:format(webp)/cdn.vox-cdn.com/uploads/chorus_image/image/73331996/681977094.0.jpg)
If you have a sore throat, you can get tested for a host of things — Covid, RSV, strep, the flu — and receive a pretty accurate diagnosis (and maybe even treatment). Even when you’re not sick, vital signs like heart rate and blood pressure give doctors a decent sense of your physical health.
But there’s no agreed-upon vital sign for mental health. There may be occasional mental health screenings at the doctor’s office, or notes left behind after a visit with a therapist. Unfortunately, people lie to their therapists all the time (one study estimated that over 90 percent of us have lied to a therapist at least once), leaving holes in their already limited mental health records. And that’s assuming someone can connect with a therapist — roughly 122 million Americans live in areas without enough mental health professionals to go around.
But the vast majority of people in the US do have access to a cellphone. Over the last several years, academic researchers and startups have built AI-powered apps that use phones, smart watches, and social media to spot warning signs of depression. By collecting massive amounts of information, AI models can learn to spot subtle changes in a person’s body and behavior that may indicate mental health problems. Many digital mental health apps only exist in the research world (for now), but some are available to download — and other forms of passive data collection are already being deployed by social media platforms and health care providers to flag potential crises (it’s probably somewhere in the terms of service you didn’t read).
The hope is for these platforms to help people affordably access mental health care when they need it most, and intervene quickly in times of crisis. Michael Aratow — co-founder and chief medical officer of Ellipsis Health, a company that uses AI to predict mental health from human voice samples — argues that the need for digital mental health solutions is so great, it can no longer be addressed by the health care system alone. “There’s no way that we’re going to deal with our mental health issues without technology,” he said.
And those issues are significant: Rates of mental illness have skyrocketed over the past several years. Roughly 29 percent of US adults have been diagnosed with depression at some point in their lives, and the National Institute of Mental Health estimates that nearly a third of US adults will experience an anxiety disorder at some point.
While phones are often framed as a cause of mental health problems, they can also be part of the solution — but only if we create tech that works reliably and mitigates the risk of unintended harm. Tech companies can misuse highly sensitive data gathered from people at their most vulnerable moments — with little regulation to stop them. Digital mental health app developers still have a lot of work to do to earn the trust of their users, but the stakes around the US mental health crisis are high enough that we shouldn’t automatically dismiss AI-powered solutions out of fear.
How does AI detect depression?
To be formally diagnosed with depression, someone needs to express at least five symptoms (like feeling sad, losing interest in things, or being unusually exhausted) for at least two consecutive weeks.
But Nicholas Jacobson, an assistant professor in biomedical data science and psychiatry at the Geisel School of Medicine at Dartmouth College, believes “the way that we think about depression is wrong, as a field.” By only looking for stably presenting symptoms, doctors can miss the daily ebbs and flows that people with depression experience. “These depression symptoms change really fast,” Jacobson said, “and our traditional treatments are usually very, very slow.”
Even the most devoted therapy-goers typically see a therapist about once a week (and with sessions starting around $100, often not covered by insurance, once a week is already cost-prohibitive for many people). One 2022 study found that only 18.5 percent of psychiatrists sampled were accepting new patients, leading to average wait times of over two months for in-person appointments. But your smartphone (or your fitness tracker) can log your steps, heart rate, sleep patterns, and even your social media use, painting a far more comprehensive picture of your mental health than conversations with a therapist can alone.
One potential mental health solution: Collect data from your smartphone and wearables as you go about your day, and use that data to train AI models to predict when your mood is about to dip. In a study co-authored by Jacobson this February, researchers built a depression detection app called MoodCapture, which harnesses a user’s front-facing camera to automatically snap selfies while they answer questions about their mood, with participants pinged to complete the survey three times a day. An AI model correlated their responses — rating in-the-moment feelings like sadness and hopelessness — with these pictures, using their facial features and other context clues like lighting and background objects to predict early signs of depression. (One example: a participant who looks as if they’re in bed almost every time they complete the survey is more likely to be depressed.)
The model doesn’t try to flag certain facial features as depressive. Rather, the model looks for subtle changes within each user, like their facial expressions, or how they tend to hold their phone. MoodCapture accurately identified depression symptoms with about 75 percent accuracy (in other words, if 100 out of a million people have depression, the model should be able to identify 75 out of the 100) — the first time such candid images have been used to detect mental illness in this way.
In this study, the researchers only recruited participants who were already diagnosed with depression, and each photo was tagged with the participant’s own rating of their depression symptoms. Eventually, the app aims to use photos captured when users unlock their phones using face recognition, adding up to hundreds of images per day. This data, combined with other passively gathered phone data like sleep hours, text messages, and social media posts, could evaluate the user’s unfiltered, unguarded feelings. You can tell your therapist whatever you want, but enough data could reveal the truth.
The app is still far from perfect. MoodCapture was more accurate at predicting depression in white people because most study participants were white women — generally, AI models are only as good as the training data they’re provided. Research apps like MoodCapture are required to get informed consent from all of their participants, and university studies are overseen by the campus’s Institutional Review Board (IRB) But if sensitive data is collected without a user’s consent, the constant monitoring can feel creepy or violating. Stevie Chancellor, an assistant professor in computer science and engineering at the University of Minnesota, says that with informed consent, tools like this can be “really good because they notice things that you may not notice yourself.”
What technology is already out there, and what’s on the way?
Of the roughly 10,000 (and counting) digital mental health apps recognized by the mHealth Index & Navigation Database (MIND), 18 of them passively collect user data. Unlike the research app MoodCapture, none use auto-captured selfies (or any type of data, for that matter) to predict whether the user is depressed. A handful of popular, highly rated apps like Bearable — made by and for people with chronic health conditions, from bipolar disorder to fibromyalgia — track customized collections of symptoms over time, in part by passively collecting data from wearables. “You can’t manage what you can’t measure,” Aratow said.
These tracker apps are more like journals than predictors, though — they don’t do anything with the information they collect, other than show it to the user to give them a better sense of how lifestyle factors (like what they eat, or how much they sleep) affect their symptoms. Some patients take screenshots of their app data to show their doctors so they can provide more informed advice. Other tools, like the Ellipsis Health voice sensor, aren’t downloadable apps at all. Rather, they operate behind the scenes as “clinical decision support tools,” designed to predict someone’s depression and anxiety levels from the sound of their voice during, say, a routine call with their health care provider. And massive tech companies like Meta use AI to flag, and sometimes delete, posts about self-harm and suicide.
Some researchers want to take passive data collection to more radical lengths. Georgios Christopoulos, a cognitive neuroscientist at Nanyang Technological University in Singapore, co-led a 2021 study that predicted depression risk from Fitbit data. In a press release, he expressed his vision for more ubiquitous data collection, where “such signals could be integrated with Smart Buildings or even Smart Cities initiatives: Imagine a hospital or a military unit that could use these signals to identify people at risk.” This raises an obvious question: In this imagined future world, what happens if the all-seeing algorithm deems you sad?
AI has improved so much in the last five years alone that it’s not a stretch to say that, in the next decade, mood-predicting apps will exist — and if preliminary tests continue to look promising, they might even work. Whether that comes as a relief or fills you with dread, as mood-predicting digital health tools begin to move out of academic research settings and into the app stores, developers and regulators need to seriously consider what they’ll do with the information they gather.
So, your phone thinks you’re depressed — now what?
It depends, said Chancellor. Interventions need to strike a careful balance: keeping the user safe, without “completely wiping out important parts of their life.” Banning someone from Instagram for posting about self-harm, for instance, could cut someone off from valuable support networks, causing more harm than good. The best way for an app to provide support that a user actually wants, Chancellor said, is to ask them.
Munmun De Choudhury, an associate professor in the School of Interactive Computing at Georgia Tech, believes that any digital mental health platform can be ethical, “to the extent that people have an ability to consent to its use.” She emphasized, “If there is no consent from the person, it doesn’t matter what the intervention is — it’s probably going to be inappropriate.”
Academic researchers like Jacobson and Chancellor have to jump through a lot of regulatory hoops to test their digital mental health tools. But when it comes to tech companies, those barriers don’t really exist. Laws like the US Health Insurance Portability and Accountability Act (HIPAA) don’t clearly cover nonclinical data that can be used to infer something about someone’s health — like social media posts, patterns of phone usage, or selfies.
Even when a company says that they treat user data as protected health information (PHI), it’s not protected by federal law — data only qualifies as PHI if it comes from a “healthcare service event,” like medical records or a hospital bill. Text conversations via platforms like Woebot and BetterHelp may feel confidential, but crucial caveats about data privacy (while companies can opt into HIPAA compliance, user data isn’t legally classified as protected health information) often wind up where users are least likely to see them — like in lengthy terms of service agreements that practically no one reads. Woebot, for example, has a particularly reader-friendly terms of service, but at a whopping 5,625 words, it’s still far more than most people are willing to engage with.
“There’s not a whole lot of regulation that would prevent folks from essentially embedding all of this within the terms of service agreement,” said Jacobson. De Choudhury laughed about it. “Honestly,” she told me, “I’ve studied these platforms for almost two decades now. I still don’t understand what those terms of service are saying.”
“We need to make sure that the terms of service, where we all click ‘I agree’, is actually in a form that a lay individual can understand,” De Choudhury said. Last month, Sachin Pendse, a graduate student in De Choudhury’s research group, co-authored guidance on how developers can create “consent-forward” apps that proactively earn the trust of their users. The idea is borrowed from the “Yes means yes” model for affirmative sexual consent, because FRIES applies here, too: a user’s consent to data usage should always be freely given, reversible, informed, enthusiastic, and specific.
But when algorithms (like humans) inevitably make mistakes, even the most consent-forward app could do something a user doesn’t want. The stakes can be high. In 2018, for example, a Meta algorithm used text data from Messenger and WhatsApp to detect messages expressing suicidal intent, triggering over a thousand “wellness checks,” or nonconsensual active rescues. Few specific details about how their algorithm works are publicly available. Meta clarifies that they use pattern-recognition techniques based on lots of training examples, rather than simply flagging words relating to death or sadness — but not much else.
These interventions often involve police officers (who carry weapons and don’t always receive crisis intervention training) and can make things worse for someone already in crisis (especially if they thought they were just chatting with a trusted friend, not a suicide hotline). “We will never be able to guarantee that things are always safe, but at minimum, we need to do the converse: make sure that they are not unsafe,” De Choudhury said.
Some large digital mental health groups have faced lawsuits over their irresponsible handling of user data. In 2022, Crisis Text Line, one of the biggest mental health support lines (and often provided as a resource in articles like this one), got caught using data from people’s online text conversations to train customer service chatbots for their for-profit spinoff, Loris. And last year, the Federal Trade Commission ordered BetterHelp to pay a $7.8 million fine after being accused of sharing people’s personal health data with Facebook, Snapchat, Pinterest, and Criteo, an advertising company.
Chancellor said that while companies like BetterHelp may not be operating in bad faith — the medical system is slow, understaffed, and expensive, and in many ways, they’re trying to help people get past these barriers — they need to more clearly communicate their data privacy policies with customers. While startups can choose to sell people’s personal information to third parties, Chancellor said, “no therapist is ever going to put your data out there for advertisers.”
Someday, Chancellor hopes that mental health care will be structured more like cancer care is today, where people receive support from a team of specialists (not all doctors), including friends and family. She sees tech platforms as “an additional layer” of care — and at least for now, one of the only forms of care available to people in underserved communities.
Even if all the ethical and technical kinks get ironed out, and digital health platforms work exactly as intended, they’re still powered by machines. “Human connection will remain incredibly valuable and central to helping people overcome mental health struggles,” De Choudhury told me. “I don’t think it can ever be replaced.”
And when asked what the perfect mental health app would look like, she simply said, “I hope it doesn’t pretend to be a human.”
