Value Expression & Alignment in AI

Updated 28 February 2026

Value Expression and Alignment is a domain that formalizes human values as measurable criteria to guide AI behavior and ensure ethical consistency.
It encompasses methods for aggregating preferences, verifying alignment through critical queries, and balancing normative foundations with empirical validation.
The field addresses multidisciplinary challenges including bias, systemic trade-offs, and dynamic calibration in multiagent AI systems.

Value expression and alignment concern the formal, mechanistic, and practical processes by which artificial agents’ behaviors are made to reflect, respect, and implement human values. The domain encompasses the representation of values, formalization of alignment as a property or objective, methods for expressing values in AI systems, criteria for evaluating alignment success, and the sociotechnical context in which these mechanisms operate. Recent literature establishes precise formalisms, taxonomies, and evaluation methodologies, revealing a technical field that is simultaneously normative, empirical, and deeply interdisciplinary.

1. Foundational Concepts: Value Expression, Alignment, and the Formal Problem

At the core of value expression is the representation of “values” as formal, often quantitative, objects—such as preference orderings over states, reward (utility) functions, or logical constraints—that can be embedded into AI agents’ decision-making processes. Value alignment is then defined as the degree to which an agent’s behavior, when governed by internal policies or external norms, advances or maintains these values in a consistent and desirable way.

Formal Representation

Let $S$ denote the set of world-states, $A$ a finite set of agent actions, and $T \subseteq S \times A \times S$ a transition relation. A value $v$ is a preference function $\operatorname{Prf}_v^\alpha(s, s') : S \times S \rightarrow [-1,1]$ mapping state pairs to preference intensities for agent $\alpha$ (Sierra et al., 2021).

Norms—behavioral constraints or rules—induce a “normative world” $(S_n, A, n, T_n)$ by restricting transitions. The degree of alignment between a norm $n$ and a value $v$ for agent $\alpha$ is given by

$\mathrm{Algn}_{n, v}^{\alpha}(S, A, T) = \frac{1}{|\text{Paths}(n)|} \sum_{p \in \text{Paths}(n)} \frac{1}{|p|} \sum_{d=1}^{|p|} \operatorname{Prf}_v^\alpha(p_{d-1}, p_d)$

Alignment can thus be operationalized as the average preference gain per transition under imposed norms (Sierra et al., 2021, Barez et al., 2023).

In multiagent settings, this formulation generalizes: value alignment may be computed for strategy profiles, aggregations across agents, or preference trade-offs across sets of values (Montes et al., 2020).

2. Historical and Normative Frameworks

Historically, value alignment has roots in Turing’s “fair play” principle, which demands that machines be evaluated by the same criteria as humans—a proto-theory of alignment as mutual normative adaptation, not mere behavioral imitation (Estrada, 2018). From this perspective, value alignment is a bidirectional process:

Constraining machine behavior to meet human norms.
Constraining human evaluative standards to ensure fair appraisal of machines.

Turing’s Imitation Game is reinterpreted as a minimal alignment test: a machine meets the “fair play” threshold if it is judged, under human standards, as indistinguishable (up to $\epsilon$ ) from a human in relevant domains. The fair-play paradigm contrasts sharply with “Moral Turing Test” approaches, which require only behavioral imitation rather than principled parity of standards (Estrada, 2018).

This normative axis motivates anchored value alignment (A-VA): systems are anchored in explicit, intrinsic normative principles (e.g., honesty, autonomy), potentially verified via modal logic or multi-objective optimization, as contrasted with mimetic alignment via imitation learning (Kim et al., 2018, Kim et al., 2019).

3. Formal and Mechanistic Methodologies

3.1. Multi-Objective and Aggregative Frameworks

Values in real systems are aggregated across agents and objectives:

Aggregation over values: $\operatorname{Prf}_V^\alpha(s, s') = \frac{1}{|V|} \sum_{v \in V} \operatorname{Prf}_v^\alpha(s, s')$ .
Aggregation over agents: $\operatorname{Prf}_v^G(s, s') = \frac{1}{|G|} \sum_{\alpha \in G} \operatorname{Prf}_v^\alpha(s, s')$ .

The alignment of a system with respect to values and stakeholders is then determined by choosing appropriate aggregation operators and evaluating the resulting alignment score (Sierra et al., 2021, Barez et al., 2023).

3.2. Verification and Evaluation

Alignment verification is formalized as a finite set of critical queries (e.g., "driver’s tests"). Efficient tests can certify with high confidence that an agent’s policy $\pi'$ is $\epsilon$ -aligned with a ground-truth reward function $R$ , using minimal numbers of policy or trajectory queries. For rational or linearly parameterized agents, two policy queries suffice in the omnipotent tester case; for others, heuristic or active-learning-based critical state queries are used (Brown et al., 2020).

3.3. Alignment Equilibria in Multiagent Systems

Equilibria analogous to Nash and Pareto optimality are extended into the value-alignment context:

Alignment equilibrium: no agent can unilaterally deviate from a strategy profile to improve their alignment score with respect to $v$ .
Pareto-optimal alignment: no other profile exists that improves at least one agent’s alignment without reducing another’s (Montes et al., 2020).

These generalize classical solution concepts to arbitrary value structures and provide a rigorous metric for social-normative synthesis.

3.4. Dynamic and Pluralistic Approaches

Edge alignment rejects scalarization and instead maintains a vector reward $R(x, y) = [r_1,\ldots, r_n]^T$ , seeking Pareto-stationary solutions and supporting plural, contextual, and democratic governance of alignment objectives. This encompasses multiobjective alignment, lexicographic and constrained approaches, pluralistic alignment modes, and explicit governance and risk-sensitive mechanisms (Bao et al., 23 Feb 2026, Zheng et al., 19 Jan 2026).

VISPA exemplifies pluralistic alignment at the architectural level: for any input, a set of maximally relevant values is selected via NLI-based relevance gating, then internal activation steering is applied to synthesize responses corresponding to each value or value subset, with both Overton (presenting spectrum), steerable (targeting a specific value), and distributional (population-matching) response modes (Zheng et al., 19 Jan 2026).

4. Value Expression: Measurement, Personalization, and Empirical Analysis

Recent advances operationalize value expression empirically, notably through large-scale annotation and LLM labeling using Schwartz’s 19-value circumplex, both in social media feeds (Jahanbakhsh et al., 17 Sep 2025, Epstein et al., 11 Nov 2025) and scenario-based datasets.

4.1. LLM Measurement and Personalization

Large annotated datasets enable the training and calibration of LLMs for value expression detection across diverse texts. Crucially, inter-rater reliability studies indicate substantial subjectivity—value expression is partly in the eye of the beholder—necessitating calibration layers that adapt LLM predictions to the value calibration profiles of individual users. Personalized models consistently predict value expressions with higher agreement to raters than inter-human agreement (Epstein et al., 11 Nov 2025).

4.2. Alignment in Downstream Systems

These measurements serve as inputs to application-layer alignment, e.g., re-ranking social media content. Feeds can be value-optimized via user-specified value weight vectors $W_u$ , ranking posts by the dot product $S(p,u) = W_u^\top V_p$ . Multi-dimensional controls allow adjustment along the circumplex, supporting explicit trade-offs and plurality (Jahanbakhsh et al., 17 Sep 2025).

Civic or societal weights can be defined for democratic value alignment, and engagement metrics can be jointly optimized in multi-objective recommender systems.

5. Alignment Dynamics, Trade-offs, and Systemic Risks

Emergent findings from the Value Alignment Tax (VAT) framework reveal that alignment interventions (prompting, fine-tuning, preference optimization) create complex, often structured, co-movements across values in large models. Gain on the target value is often accompanied by measurable off-target shifts, and certain values act as “coordination hubs” subject to concentrated alignment stress. System-level metrics—normalized VAT, value coupling matrices, and centralization indices—quantify these trade-offs and amplify systemic risk: as models become highly entangled in value space, isolated alignment becomes impossible and minor interventions can produce disproportionate effects (Chen et al., 12 Feb 2026).

6. Elicitation, User-Driven and Bottom-Up Alignment, and Contextualization

Expression and alignment cannot be decoupled from elicitation: enabling humans to articulate, reflect, and revise their value frameworks. Interactive-Reflective Dialogue Alignment scaffolds users' ability to articulate subjective values and produces personalized reward models through active learning and hypothesis reflection, succeeding even when aggregated alignment to group consensus is suboptimal due to value pluralism (Blair et al., 2024).

User-driven value alignment frameworks document user strategies to correct AI misalignment in real time, ranging from technical interventions to character-driven steering, and illustrate the sociotechnical complexity of alignment in lived contexts (Fan et al., 2024).

Bottom-up alignment, grounded in ISO/IEEE value ontologies, identifies context-specific values and misalignments from real-world conversational logs rather than propagating abstract principles, improving the empirical grounding and practical utility of alignment frameworks (Motnikar et al., 26 Jun 2025).

7. Open Problems, Taxonomies, and Future Research Directions

Despite the formal and empirical progress, core challenges persist:

The naturalistic fallacy: mimetic alignment without normative anchors propagates human biases and lacks ethical legitimacy (Kim et al., 2018, Kim et al., 2019).
Concept alignment is a prerequisite for value alignment: unaligned conceptual spaces lead to systematic and profound misalignment (Rane et al., 2023).
Macro-alignment (multi-agent, societal value aggregation) remains under-theorized; Arrow-impossibility and instability from stakeholder heterogeneity persist (McKinlay et al., 17 Sep 2025).
Standardized, reproducible alignment benchmarks for diverse contexts, cultures, and time-evolving values are lacking.
Empirical, human-in-the-loop evaluation and dynamic calibration are necessary for sustaining alignment as values and contexts change.

Emerging taxonomies arrange alignment work by (i) normative foundation, (ii) technical implementation, and (iii) calibration procedures, complemented by classifications of ethical theory (consequentialism, deontology, virtue ethics, hybrids), stakeholder roles, and temporal scales of value change (McKinlay et al., 17 Sep 2025). Recommendations emphasize multi-objective, pluralistic, context-aware architectures, robust empirical and audit mechanisms, and participatory governance (Bao et al., 23 Feb 2026, Zheng et al., 19 Jan 2026).

The synthesis of technical, empirical, and normative advances demonstrates that value expression and alignment span a formal representational substrate, rigorous evaluation procedures, empirical and participatory mechanisms, and dynamic, multi-agent system design. Robust alignment requires commitment to explicit value foundations, interactive elicitation, architectural capacity for pluralism, and perpetual openness to systemic dynamics—the necessary conditions for aligning artificial systems with complex, evolving human norms and values.