Co-Constructing Alignment: A Participatory Approach to Situate AI Values

TL;DR

This paper reframes AI value alignment as a situated, interactional practice co-constructed by users and AI during use, rather than a static property learned predeployment. It develops a participatory workshop methodology—combining misalignment diaries, generative design prompts, and artefact-driven reflexive analysis—to surface local misalignments and envision actionable co-construction roles and interfaces for large language models used as research assistants. The findings reveal misalignment as concrete task and interaction breakdowns, highlight user-led strategies for adjustment and restraint, and outline design opportunities that surface model positioning, feedback, and collective responsibility. Collectively, the work argues for runtime, distributed alignment governance that respects user agency and situates alignment in everyday practice, with implications for interface design and back-end mechanisms that can translate user input into real-time behavioural adjustments.

Abstract

As AI systems become embedded in everyday practice, value misalignment has emerged as a pressing concern. Yet, dominant alignment approaches remain model centric, treating users as passive recipients of prespecified values rather than as epistemic agents who encounter and respond to misalignment during interactions. Drawing on situated perspectives, we frame alignment as an interactional practice co-constructed during human AI interaction. We investigate how users understand and wish to contribute to this process through a participatory workshop that combines misalignment diaries with generative design activities. We surface how misalignments materialise in practice and how users envision acting on them, grounded in the context of researchers using Large Language Models as research assistants. Our findings show that misalignments are experienced less as abstract ethical violations than as unexpected responses, and task or social breakdowns. Participants articulated roles ranging from adjusting and interpreting model behaviour to deliberate non-engagement as an alignment strategy. We conclude with implications for designing systems that support alignment as an ongoing, situated, and shared practice.

Figures (7)

• Figure 1: Overview of the three workshop phases. Phase 0 uses a diary to sensitise participants to different forms of misalignment in AI interactions. Phase 1 foregrounds the negotiation of situated values shaping misaligned AI interactions. Phase 2 moves from problem articulation to solution-making through discussion of actions and interface interventions toward a self-defined alignment goal.
• Figure 2: Example of a misaligned interaction traced from initial diary entry to concern and values at stake across Phases 0–2. The example from P7 documents an initial prompt, a misaligned response, an intuitive intervention, and the final outcome when using ChatGPT to assess workshop safety, followed by a reflection on why the misalignment mattered and values were implicated.
• Figure 3: From a shared alignment goal to an action metaphor in Steps 4–6. P7 and Group 3 define an alignment goal emphasising nuance and reflexivity, which P7 further translates into user actions through a goat-herding metaphor.
• Figure 4: P7-envisioned interface for Step 7, supporting reflexive alignment through visible model positionality.
• Figure A1: Participant demographic counts for our participatory workshop with researchers ($N=12$).