Apps such as Instagram and TikTok keep us glued to our smartphones by design. An artificial intelligence system that reminds us to take breaks can help slash that screen time.
A random notification to stop doomscrolling won鈥檛 always tear someone away from their phone. But machine learning can personalise that intervention so it arrives at the moment when it is most likely to work, says at the Massachusetts Institute of Technology.
Xu and his colleagues recruited university students to help train and test several versions of an AI reminder system to reduce excessive app use. For the first two weeks, the study monitored students鈥� usage of entertainment, social media and shopping apps to see how often and for how long each person used each app. The system also periodically asked people if they agreed or disagreed with statements such as 鈥淚 think I shouldn鈥檛 use TikTok now鈥�.
Advertisement
Each person鈥檚 self-reported indications of possible smartphone overuse helped calculate a simple probability for when to provide reminders in a basic intervention system. For example, a person who reported overusing their apps during 60 per cent of the initial system check-ins might later get a pop-up reminder 60 per cent of the time.
This self-reported data also helped train multiple AI models to potentially make more accurate predictions about the most effective time for reminders. One personalised AI model trained only on this initial dataset, and two adaptive AI models continued learning from user behaviour during their deployment. One of the adaptive AI models also showed explanations for why users should stop scrolling, such as including the amount of time the subject had spent on the app or pointing out that it is late at night.
Next, the researchers tested the different reminder systems for four weeks. When using an app, the test subjects received periodic pop-up reminders prompting them to enter a sequence of numbers if they wanted to continue scrolling. The pop-ups also asked for optional feedback to further refine AI predictions by asking users to click 鈥渁gree鈥� or 鈥渄isagree鈥� on statements about their app usage. Then the system tracked whether people quit or continued using the app.
The adaptive AI reminders proved most successful, reducing overall app visit frequency by up to 9 per cent. People were almost 30 per cent more likely to quit using an app when prompted by an adaptive AI reminder compared with the probability-based intervention, says at the Korea Advanced Institute of Science and Technology, a coauthor on the study.
The AI interventions seem 鈥減romising鈥� but are also 鈥渧ery preliminary鈥�, says at Stanford University in California. The high drop-out rates 鈥� 49 out of 127 students originally recruited stopped participating 鈥� could have also skewed the results. But she praised the overall research direction, despite the measured impact being 鈥渟mall and short-term鈥�.
鈥淎 more powerful version of this tool [could] appeal to the user鈥檚 emotion, by getting them to talk about why they want to change their phone usage behaviours up front, and then incorporating that information into the AI feedback later,鈥� says Lembke.
arXiv
