Earning Autonomy: A Confidence-Based Approach to Preserving Human Control in Smart-Homes
Authors
Holter Brubakken, Sondre ; Damgaard, Ulrikke
Term
4. term
Education
Publication year
2026
Abstract
This thesis examines how smart-home systems can transition from user-initiated commands to more autonomous and proactive behavior while preserving human control. Building on Shneiderman’s Human-Centered AI (HCAI) framework and van Berkel et al.’s continuum from intermittent to proactive interaction, the work introduces a smart-home architecture centered on a confidence-based mediator. A Large Language Model (LLM) grounds natural language into device actions or higher-level intents and interprets user overrides as either temporary exceptions or genuine preference changes. The mediator combines this interpretation with sensor readings, device states, a dynamically updated confidence score, and contextual validation to decide whether actions should be merely presented (Exploration), suggested for confirmation (Preference), or executed and initiated autonomously (Autonomous). The system supports cold-start operation and gradually builds confidence through repeated interaction while reducing effective confidence when the current context contradicts previously learned patterns, such as rain conflicting with a window-opening preference. The architecture was first technically verified through scripted tests, which showed that three autonomy configurations (Conservative, Standard, Aggressive) behaved in line with the formal specification for confidence accumulation, reversibility after overrides, and contextual suppression. It was then evaluated in a between-subjects user study (n=101) using a browser-based simulation, where the three configurations were compared via Likert-scale ratings analyzed with non-parametric tests and inductive thematic analysis of open-ended responses. The Standard configuration received the highest scores on all dimensions and was rated significantly safer than both Conservative and Aggressive and more trustworthy than Conservative, while differences in perceived Autonomy were not significant. The two extreme configurations failed in different ways: Conservative was seen as adapting too slowly and burdening users with excessive confirmations, whereas Aggressive acted prematurely on insufficient evidence, undermining perceived competence and control respectively. A key qualitative finding was that users did not view autonomy as a uniform, system-wide property but evaluated it at the level of individual actions: low-risk actions such as lighting adjustments were accepted earlier, while higher-risk actions involving windows, curtains, or locks required substantially more evidence before autonomous behavior was seen as legitimate. Overall, the findings suggest that the main challenge is not enabling autonomous and proactive smart homes per se, but determining when such behavior becomes legitimate from the user’s perspective, pointing toward risk-sensitive, action-specific autonomy thresholds rather than a single global policy in future systems.
Denne afhandling undersøger, hvordan smarthome-systemer kan bevæge sig fra brugerstyrede kommandoer til mere autonom og proaktiv adfærd uden at mindske menneskelig kontrol. Med udgangspunkt i Shneidermans Human-Centered AI (HCAI)-ramme og van Berkels kontinuum fra intermittent til proaktiv interaktion udvikles en smart-home arkitektur centreret om en såkaldt confidence-baseret mediator. Et Large Language Model (LLM) oversætter naturligt sprog til konkrete handlinger og fortolker brugerens forklaringer ved overrides som enten midlertidige undtagelser eller egentlige præferenceændringer. Mediatoren kombinerer denne fortolkning med sensordata, en løbende opdateret selvtillids-score og kontekstuel validering for at afgøre, om handlinger skal vises som muligheder (Exploration), foreslås til bekræftelse (Preference) eller udføres og initieres autonomt (Autonomous). Systemet kan bruges fra et koldstart-scenarie og lader sin selvtillid vokse gennem gentagne interaktioner, men kan sænke effektiv selvtillid, hvis den aktuelle kontekst modsiger tidligere lærte mønstre, eksempelvis ved regn og vinduesåbning. Arkitekturen blev først teknisk verificeret gennem scriptede tests, som viste, at alle tre autonomikonfigurationer (Konservativ, Standard og Aggressiv) fulgte den formelle specifikation for bl.a. tillidsopbygning, tilbageførsel efter overrides og kontekstuel undertrykkelse. Dernæst blev den evalueret via et between-subjects brugerstudie (n=101) i en browserbaseret simulation, hvor de tre konfigurationer blev sammenlignet ved hjælp af Likert-skalaer analyseret med ikke-parametriske test og tematisk analyse af åbne svar. Standard-konfigurationen opnåede de højeste bedømmelser på alle dimensioner og blev vurderet signifikant mere sikker end både Konservativ og Aggressiv samt mere troværdig end Konservativ, mens forskellen i oplevet Autonomi ikke var signifikant. De to ekstreme konfigurationer udviste forskellige svagheder: den Konservative blev oplevet som for langsom og krævende for mange bekræftelser, mens den Aggressive handlede for tidligt på utilstrækkelig evidens, hvilket underminerede henholdsvis oplevet kompetence og kontrol. Et centralt kvalitativt fund var, at brugerne ikke vurderede autonomi som en samlet systemegenskab, men pr. handling: lavrisiko-aktiviteter som lysjustering blev accepteret tidligt, mens højere risikohandlinger ved fx vinduer, gardiner og låse krævede væsentligt mere evidens, før autonom adfærd blev opfattet som legitim. Samlet peger resultaterne på, at udfordringen ikke blot er at skabe autonome og proaktive smarthomes, men at afgøre hvornår denne autonomi opleves som berettiget af brugeren, og at fremtidige systemer bør arbejde med risikofølsomme, handlingsspecifikke autonomitærskler frem for én global politik.
[This abstract has been generated with the help of AI directly from the project full text]