Explainability for the People: Designing User-Centred, Post-Hoc Explainability through Participatory Provotyping
Authors
Barros Martins de Jesus, João ; Roumeliotis, Timon
Term
4. term
Education
Publication year
2026
Abstract
This thesis addresses the gap between ethical aspirations and practical use of explainability in large language models, where prevailing techniques serve developers and often yield fluent but not necessarily faithful explanations. The authors reconceptualize explainability as an interactional practice that emerges between a user, the interface, a given answer, and the situational context, rather than as an inherent property of the model. To explore this, they created ExplainGround—a low-fidelity, model-agnostic provotype that adds a post-hoc explainability layer with a five-item prompt set and two user interaction modes on top of a chat interface. The tool was evaluated in two participatory design workshops with 12 participants spanning technical and non-technical backgrounds. Data (transcripts, shared board images, and usage logs) were analyzed using reflexive thematic analysis. Six themes emerged: explanations were negotiated in interaction; their value increased with user control; satisfaction depended on the relationship between user expertise and model performance; a commonly technical issue could be reframed as a manageable, user-centered design problem; ExplainGround functioned as a productive research instrument; and the findings argue against pursuing computational transparency in favor of explainability-oriented interfaces. Overall, the thesis suggests that meaningful explainability for everyday users is best advanced through user-centered interaction design and participatory provotyping, rather than attempts to expose a model’s internal workings.
Denne afhandling adresserer kløften mellem etisk ambition og praktisk anvendelse af forklarlighed i store sprogmodeller, hvor eksisterende metoder primært er målrettet udviklere og ofte producerer flydende, men ikke nødvendigvis troværdige forklaringer. Forfatterne behandler forklarlighed som en interaktionel praksis, der opstår mellem bruger, interface, svar og situation, frem for en indre egenskab ved modellen. For at undersøge dette blev ExplainGround udviklet—et lav-fidelitets, model-agnostisk provotypeværktøj, der lægger et post-hoc forklaringslag med et sæt på fem prompttyper og to former for brugerinteraktion oven på en chatgrænseflade. Værktøjet blev afprøvet i to participatoriske designworkshops med 12 deltagere med både tekniske og ikke-tekniske baggrunde. Data (transskriptioner, billeder af fælles boards og brugslogfiler) blev analyseret ved hjælp af refleksiv tematisk analyse. Seks centrale fund fremkom: forklaringer blev forhandlet frem i interaktionen; forklaringernes nytteværdi hang sammen med, hvor meget kontrol brugeren havde over dem; tilfredshed afhang relationelt af brugerens ekspertise og modellens præstation; et ofte teknisk indrammet problem kunne omformuleres til et mere håndterbart, brugercentreret designproblem; ExplainGround fungerede som et effektivt forskningsinstrument; og fokus bør flyttes fra beregningsmæssig transparens til forklaringsorienterede interfaces. Samlet peger afhandlingen på, at meningsfuld forklarlighed for almindelige brugere bedst fremmes gennem brugercentreret interaktionsdesign og deltagende provotyping frem for at stræbe efter fuld indsigt i modellens indre mekanismer.
[This apstract has been generated with the help of AI directly from the project full text]