Constructive Alignment: Governing Preference Dynamics in AI

Constructive Alignment, a 23-page paper by Max Kanwal and Caryn Tran submitted to arXiv on 1 Apr 2026 (arXiv:2607.00001), reframes AI alignment as a control problem over evolving human preference trajectories. The paper, listed as Proceedings of the AAAI-26 Workshop on Machine Ethics and containing one figure, argues that preferences are layered and constructed through interaction rather than fixed targets to be inferred.

What is Constructive Alignment?

Constructive Alignment reframes alignment as governing how AI systems shape the evolution of human values and evaluative states, not merely satisfying static preferences. The authors draw on behavioral economics, psychology, and constructivist social theory to argue that adaptive, persistent, personalized, and socially embedded systems participate in shaping what people attend to, value, and endorse over time.

The paper rejects the common assumption that human preferences are fixed targets. Instead it models preferences as layered state variables that change through interaction with technology. That shift moves the problem from optimizing for a static objective to managing long-term value formation.

How do the authors formalize preferences and interaction?

They formalize the view using a control-theoretic framework in which system actions and interaction design jointly influence both world states and human evaluative states. The opening sentences of this section state the core: preferences are layered state variables that evolve under interaction, and alignment becomes a control problem over those evolving trajectories.

Under this framework, designers and systems are treated as agents that can alter trajectories through action selection and interaction design choices. The model links traditional world-state control with an explicit representation of human evaluative states, making the influence of interaction a first-class object of analysis. The paper emphasizes governance goals: ensure value trajectories remain coherent, reflectively endorsed, epistemically grounded, bounded against manipulation, and empowering under uncertainty.

How does Constructive Alignment change alignment practice?

The paper shifts the focus from controlling only the AI agent's outputs to regulating how systems influence human evaluative states over time. The authors write that "alignment is not primarily about controlling AI behavior, but about regulating how AI systems influence the evolution of human preferences," positioning interaction design, persistence, and personalization as leverage points.

Practically, this implies different evaluation targets. Rather than measuring compliance with a fixed reward or preference model, systems would be evaluated on properties of preference trajectories: coherence, reflective endorsement, epistemic grounding, resistance to manipulation, and empowerment under uncertainty. The paper does not prescribe a single implementation; it provides a conceptual and formal foundation for treating preference formation as part of the control loop.

Why it matters

Persistent, personalized, and socially embedded AI already shapes attention and values. By modelling preferences as dynamic and constructed, Constructive Alignment forces alignment research to confront influence pathways that current static-target approaches ignore. Governing preference dynamics changes which interventions count as alignment failures: downstream shifts in values become a design outcome to regulate, not an externality to accept.

This reframing also reallocates responsibility. Designers, platform operators, and policy actors gain explicit standing in alignment discussions because interaction design choices become formal levers in the control-theoretic model. The paper ties theoretical claims to this practical stake, arguing for governance that preserves reflectivity and epistemic standards.

What to watch

Watch for empirical work that operationalizes the paper's control-theoretic framework and for prototype systems that model human evaluative states as layered variables. Also look for debates or workshop follow-ups at venues around AAAI-26 that test the five governance criteria the authors propose: coherence, reflective endorsement, epistemic grounding, boundedness against manipulation, and empowerment under uncertainty.

Reference: Max Kanwal and Caryn Tran, "Constructive Alignment: Governing Preference Dynamics in Human-AI Interaction," arXiv:2607.00001, submitted 1 Apr 2026, 23 pages, 1 figure; Proceedings of the AAAI-26 Workshop on Machine Ethics.