Researchers at Stanford University have developed an innovative health coaching app, named Bloom, featuring a large language model (LLM) known as Beebo. The study, conducted with 54 subjects over four weeks, revealed that although users of the LLM-enhanced version of the app did not significantly increase their physical activity compared to those using a standard version, their mindsets about fitness underwent notable changes.
Matthew Jörke, a key researcher in the project, explained that participants using Beebo began to view physical activity as attainable and beneficial. “People start recognizing that physical activity is something they can do, that it’s good for them, and it makes them feel better,” he stated. This shift in perception aligns with a growing trend of individuals seeking guidance from LLMs in various life areas, including health.
To create Beebo’s coaching style, the researchers, including senior author James Landay, the Denning Director of the Stanford Institute for Human-Centered AI, employed a human-computer interaction approach. Initially, they interviewed 12 health experts and 10 potential users to uncover essential traits for an effective LLM coach. Key principles emerged: a facilitative approach to encourage users to take ownership of their health, personalized advice suited to individual circumstances, and a nonjudgmental, supportive tone.
Additionally, Beebo was designed to integrate with Stanford Active Choices, a proven counseling program aimed at enhancing physical activity among various demographics. The program’s initial step involves an in-depth interview to extract relevant information about the client’s goals, past experiences, and barriers to activity. Bloom imitates this onboarding conversation through a technique called motivational interviewing, which aims to help users identify their intrinsic motivations.
However, the team faced challenges when they attempted to simply prompt a standard LLM for the initial interview. Jörke noted that the model struggled to remain on topic and often dispensed unsolicited advice. Instead of relying on a large dataset of onboarding conversations to refine the LLM, the researchers experimented with different prompting strategies to guide the model in maintaining focus and adhering to motivational interviewing principles.
The final design incorporated two prompt chains. The first, a dialog state prompt chain, ensures complete coverage of the topics associated with Stanford Active Choices before proceeding. The second consists of a two-step process where one AI agent selects an appropriate conversational strategy, while a second crafts the response accordingly. Jörke emphasized the necessity of this additional structure, stating, “The LLM needs that extra scaffolding to use motivational interviewing the way we want it to.” Safety filters were also integrated to prevent harmful content, such as negative feedback or inappropriate comments regarding body image.
In the comparative study, half of the participants engaged with the standard Bloom app, which offered common health app functionalities, including goal setting and activity reminders. In contrast, the treatment group benefited from Beebo’s extensive features, including personalized workout plans, real-time reminders, and motivational check-ins. Jörke remarked, “We specifically wanted to study what a nonprescriptive LLM would add on top of a high-quality health app. And remarkably, it wasn’t a whole lot more exercise. It was a change in mindset.”
Users of Beebo reported increased satisfaction with their activity levels and a strengthened belief in the sufficiency and benefits of their efforts. Some even noted a newfound appreciation for everyday activities, like gardening, as valid forms of exercise. Jörke explained that Beebo gradually helped users recognize their capabilities and the ease of maintaining an active lifestyle.
Landay highlighted the effectiveness of certain baseline app features, like the ambient display, which have proven impactful in prior studies but are rarely found in commercial health apps. “By including them in the baseline, we set the comparison up to be quite hard for the LLM version of Bloom to beat,” he noted, emphasizing the promising nature of the study’s findings.
With potential applications in areas beyond fitness, such as sleep and nutrition, Jörke sees opportunities for expanding Bloom’s framework. However, he acknowledges the need to create expert-informed dialog state prompt chains tailored to various contexts. “We need to develop a simpler way to prompt or fine-tune a model so that it does a better job of balancing the need to strategically gather relevant information with the urge to deliver advice and recommendations,” he stated.
Concerns regarding the anthropomorphism of AI agents also surfaced, particularly the risks of users developing inappropriate attachments. Nonetheless, Jörke observed that users did not seek medical advice or companionship from Beebo; they interacted with it as a distinct coaching entity. “The metaphors that we use for these chatbots and how we communicate their capabilities to people might actually make their interactions better overall,” he concluded.
While Bloom remains a research prototype, the team plans to make it publicly available soon, responding to growing interest. Emma Brunskill, a researcher involved in the project, remarked, “I think this project is tapping into a deep need. Everyone knows that exercise is important, and people want to be more active, but they are also balancing an enormous number of demands on their time. Bloom tries to address this in a way that centers human agency.”
See also
AI Study Reveals Generated Faces Indistinguishable from Real Photos, Erodes Trust in Visual Media
Gen AI Revolutionizes Market Research, Transforming $140B Industry Dynamics
Researchers Unlock Light-Based AI Operations for Significant Energy Efficiency Gains
Tempus AI Reports $334M Earnings Surge, Unveils Lymphoma Research Partnership
Iaroslav Argunov Reveals Big Data Methodology Boosting Construction Profits by Billions
