State Your Intention to Steer Your Attention: An AI Assistant for Intentional Digital Living

TL;DR

The paper tackles distracting behavior on digital devices by introducing the Intent Assistant (INA), an AI-powered system that elicits user intentions, monitors on-screen activity with LLM-based alignment detection, and issues timely, gentle nudges to maintain focus. It advances a formative study, a public IntentionBench dataset, and a three-week in-the-wild evaluation showing reductions in off-task behavior and higher intention alignment compared with baselines. Its contributions include a context-aware detection framework, an intention-clarification dialogue, and a feedback-driven refinement loop, all implemented in a cross-device, privacy-conscious pipeline. The work demonstrates the potential of proactive, autonomy-preserving AI assistants to support mindful digital living while highlighting challenges around notification burden, detection accuracy, privacy, and long-term adoption.

Abstract

When working on digital devices, people often face distractions that can lead to a decline in productivity and efficiency, as well as negative psychological and emotional impacts. To address this challenge, we introduce a novel Artificial Intelligence (AI) assistant that elicits a user's intention, assesses whether ongoing activities are in line with that intention, and provides gentle nudges when deviations occur. The system leverages a large language model to analyze screenshots, application titles, and URLs, issuing notifications when behavior diverges from the stated goal. Its detection accuracy is refined through initial clarification dialogues and continuous user feedback. In a three-week, within-subjects field deployment with 22 participants, we compared our assistant to both a rule-based intent reminder system and a passive baseline that only logged activity. Results indicate that our AI assistant effectively supports users in maintaining focus and aligning their digital behavior with their intentions. Our source code is publicly available at https://intentassistant.github.io

Figures (11)

• Figure 1: Overview of INA. (1) The process begins with a user's initial, often abstract, intention. (2) The system engages in an LLM Q&A interaction to (3) establish a clarified user intention. (4) An LLM detector then analyzes on-screen activity, using a distraction score to classify the user's state as on-task or off-task. (5) Based on this state, the system delivers a gentle nudge or positive reinforcement. (6) The user can provide feedback on this intervention, which (7) the LLM uses to refine its model, improving detection accuracy for future interactions.
• Figure 2: (1-2) The user inputs an intention and presses 'Set' to initiate a session. (3-5) A chat interface is used for a brief Q&A clarification process, where the user responds to the LLM's questions. (6) The user can also press 'Start' to skip this clarification and begin the session immediately. (7) During the session, INA displays the user's current state with colored notifications: green for on-task and red for off-task. (8-9) The user can hover over a notification to provide feedback, marking it as correct or incorrect and optionally adding a reason. (10) The 'Stop' button can be pressed at any time to end the session.
• Figure 3: An illustration of the IntentionBench construction process. We first collect focused sessions, where each session corresponds to executing a distinct task instruction (e.g., "Study HCI"). Then, two focused sessions are combined to synthesize a mixed session by concatenating and reordering their segments. The user intention is inherited from one session (on-task), while the other constitutes off-task segments, mimicking natural transition points of user workflows.
• Figure 4: Overall procedure of our three-week, within-subjects field study. Participants attended a 30-minute online orientation session before the study. Participants used one application type on their computers each week. The assignment of each application type—INA, logging only, or simple reminder—was based on a random order. At the end of each week, participants completed a post-survey and a 10-20 minute semi-structured video interview.
• Figure 5: (a) Mean LLM-estimated off-task ratio. (b) Mean intention alignment rating based on end-of-session self-reports. Error bars denote 95% confidence intervals across participants (user-averaged). ***, **, and *, indicate significance of p < 0.001, p < 0.01, p < 0.05, respectively. Arrows indicate whether higher ($\uparrow$) or lower ($\downarrow$) values are more favorable.