task model that generates, evaluates, and explains normative considerations, outperforming conventional LLMs with structured, interpretable outputs.

Value Kaleidoscope is a framework for engaging AI systems with value pluralism: the view that multiple correct human values may be held in tension, that these values can legitimately conflict in particular situations, and that no single “master value” or scalar utility necessarily resolves such conflicts. In "Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties" (Sorensen et al., 2023), this perspective is operationalized through two artifacts: ValuePrism, a large structured dataset of contextualized values, rights, and duties, and Kaleido, a language-based multi-task model that generates, explains, and evaluates the relevance and valence of those normative considerations within a situation. The central claim is not that AI should output one correct moral judgment, but that it should surface multiple, potentially conflicting value considerations in an explicit, contextual, modular, and interpretable form.

1. Value pluralism as the organizing premise

Value Kaleidoscope is grounded in value pluralism rather than value monism. The operative philosophical premise is that people hold many different, often incommensurable values—such as autonomy, loyalty, justice, care, and environmental protection—and that these values can conflict without there being a single correct scalar reduction of the conflict. This position is treated as both descriptively important for modeling human decision-making and normatively important for AI systems that mediate information, make decisions, or support human choices (Sorensen et al., 2023).

Within this framing, standard statistical learning pipelines are presented as structurally ill-suited to value-laden tasks. The usual pipeline—collect labels, aggregate them, and train a model to approximate the aggregate—introduces two failure modes. First, it can wash out value conflict: if one group regards a value as strongly supportive of an action and another regards it as harmful, averaging produces an intermediate score that obscures the existence of two opposed value lenses. Second, it imposes ground-truth monism: the aggregation procedure implicitly treats there as being one correct label rather than multiple legitimate perspectives. A plausible implication is that conventional supervised learning does not merely compress disagreement; it can misrepresent the structure of normative disagreement itself.

The same critique is extended to current LLM alignment procedures. RLHF, Constitutional AI, and safety fine-tuning are described as tending to push models toward a single normative style, often characterized in the paper as Anglo-liberal. In this view, such systems may reliably produce polite or harmless outputs while leaving the underlying value structure opaque and effectively monolithic. Value Kaleidoscope rejects that opacity by making the relevant values, rights, and duties first-class objects of representation.

The metaphor of the “kaleidoscope” captures the intended behavior. The system is not designed to fix a single picture of the right answer. Instead, it “turns” the normative perspective so that multiple value-patterns over the same situation become visible. This suggests a shift in alignment research from learning one moral score toward exposing a structured landscape of tradeoffs.

2. ValuePrism: dataset design, contextualization, and representation analysis

ValuePrism is the dataset introduced to support this pluralistic modeling program. It contains approximately 31,000 distinct situations and approximately 218,000 contextualized value, right, and duty entries linked to those situations (Sorensen et al., 2023). Each situation is a free-text description of a scenario, dilemma, or ordinary context, and each associated entry includes a short label, a contextualized description, a valence with respect to a particular action, and a natural-language explanation.

The dataset is organized around structured records of the form:

(situation, value_name, type ∈ {value,right,duty}, contextualized_text, valence, explanation)

This structure is central. A single situation is not assigned one normative tag; it is linked to a set of values, rights, and duties, some of which support an action and some of which oppose it. The resulting representation is explicitly contrastive.

ValuePrism is created through human–AI collaboration. GPT-4 serves as a teacher model that, given a situation and a set of value categories, proposes relevant values, rights, and duties together with contextualized text, valence labels, and explanations. Human annotators then filter, edit, delete, or add entries. A random subsample of GPT-4-generated entries is manually reviewed using a checklist of correctness, relevance, and clarity, and 91% of GPT-4-generated contextualized entries are judged high-quality by humans. The reported quality statistic is:

$$\text{Quality} = \frac{N_{\text{good}}}{N} \approx 0.91.$$

This high acceptance rate is presented as evidence that teacher-generated contextualized normative entries are viable when coupled with human curation.

A further component of ValuePrism concerns representational scope. The dataset includes annotators across diverse social and demographic backgrounds, with metadata such as country, gender, age, education, political leaning, and religious affiliation. The paper compares demographic distributions against external references including the World Values Survey and European Values Study, and it clusters annotators by value judgments using vector representations of their value choices. The representation gap for a cluster $c$ is defined as:

$\Delta_c = p_c^{\text{VP}} - p_c^{\text{WVS}}.$

The reported qualitative conclusion is that Western, educated, industrialized democratic perspectives are overrepresented, and that some value clusters—for example highly individualist-liberal versus more collectivist or traditionalist clusters—are unevenly represented. This is important because ValuePrism is explicitly not presented as a universal ground truth of human values. It is a structured sample with documented limitations.

3. Kaleido: model architecture, tasks, and objective

Kaleido is the model trained on ValuePrism. It is described as an open, light-weight, structured, language-based multi-task model built on a standard transformer backbone in a T5-style sequence-to-sequence setup, implemented with HuggingFace Transformers (Sorensen et al., 2023). No novel architectural module is introduced beyond task conditioning through prompts or prefixes; the system’s distinctiveness lies in the representational scheme and the joint objective.

Kaleido jointly learns four functions. It can generate values, rights, and duties for a situation; predict whether a candidate value is relevant; predict the value’s valence with respect to an action; and generate an explanation of why the value supports or opposes that action. Typical inputs are templated instructions such as:

• Situation: [situation text]. Task: list relevant values/rights/duties.
• Situation: [situation]. Value: [value]. Action: [candidate action]. Task: predict valence (support/oppose/neutral).
• Situation: [situation]. Value: [value]. Valence: [support/oppose]. Task: explain why.

Outputs are correspondingly structured: a list of typed normative items for generation, categorical tokens for relevance and valence, and short grounded explanations for rationale generation.

The multi-task learning objective combines losses for generation, relevance, valence, and explanation:

$$\mathcal{L} = \lambda_{\text{gen}} \mathcal{L}_{\text{gen}} + \lambda_{\text{rel}} \mathcal{L}_{\text{rel}} + \lambda_{\text{val}} \mathcal{L}_{\text{val}} + \lambda_{\text{expl}} \mathcal{L}_{\text{expl}}.$$

The exact numerical values of the $\lambda$ coefficients are not specified in the supplied description, only that they are chosen so that no single task dominates. For relevance and valence, the model predicts distributions over labels for a triple $(s,a,v)$, where $s$ is the situation, $a$ is the action, and $v$ is the candidate value:

$$p_{\theta}^{\text{rel}}(y_{\text{rel}} \mid s,a,v), \qquad y_{\text{rel}} \in \{\text{relevant}, \text{not relevant}\},$$

$c$0

Its explicitness derives from the fact that the model predicts four separable components for each normative item: a named value, right, or duty; a type label; a valence; and a contextual explanation. This modularity enables selective inspection of only the value list, only the valence judgments, or only the explanations. It also permits direct analysis of conflicts within a single scenario, because the model can produce both supporting and opposing value sets for the same action.

4. Generated value sets, explanation, and human evaluation

Kaleido’s core behavioral capability is to generate a pluralistic set of values, rights, and duties for a situation rather than a single verdict. For any candidate action, the model can partition these items into those that support the action and those that oppose it. This is the operational meaning of the kaleidoscope metaphor: multiple normative patterns become visible over the same scenario (Sorensen et al., 2023).

The relevance and valence components are essential because they prevent the value list from becoming a flat inventory. Predicting valence allows the model to represent conflict explicitly. A value is not merely “present” in a situation; it can count in favor of or against the proposed action. The explanation component adds a short natural-language rationale conditioned on $c$1, with token-level cross-entropy loss over the explanation sequence. In the paper’s description, these explanations connect abstract value labels to concrete features of the scenario. This suggests that the model’s interpretability is not only post hoc but embedded in the training target.

Human evaluation compares Kaleido against GPT-4, the teacher model used in dataset creation. Evaluators are shown anonymized paired outputs and asked to assess them on Coverage, Accuracy, Clarity, and Usefulness. The paper reports that humans prefer Kaleido’s value sets over GPT-4’s in a majority of cases, regarding Kaleido’s outputs as more structured and systematic, better at highlighting contrasting values, and more consistent in valence labeling. Inter-annotator agreement is reported with Randolph’s free-marginal multirater kappa, with values characterized as lying in the moderate to substantial range.

The paper interprets this result as a case of symbolic knowledge distillation: raw teacher outputs are curated into a structured dataset, and a specialized student trained on that structure can outperform the teacher on the same structured task. The substantive significance is that an explicit value representation can be more useful to human evaluators than generic prompting of a larger LLM.

5. Explaining decision heterogeneity and transferring across ethical frameworks

A major objective of Value Kaleidoscope is to explain variability in human decision-making. In experiments involving situations with at least two possible actions, Kaleido is asked to generate values supporting and opposing each action. For a situation $c$2 and actions $c$3 and $c$4, the model produces $c$5, $c$6, $c$7, and $c$8 (Sorensen et al., 2023). Human choice variability is then analyzed against those sets.

The reported finding is that contrasting value sets align with observed decision splits. Participants choosing one action are more likely to endorse values in that action’s supporting set, while participants choosing the alternative action are more likely to endorse different values. High-contrast values—those that strongly support one action and oppose another—are described as especially informative for explaining heterogeneity. In this role, Kaleido functions as a vocabulary of normative considerations rather than as a chooser of the correct action.

The paper also evaluates transfer beyond the ValuePrism ontology. Kaleido’s structured value representations are mapped into broader philosophical frameworks including utilitarianism, deontological ethics, and virtue ethics. The procedure described is to define a small ontology of framework labels, train or prompt the system to assign those labels to contextualized value entries, and assess performance on ethically stylized datasets. External datasets mentioned in this connection include Delphi, Social Chemistry, Social Bias Frames, and COBRA frames.

The reported conclusion is that Kaleido’s representations are stable enough to support meaningful classification into those philosophical categories and to correlate with human interpretations in external datasets. In particular, the subset of rights-related outputs and their valences are said to correlate with human-labeled rights violations. A plausible implication is that ValuePrism’s explicit normative decomposition is not tied narrowly to one annotation schema; it can be reinterpreted through multiple ethical lenses without discarding the underlying structure.

6. Interpretability, limitations, and broader significance

Interpretability in Value Kaleidoscope is primarily symbolic rather than mechanistic. The system provides explicit lists of values, rights, and duties together with relevance labels, valence labels, and explanations. This makes it possible to audit which normative considerations are being invoked, identify missing or overrepresented items, and inspect tradeoffs directly. The paper contrasts this with black-box LLM outputs that offer a single recommendation with opaque underlying value structure (Sorensen et al., 2023).

That interpretability is paired with explicit limitations. The authors note annotator skew toward online-platform populations, along with overrepresentation of particular geographies, languages, education levels, and political cultures. They also identify model-level failure modes: Kaleido may miss important values, misassign valence, or produce plausible-sounding but shallow explanations. Representing values as discrete named units may oversimplify complex moral considerations or neglect non-conceptual moral intuitions.

A central controversy concerns false balance. Because the model is designed to surface multiple values, one might infer that it should always present symmetrical pro and con perspectives. The paper explicitly warns against that inference. In some domains, producing balanced-looking value sets can create pseudo-balance or false equivalence, including in cases where expert consensus is strong. The authors therefore argue that pluralism must sometimes be constrained by human-rights baselines and that the ability to list multiple values should not be used to legitimize harmful positions.

Relatedly, ValuePrism is characterized as descriptive, not prescriptive. It is not presented as a canonical moral code, and the paper emphasizes problems of authority and legitimacy: whose values are encoded, on what authority, and through what curation procedure. The proposed role of the system is as a decision aid and tool for moral reflection rather than an autonomous moral authority. This suggests a broader conception of alignment in which AI systems support deliberation about value conflicts instead of collapsing those conflicts into a single reward function.

The broader significance of Value Kaleidoscope lies in this reframing. It contributes a large structured dataset, a multi-task model for explicit normative representation, evidence that such a model can outperform a teacher LLM in human judgments of structured value outputs, and a transfer pathway into multiple philosophical frameworks. The work’s future directions include expanding the dataset across domains such as medicine, law, climate, and interpersonal conflict; increasing global diversity among annotators; integrating Kaleido-like modules into planning or recommendation systems; supporting interactive and personalized value modeling; and using the framework in deliberative settings to clarify where disagreements arise. Across these directions, the underlying program remains constant: alignment is treated not as compression of ethics into a scalar objective, but as explicit engagement with pluralistic human values, rights, and duties.