Smartphones and Teen Depression: There Might Be an App for That

Two of the most striking features of modern teenage life are ubiquitous smartphone usage and rising rates of depression, a correlation that has led many parents, teachers, clinicians, and policymakers to question whether phones are a problem. But what if they can be part of the solution?

That’s the provocative idea behind a new study by Columbia researchers, who found that AI’s large language models, which power popular ChatGPT, Claude, and Gemini chatbots, can analyze teens’ smartphone data and predict the emergence of depressed moods within days. The team is now working to develop a smartphone application that can turn those predictions into interventions, nudging teens at risk of depression toward more positive thought patterns.

“Given how quickly and efficiently AI can analyze smartphone text entries, an AI-powered phone app could deliver just-in-time interventions to teens to prevent depressive episodes,” says study leader Isaac Treves, a postdoctoral research scientist in the Department of Psychiatry at Columbia University Vagelos College of Physicians and Surgeons.

AI outsmarts older algorithms

With his adviser Randy P. Auerbach, Treves has been searching for methods that can better predict—and even prevent—depression in teens. Treves turned his attention to rumination, an excessive focus on negative thoughts and emotions that is a common trait in people who develop depression. Auerbach and Treves hypothesized that teens’ smartphone data could be an ideal source for detecting negative self-talk, often an indicator of rumination, and to see whether it would predict upcoming mood changes.

In the study, the researchers recruited over 200 teenagers who installed a smartphone app that recorded every word they typed—from browser searches and text messages to social media posts and emails. The resulting dataset was enormous: More than 4.5 million entries were collected, with each teen logging more than 17,000 entries on average during the 12-month collection phase.

The teens were surprisingly open to the idea. “If teenagers know that the monitoring is intended to support their mental health, and automated, that there isn’t someone looking over their shoulder in real time, they are more likely to accept it,” says Treves.

It’s not the first time that scientists have mined smartphone data to try to detect depression, but previous efforts used relatively simple, rule-based algorithms that look for words like ‘hate’ or ‘kill.’ “They are often inaccurate when looking at teen smartphone language,” says Treves.

By contrast, when labelling the sentiment of the teens’ text, the researchers’ best AI model agreed with human experts 84% of the time. The study also showed that AI-detected negative sentiment could predict increased depressive symptoms and worse moods in the following days.

Future app?

The model required fine tuning by experienced clinicians to achieve those results, and Treves emphasizes that human involvement will remain a critical part of the process, even if a smartphone app is rolled out.

Because of the concern about teen depression and suicide, computer and smartphone monitoring is a fraught subject. “It’s already a problem. School districts monitor teens’ language on shared devices to look for suicide-related concerns and then send emergency teams to people’s houses,” says Treves. “Sometimes that has helped people get the crisis services they need, but other times it has led to traumatizing and wasteful emergency responses.”

Treves envisions a different, less invasive model where the smartphone app would complement traditional therapy. The smartphone app could use signals like negative self-talk to provide helpful nudges toward more positive activities and thought patterns. In therapy sessions, clinicians could review activities, look at signals from language to understand smartphone use, and provide more intensive interventions if necessary.

Now, the researchers are hoping to obtain funding to build the smartphone app. The app itself will have two types of interventions: distraction and problem-solving. “The idea is that when you’re in these moments of self-criticism and rumination, the app recognizes that and helps people find meaningful, absorbing activities that take their mind off their problems, followed by concrete problem-solving techniques to address the problems,” Treves says.

The proposed app’s steady but hidden observation sets Treves’s approach apart from existing mental health apps that interrupt their users with daily questions. “We know that depressed teens don’t want to respond all the time. They’re often resistant to suggestions, and they learn to adjust their responses to the app to evade those suggestions,” Treves says. “Our tool could be more acceptable to teens and potentially more effective.”