Collaborative Diary Agents Require User Control and Grounding to Support Health Reflection

Conversational systems driven by large language models require user-directed revision, contextual grounding, and clear boundaries of ownership rather than passive automation when assisting health-related reflection, according to an exploratory investigation published in the Journal of Ambient Intelligence and Humanized Computing. Researchers Jayalakshmi Baskar, Vera C. Kaelin, and Helena Lindgren examined how individuals interact with artificial intelligence when constructing personal health diaries. The study established that while modern generative models can readily synthesize fluent narratives about daily activities, users frequently experience ungrounded or autonomous outputs as alienating, generic, and unrepresentative of their lived experiences. By applying an activity-theoretical lens to the co-construction of reflective texts, the investigation demonstrated that the true utility of an artificial intelligence diary lies not in generating finished records independently, but in serving as a flexible partner that invites the person to question interpretations, correct misunderstandings, and negotiate personal meaning through sustained, iterative dialogue.

The conceptual framework underpinning the research is rooted in Cultural-Historical Activity Theory, complemented by Walton’s argumentation theory. Within this theoretical structure, the collaborative task of producing a purposeful diary entry constitutes a central, mediated activity governed by actors, tools, divisions of labor, and rules. Instead of framing generative artificial intelligence as an autonomous authority that outputs self-contained summaries, the authors conceptualized the diary entry as an evolving artifact continuously shaped through mutual communication. Everyday planning and retrospective reflection represent purposeful, socially situated activities influenced by personal motives and environmental context. Shifting an individual’s focus from writing text to addressing obstacles in model behavior or recalling past events mirrors structural shifts in ongoing human activity. Grounding diary narratives within explicit argumentation schemes—specifically argument from value, position to know, sufficient condition, and expert opinion—provided an internal architectural structure to tie generated assertions directly to reported personal experiences rather than speculative machine assumptions.

To formulate concrete requirements for such collaborative systems, the research team instituted a two-phase participatory design methodology centered on the domain of occupational therapy. The researchers recruited six female occupational therapy experts residing in three European countries, all of whom were non-native English speakers possessing specialized knowledge in rehabilitation, daily activity reflection, and health technology. The study deliberately engaged domain specialists to obtain rigorous professional insights into reflection on daily habits, acknowledging that the small, focused sample was formative and not intended to represent the wider population of occupational therapists or general end users. Phase 1 comprised individual semi-structured co-design sessions lasting approximately sixty minutes each, conducted remotely over Microsoft Teams with two or three researchers present. These collaborative sessions explored interface mock-ups, potential purposes, interaction workflows, and organizational structures for distinguishing between retrospective reflection on past events and prospective planning for upcoming routines.

Qualitative analysis of the Phase 1 co-design sessions using inductive content analysis identified nine primary thematic categories governing interaction: purpose-driven and context-dependent use; structured yet flexible dialogue; personalization and adaptive interaction; reflection as recognition, learning, and pattern awareness; the relationship between reflection and planning; balancing motivation, goals, and psychological safety; usability and accessibility; privacy, ownership, and controlled sharing; and engagement and long-term use. Across these dimensions, collaboration emerged not as an isolated functional requirement, but as an overarching interactional condition. Experts emphasized that the agent must act as an external memory capable of supporting self-awareness while guiding users without imposing rigid reporting constraints. Crucially, the occupational therapy specialists noted that reflection fulfills three interrelated roles: acknowledging completed accomplishments to counter distorted perceptions of inactivity, learning how stress and routines influence outcomes, and detecting longitudinal behavioural patterns across daily life.

Balancing motivational goal-setting against emotional safety surfaced as an essential ethical consideration during the initial design phase. The participating clinicians highlighted that aggressive goal-setting risks creating undue psychological pressure, potentially inducing guilt or reinforcing perceived functional decline when plans remain unfulfilled. Consequently, the experts advised that diary systems must avoid negative framing, introduce small and achievable steps, treat unfulfilled intentions as normal opportunities for learning rather than personal failures, and ground future planning in exploratory possibilities rather than rigid productivity expectations. Accessibility requirements identified by the panel included plain language, concise summaries, large readable text, voice input support, layered information views, and multimodal presentation formats. Furthermore, the experts maintained that diary records represent deeply private reflections; data sharing with healthcare professionals or researchers should remain strictly opt-in and under the explicit, granular control of the person rather than enabled by default.

Following Phase 1, the researchers translated these design considerations into a functional software module integrated into BlumeAI, a specialized mobile health platform designed to track lifestyle activities, social participation, recovery, and physical exercise. The technical pipeline utilized the OpenAI Chat Completions application programming interface powered by the gpt-4o-mini language model, operating at a controlled decoding temperature of 0.1 to limit random generation. When an individual initiated a diary session, the mobile application prompted them to choose between retrospective reflection on the preceding week or prospective planning for the coming week, supported by selectable example purposes or custom text input. The client frontend then transmitted this selected objective, the temporal direction, up to three logged activity descriptions from the BlumeAI activity module, and up to six background-question answers concerning stress, tiredness, wellbeing, lifestyle habits, and physical barriers to the backend server to ground the generation process.

The system’s interaction architecture deliberately enforced a strict spatial and operational separation between the evolving diary narrative and the collaborative communication channel. In the mobile user interface, the generated diary draft was displayed in a distinct diary-post space, while a separate dialogue space provided a free-text messaging field. The diary-post space was deliberately not directly editable by the user; instead, all additions, removals, and textual corrections were mediated through conversational exchanges with the diary agent. During iterative revisions, the backend submitted the existing text draft alongside the user’s latest dialogue prompt to the language model, instructing it to integrate requested revisions while preserving relevant existing context. Although this architectural separation successfully supported progressive narrative drafting, the Phase 2 research prototype did not incorporate persistent database storage for finalized entries, nor did it include automatic algorithms for generating context-sensitive clarification questions when input details were ambiguous.

In Phase 2, the six occupational therapy experts evaluated the implemented prototype by interacting with the BlumeAI application over approximately one week. Five experts tested the diary agent independently, subsequently providing evaluative feedback through follow-up interviews or detailed written responses, while the sixth participant explored the mobile prototype under direct researcher guidance during a virtual session. Subsequent inductive analysis of their experiences revealed four overarching categories: reflection, ownership, and personalization; explainability and collaborative dialogue; reflection versus planning; and adaptive coaching and trust. Overall, the domain experts confirmed that interacting with the prototype demonstrated clear potential for assisting individuals in revisiting life events and structuring routines, but their hands-on experiences simultaneously revealed notable interactional tensions surrounding automated assumptions, system repetition, and the psychological boundaries of authorship.

Personalization and contextual grounding emerged as substantial operational challenges during prototype testing, with five experts reporting instances where model outputs felt superficial, repetitive, or disconnected from their real lives. Participants observed that the system frequently fell back on stereotypical wellness recommendations, repeatedly suggesting activities like mindfulness or physical exercise even when other personalized activities were more relevant to the logged context. Furthermore, because narrative generation was initially performed entirely by the language model, participants expressed that the resulting text often felt detached from their personal identity, prompting one expert to observe that when an automated agent continually writes the entries, the artifact ceases to feel like one’s personal diary. In response, experts adopted active compensatory strategies, such as explicitly instructing the agent to retain their exact phrasing, and strongly recommended that future iterations allow users to contribute original text directly to safeguard narrative ownership.

The evaluations underscored an urgent need for algorithmic explainability and transparent reasoning during human-agent collaboration. When the agent generated an unsubstantiated claim asserting that a participant had engaged in limited physical activity, the expert immediately challenged the assertion, demanding to know how the system had reached that conclusion and insisting that the underlying logged data points be identified. Two experts also experienced the agent as insufficiently responsive or overly focused on prematurely concluding the task, noting that the conversational interface repeatedly prompted them to finalize and save the text rather than sustaining an open-ended, reflective conversation. These experiences demonstrated that collaborative agents must permit users to challenge model inferences, correct misunderstandings through dialogue, and explore motivations, rather than prioritizing rapid narrative closure.

The study authors documented several methodological limitations, noting that their sample comprised six European occupational therapy professionals and did not include older adults, who represent a primary intended demographic for future diary deployment. Technically, the exploratory prototype evaluated only a specific set of prompts, a single language model, and a constrained temperature setting, without retaining fine-grained request telemetry, token expenditure records, or quantitative interaction metrics. Persistent storage and context-sensitive clarifications remained unaddressed in the evaluated build. Within the cultural-historical activity framework, the researchers concluded that true human-agent collaboration requires treating the user as an active partner endowed with structural control over the narrative, ensuring that future health-supportive artificial intelligence systems remain grounded, explainable, and fundamentally adaptable to human agency.

Subject of Research: Participatory design and activity-theoretical evaluation of an LLM-based collaborative health diary agent

Article Title: Understanding human-diary agent collaboration through an activity-theoretical framework

Article References: Baskar, J., Kaelin, V. C., & Lindgren, H. (2026). Understanding human-diary agent collaboration through an activity-theoretical framework. Journal of Ambient Intelligence and Humanized Computing. https://doi.org/10.1007/s12652-026-05139-7

Image Credits: AI Generated

DOI: 10.1007/s12652-026-05139-7

Keywords: conversational AI, collaborative diary agent, large language models, activity theory, participatory design, occupational therapy, BlumeAI, reflective dialogue, activity planning, explainable AI, narrative co-construction, contextual grounding, gpt-4o-mini, health informatics, user ownership, argumentation theory