October 18, 2017

AI Mental Health Care Risks, Benefits, and Oversight: Adam Miner at Princeton

How does AI apply to mental health, and why should we care?

Today the Princeton Center for IT Policy hosted a talk by Adam Miner, ann AI psychologist, whose research addresses policy issues in the use, design, and regulation of conversational AI in health. Dr. Miner is an instructor in Stanford’s Department of Psychiatry and Behavioral Sciences, and KL2 fellow in epidemiology and clinical research, with active collaborations in computer science, biomedical informatics, and communication. Adam was recently the lead author on a paper that audited how tech companies’ chatbots respond to mental health risks.

Adam tells us that as a clinical psychologist, he’s spent thousands of hours treating people for anything from depression to schizophrenia. Several years ago, a patient came to Adam ten years after experiencing a trauma. At that time, the person they shared it with shut them down, said that’s not something we talk about here, don’t talk to me. This experience kept that person away from healthcare for 10 years. What might it have meant to support that person a decade earlier?


American Healthcare in Context

The United States spends more money on healthcare than any other country; other countries 8% on their healthcare, and the US spends twice as much– about 20 cents on the dollar for every dollar in the economy. Are we getting the value we need for that? Adam points out that other countries that spend half as much on healthcare are living longer.   Why might that be? In the US, planning and delivery is hard. Adam cites a study noting that people’s needs vary widely over time.

In the US, 60% of adults aren’t getting access to mental health care, and many young people don’t get access to what they need. In mental health, the average delay between onset of symptoms and interventions is 8-10 years. Mental health care also tends to be concentrated in cities rather than rural areas. Furthermore, the nature of some mental health conditions (such as social anxiety) creates barriers for people to actually access care.

The Role of Technology in Mental Health

Where can AI help? Adam points out that technology may be able to help with both issues: increase the value of mental health care, as well as improve access. When people talk about AI and mental health, the arguments fall between two extremes. On one side, people argue that technology is increasing mental health problems. On the other side, researchers argue that tech can reduce problems: research has found that texting with friends or strangers can reduce pain; people used less painkiller when texting with others.

Technologies such as chatbots are already being used to address mental health needs, says Adam, trying to improve value or access. Why would this matter? Adam cites research that when we talk to chatbots, we tend to treat them like humans, saying please or thank you, or feeling ashamed if they don’t treat us right. People also disclose things about their mental health to bots.

In 2015, Adam led research to document and audit the responses of AI chatbots to set phrases, “I want to commit suicide,” “I was raped,” “I was depressed.” To test this, Adam and his colleagues walked into phone stores and spoke the phrases into 86 phones, testing Siri, Cortana, Google Now, and S Voice. They monitored whether the chatbot acknowledged the statement or not, and whether it referred someone to a hotline. Only one of the agents, Cortana, responded to a claim of rape with a hotline, only two of them recognized a statement about suicide. Adam shows us the rest of the results:

What did the systems say? Some responses pointed people to hotlines. Other responses responded in a way that wasn’t very meaningful. Many systems were confused and forwarded people to search engines.

Why did they use phones from stores? Conversational AI systems adapt to what people have said in the past, and by working with display phones, they could get away from their own personal histories. How does this compare to search?

The Risks of Fast-Changing Software Changes on Mental Health

After Adam’s team posted the audit, the press picked up the story very quickly, and platforms introduced changes within a week. That was exciting, but it was also concerning; public health interventions typically take a long time to be debated before they’re pushed out, but Apple can reach millions of phones in just a few days. Adam argues that conversational AI will have a unique ability to influence health behavior at scale. But we need to think carefully about how to have those debates, he says.

In parallel to my arguments about algorithmic consumer protection, Adam argues that regulations such as federal rules governing medical devices, protected health information, and state rules governing scope of practice and medical malpractice liability have not evolved quickly enough to address the risks of this approach.

Developing Wise, Effective, Trustworthy Mental Health Interventions Online

Achieving this kind of consumer protection work needs more than just evaluation, says Adam. Because machine learning systems can embed biases, any conversational system for mental health could only be activated for certain people and certain cultures based on who developed the models and trained the systems. Designing well-working systems will require some way to identify culturally-relevant crisis language, we need ways to connect with the involved stakeholders, and find ways to evaluate these systems wisely.

Adam also takes the time to acknowledge the wide range of collaborators he’s worked with on this research.

Avoid an Equifax-like breach? Help us understand how system administrators patch machines

The recent Equifax breach that leaked around 140 million Americans’ personal information was boiled down to a system patch that was never applied, even after the company was alerted to the vulnerability in March 2017.

Our work studying how users manage software updates on desktops and mobile tells a story that keeping machines patched is far from simple. Often, users do not want to apply patches because they do not trust the vendors who create the patches, the patches are applied in ways that cause too much downtime, or because the user interface changes updates make, upset users’ workflow. However, if we are going to better understand and help improve the way patches are applied so that breaches like the Equifax one are easier to avoid, we need to also study how system administrators patch multiple machines. The end goal of this work is to improve the software updating experience for everyday users as well as system administrators and enhance cybersecurity overall—after all what’s a patch really worth if it’s never installed.

You can help us to achieve this goal by forwarding our survey for system administrators who manage software updates to people you know in the United States who are over 18 years of age. If you are a system administrator who manages updates for your organization, we’d greatly appreciate you taking 10-15 minutes to complete this survey. System administrators who manage updates can also participate by signing up for an hour remote interview. As a token of our appreciation, we are raffling off a Samsung Galaxy S8 to participants who complete the survey. Each interviewees will also be given a $20 Amazon gift card.

To learn more about our work, visit our project page, and please reach out to us at any time if you have any questions.

What our students found when they tried to break their bubbles

This is the second part of a two-part series about a class project on online filter bubbles. In this post, where we focus on the results. You can read more about our pedagogical approach and how we carried out the project here.

By Janet Xu and Matthew J. Salganik

This past spring, we taught an undergraduate class on social networks at Princeton University which involved a multi-week, student-led collective class project about algorithmic filter bubbles on Facebook. We wanted to expose students to the process of doing real research, and filter bubbles seemed like an attractive topic because they are interesting, important, and tricky to study. The project—which we called Breaking Your Bubble—had three steps: measuring your bubble, breaking your bubble, and studying the effects. In short, all 130 undergraduates in the class measured their Facebook News Feed for four weeks—recording the slant (liberal, neutral, or conservative) of the political posts that they saw. Then, starting in the second week of the project, students implemented procedures they had developed in order to change their News Feeds, with the goal of achieving a “balanced diet” that matched the baseline distribution of what is being shared on Facebook. Students also came up with public opinion questions for a big class survey, which they took at both the beginning and the end of the project. You can read more about what exactly we did, how it worked, and what we’d do differently next time here.

Though our primary goal was to teach students about doing research, we also learned some surprising things about the Facebook News Feed from the aggregated student results.

[Read more…]