Do artificially intelligent chatbots work to increase physical activity?

Intervention group participants will receive access to a smartphone app that aims to increase their physical activity using a digital assistant (i.e., chatbot or conversational agent) that applies reinforcement learning (a type of machine learning) to personalise content and the timing it is delivered. The app works in conjunction with physical activity trackers (trail participants will receive a Fitbit). The intervention consists of a stand-alone smartphone app that works in conjunction with a physical activity tracker (i.e., Fitbit). The main feature of the app is the physical activity digital assistant (i.e., chatbot or conversational agent). The name of the app in the app stores is ‘MoveMentor’ and the total duration of the intervention period is 6 months. Engagement with the intervention will be monitored use app usage statistics. The digital assistant interacts with participants in 3 main ways: 1) conversations; 2) nudges and 3) question and answer (Q&A). The conversations provide participants with educational content in relation to physical activity delivered through interactive ‘conversations’ or ‘chats’. The aim of the conversations is to encourage participants to become more physically active and meet the national physical activity guidelines, as well as to support them in overcoming any difficulties in becoming more active. Example topics of conversations are: ‘Developing a lasting activity habit’, ‘What stops you from being active?’, and ‘Make the most of your surroundings’. During the conversations, the digital assistant will ask participants physical activity questions and participants will be provided with personalised advice, depending on the answers they provide. The conversations are designed to last no longer than 5 or 10 minutes. There is not a set frequency/duration of time that participants must spend with the conversations, they are free to complete as many of the conversations as they see fit whenever it suits them. Participants will also receive activity 'nudges' delivered as a smartphone app notifications. The timing and content of the nudges is determined by machine learning algorithms that learn what works for whom under what circumstances through trial and error. The nudges that work will be reinforced, those that don’t will be phased out. More specifically, a ‘contextual bandit’ approach is applied to deliver the right nudges at the right time, this is a type of reinforcement learning. Contextual data will be derived from participants app preferences and settings (derived using an onboarding survey), participant interactions with the app and digital assistant, physical activity data from the activity trackers, nudge ratings (i.e., like or dislike), and weather data (to determine whether indoor or outdoor activity is recommended based on the participants' location at that specific time). An extensive library of nudges has been developed to send to users through push notifications. The library consists of a range of nudge categories. Examples of nudge categories are: daily morning check in, weekly action plan review, activity suggestion , motivational, educational, streaks, milestones and more. Finally, participants can engage with the digital assistant by asking it physical activity-related questions. Answers to questions that have not been pre-programmed (e.g., I have a muscle tear in my calf, how can I still be active?), are being delivered by PaLM2, Google’s machine learning and natural language generation model (it is a large language model comparable to ChatGPT). There is not a set number of Q&A sessions participants must complete. Whenever they have a physical activity-related question, they can just ask the digital assistant as they see fit; this could be a lot of questions or no questions at all. Beyond interactions with the digital assistant and personalisation features, the MoveMentor apps include additional features of which ongoing self-monitoring, adaptive goal setting and action planning are the most important. Every morning, participants receive a daily activity goal. The goal is designed to slowly increase over time, aligned with how well participants are performing, until it reaches a participant’s ‘long-term activity’ goal at which point the goal stops increasing. If participants are consistently not meeting their daily activity goal, then the goal will gradually decrease. Participants are encouraged to engage in Action Planning conversations. The action plan helps participants define what physical activities they will engage in, where they will do them, how many times per week, what days of the week, what time of day, how long activity sessions will last and identify any people they may be active together with.