Bloom: Designing for LLM-Augmented Behavior Change Interactions

This paper introduces Bloom, an LLM-integrated health coaching application that, while not immediately outperforming a non-LLM control in short-term physical activity levels, significantly enhances psychological outcomes like enjoyment and self-compassion, suggesting LLMs are particularly effective at fostering the mindset shifts necessary for long-term behavior change.

The Gist

Imagine you want to get fitter, but you've tried apps before that felt like a strict drill sergeant or a boring spreadsheet. They tell you to "do 30 minutes of cardio" and then just stare at you if you miss a day.

The paper you're asking about introduces Bloom, a new kind of health app that tries a different approach. Instead of just tracking your steps, it pairs a standard fitness tracker with a smart, empathetic AI coach named Beebo.

Here is the story of what they did, what they found, and what it means for you, explained simply.

🌱 The Big Idea: The Garden vs. The Spreadsheet

Think of traditional fitness apps as a spreadsheet. They give you numbers: "You walked 5,000 steps. Goal: 10,000. You failed." It's cold, rigid, and often makes you feel guilty.

Bloom is more like a digital garden.

• The Coach (Beebo): Beebo is an AI chatbot that talks to you like a supportive friend, not a boss. It uses a technique called "motivational interviewing," which means it asks you what you want to achieve and helps you figure out how to get there, rather than just ordering you around.
• The Garden: As you complete your workouts, a flower in your phone's "garden" grows. If you miss a day, the flower doesn't die; it just waits for you to come back. It's a gentle, visual reminder that feels good to look at, rather than a red "X" on a calendar.
• The Mix: The app combines this friendly chat with the "boring but useful" stuff: tracking your steps, making a weekly plan, and sending you reminders.

🧪 The Experiment: The "Magic" vs. The "Standard"

The researchers wanted to know: Does having this chatty, smart AI coach actually help people get healthier compared to just having a standard app?

They set up a four-week experiment with 54 people:

1. The "Bloom" Group: Got the full app with Beebo (the AI coach), the garden, and the chat.
2. The "Control" Group: Got a version of the app without the AI. They had to type their own goals, saw the same garden, and got generic text reminders (like "Time to walk!").

📊 What Happened? (The Results)

The results were surprising and nuanced.

1. The "Gym Rat" Effect (Physical Activity)
Both groups got much more active! In fact, the number of people meeting the recommended weekly exercise goals doubled in both groups.

• The Twist: The group without the AI coach actually got slightly more active in the short term. They jumped into action quickly. The AI group didn't get more steps than the non-AI group.

2. The "Mindset" Shift (The Real Win)
This is where the AI shined. While the non-AI group just got more steps, the Bloom group changed how they thought about exercise.

• Enjoyment: They started to actually like moving their bodies more.
• Self-Compassion: If they missed a workout, they didn't beat themselves up. They thought, "Oh well, I'll try again tomorrow," instead of "I'm a failure."
• Belief: They truly believed that exercise was good for them, not just a chore they had to do.
• Flexibility: They were better at adjusting their plans when life got crazy (e.g., "It's raining, let's walk inside the mall instead of outside").

3. The "Friend" Factor
People in the Bloom group spent five times more time in the app than the control group. They felt like Beebo was a partner. One user said, "It felt like talking to a person." Even though they knew it was a robot, the way Beebo spoke made them feel supported and less alone.

💡 The Takeaway: Why This Matters

The researchers realized something important: AI isn't necessarily better at making you run faster today, but it might be better at making you want to run forever.

Think of it like this:

• The Standard App is like a personal trainer who pushes you hard for one session. You get a great workout, but you might dread going back next week.
• Bloom (The AI) is like a wise friend who helps you build a habit. They don't force you to run a marathon tomorrow, but they help you believe you can run, and they help you find joy in the process.

🚧 The Hiccups

It wasn't perfect. Sometimes Beebo was a bit too chatty or repeated the same phrases. Sometimes it got confused about technical details (like why your watch wasn't syncing). But overall, the "human-like" connection was powerful enough to change people's minds about exercise.

🌟 The Bottom Line

If you've ever quit a fitness app because it made you feel guilty or bored, Bloom suggests a new path. By combining a smart, empathetic AI with a gentle, visual garden, we might be able to help people build lasting habits not by forcing them to move, but by helping them fall in love with the idea of moving.

The AI didn't just count steps; it helped people change their story about who they are and what they are capable of. And that might be the most powerful tool for change we have.

Technical Summary

Here is a detailed technical summary of the paper "Bloom: Designing for LLM-Augmented Behavior Change Interactions."

1. Problem Statement

Despite the proven efficacy of Large Language Models (LLMs) in improving conversational flexibility for health coaching, existing research has narrowly focused on text-only interactions (chatbots, summaries, or nudges). This approach overlooks decades of Human-Computer Interaction (HCI) research demonstrating that effective behavior change relies on multimodal interactions (e.g., ambient displays, data visualizations, just-in-time nudges) and established design patterns (goal setting, self-tracking).

Furthermore, there is a lack of rigorous, real-world field studies comparing LLM-augmented systems against established non-LLM baselines. Most prior work assumes LLM efficacy without isolating the specific value added by LLMs versus traditional rule-based or UI-driven interventions. The core challenge is determining how to integrate LLMs not as replacements for existing behavior change strategies, but as augmentations that leverage qualitative context to personalize and enhance these established interactions.

2. Methodology

System Design: Bloom

The authors developed Bloom, an iOS application designed to promote physical activity (PA) by integrating an LLM health coach ("Beebo") with six evidence-based behavior change interactions derived from the Stanford Active Choices program and HCI literature:

1. Goal Setting & Planning: Collaborative weekly planning using Motivational Interviewing (MI) strategies.
2. Activity Tracking: Automatic logging via Apple HealthKit.
3. Data Visualizations: Charts annotated with LLM-generated summaries.
4. Ambient Display: A "garden" metaphor on the lockscreen that grows as activities are completed.
5. Push Notifications: Personalized, LLM-generated reminders and reflections.
6. LLM Chat: An onboarding, check-in, and "at-will" chat interface.

Technical Architecture:

• Frontend: iOS app built with React Native/Swift, integrating Apple HealthKit.
• Backend: Python/FastAPI on Google Cloud Run.
• LLM: GPT-4o (via OpenAI API) for inference.
• Context Management: The system maintains a summary-based memory of past conversations to provide qualitative context (goals, barriers, values) for personalization.
• Safety: A custom safety layer using a taxonomy of harms (Bodily Harm, Body Image, Mental Health, Negative Mindsets, Inaccurate Advice) with few-shot classifiers and revision prompts to filter and rewrite harmful outputs.

Study Design

• Type: Four-week, between-subjects, randomized field study.
• Participants: $N=54$ (26 Treatment, 28 Control). Participants were iPhone/Apple Watch owners with low-to-moderate baseline PA levels.
• Conditions:
  • Treatment (Bloom): Full system with LLM chat, personalized summaries, and LLM-generated notifications/plans.
  • Control (No-LLM): Removed the LLM agent. Used UI-based plan creation, template-based notifications, and no chat. This served as a "strong pre-LLM status quo" baseline.
• Data Collection:
  • Objective: Wearable data (steps, energy, distance, exercise time) via HealthKit.
  • Subjective: Pre/post surveys (IPAQ, mindset scales, usability), weekly/daily surveys, and semi-structured interviews.
  • System Logs: Chat logs, plan edits, and app usage duration.
• Analysis: Mixed-methods approach including reflexive thematic analysis for interviews, descriptive statistics for surveys, and linear mixed-effects modeling for wearable data.

3. Key Contributions

1. The Bloom System: A novel architecture integrating an LLM coaching agent with established behavior change UI elements. It demonstrates how LLMs can leverage qualitative context (from chat) to personalize quantitative interactions (plans, notifications, visuals).
2. Safety Benchmark & Evaluation: The authors conducted a redteaming evaluation with domain experts to create a taxonomy of harms specific to PA coaching. They released a safety benchmark dataset (600 examples) and demonstrated that their safety filters achieved high recall (>0.96) in detecting harmful content, with a revision prompt reducing residual harmfulness to <0.2%.
3. Empirical Field Study Findings: A rigorous comparison of an LLM-augmented system against a non-LLM control, providing evidence that LLMs shift psychological outcomes (mindsets) more effectively than short-term behavioral metrics.
4. Design Implications: Concrete guidelines for designing LLM-augmented behavior change systems, emphasizing the role of LLMs in fostering agency through qualitative context and the trade-offs of relational cues.

4. Results

Psychological & Motivational Outcomes (Qualitative & Survey)

• Mindset Shifts: The Treatment group showed significantly greater improvements in PA adequacy mindset (belief that activity is beneficial) and self-compassion compared to the Control group.
• Enjoyment & Agency: Treatment participants reported higher enjoyment of exercise and a stronger sense of autonomy. They viewed Beebo as an "accountability partner" rather than a rigid tool.
• Qualitative Themes:
  • Accountability: Participants in both groups relied on plans and ambient displays for accountability, but the Treatment group felt Beebo reinforced this through flexible, empathetic support.
  • Flexible Planning: The LLM's ability to reschedule or modify plans based on natural language requests was highly valued.
  • Reframing: Beebo helped users reframe everyday activities (e.g., gardening) as valid exercise.

Behavioral Outcomes (Wearable Data)

• Overall Increase: Both groups significantly increased their PA levels compared to baseline (doubling the proportion of participants meeting the 150 min/week guideline).
• Treatment vs. Control:
  • Short-term PA: There was no statistically significant difference in short-term physical activity levels between the LLM and Control groups.
  • Trends: Descriptive patterns suggested the Control group had a larger initial spike in activity followed by a steeper decline, whereas the Treatment group showed smaller initial gains but more stable persistence over the four weeks.
• Plan Quality: The Treatment group created more balanced plans (covering aerobic, strength, and flexibility) and made more edits to their plans, indicating higher engagement.

Engagement & Usability

• Usage Time: Treatment participants spent 5.6 times more time in the app than Control participants. This increase was observed across all screens, not just the chat, suggesting the LLM enhanced the overall experience.
• Usability: Usability scores (SUS) declined over time for both groups, but the decline was steeper and more variable in the Treatment group, likely due to the complexity of integrating non-deterministic LLMs with rigid UI flows.

5. Significance and Implications

• LLMs as Mindset Shapers: The study challenges the assumption that LLMs must immediately drive higher activity counts. Instead, it suggests LLMs are uniquely effective at shifting mindsets (enjoyment, self-efficacy, belief in benefits) which are precursors to long-term behavior maintenance.
• Qualitative Context as a Lever: The primary value of the LLM was not just generating text, but using qualitative context (goals, barriers, life circumstances) to personalize quantitative interventions. This enabled non-prescriptive, agency-promoting interactions that rule-based systems could not achieve.
• Relational Cues: The "human-like" nature of Beebo increased engagement and accountability but introduced risks of over-reliance or attachment. The authors argue for a balanced approach where relational cues are paired with other engagement strategies (like ambient displays) to mitigate risks.
• Design Trade-offs: Integrating LLMs introduces challenges in consistency (hallucinations, repetitive responses) and usability (managing user expectations). Future systems need robust tool-use mechanisms and clearer communication of agent capabilities.

In conclusion, Bloom demonstrates that integrating LLMs into established behavior change frameworks can create more supportive, personalized, and engaging health applications. While short-term activity gains may not differ significantly from traditional tools, the LLM's ability to foster positive psychological shifts and sustained engagement offers a promising pathway for long-term health behavior change.