Should We Give Chatbots a Chance for Semi-Therapy?
How many times in life you have been told “you should talk to a therapist or psychiatrist”, but have you ever been referred to a chatbot instead? I doubt if a GP or a close friend would make that suggestion to someone that is experiencing emotional distress or trauma since it does not show full sympathy or respect for someone else’s pain. right?
Chatbots — also known as “conversational agents” — are software applications that mimic written or spoken human speech for the purposes of simulating a conversation or interaction with a real person. At the heart of chatbot technology lies Natural Language Processing (and Natural Lanuguage Understaning), the same technology that forms the basis of the voice recognition systems such as Apple’s Siri and Microsoft’s Cortana. Here you can find the 10 most innovative applications of chatbot technology.
There is a growing body of research looking into how conversational chatbots or other kinds of digital means can be safely integrated within psychiatry and mental wellbeing [1]. Nonetheless, not many of us have a positive stance towards getting support from a machine in the midst of a life crisis. Even if receiving pretend empathy from a machine makes us feel better, we may come to think that we are putting ourselves in a situation that diminishes us as human beings [2]. This reminds me of the movie Her, the story of a talented writer who develops an unexpected rapport with his operating system as it evolves into a bona fide companion.
Okay, let’s say we don’t want any sort of pretend empathy or destructive machine intimacy (I hope we are not excluding a surreal partner). How about we look at chatbots from a different perspective? Let’s say you are going through a complex or unfamiliar situation that is painful and overwhelming. For whatever reasons, you cannot immediately reach out to a wise man, or a trustworthy friend, or a therapist that can assist you in alleviating distress. Instead, you only have access to a machine that you are aware hasn’t lived a human life, but instead, its brain has been exposed and trained to thousands of ‘non-comfortable’ situations that real people have gone through in real life. How do you feel about that?
Training is an important and intuitive concept in machine learning. It simply means determining good values from labeled examples [3]. In supervised learning, often a machine learning algorithm builds a model by examining many examples and attempts to find a model that minimizes loss, the penalty for a bad prediction. In our context, examples can be therapeutic conversations (in the form of video, audio, transcriptions, questionnaires, etc) and various labels can be assigned (either manually or computationally) based on target disorder, the severity of pain, patient experience, treatment regime, what works, what didn’t work to build a corpus of labeled examples. If the aim is building a chatbot for the diagnose and treatment of mental health, the loss function can be interpreted as minimizing incorrect treatment outcomes.
With this little induction, you may be willing to give chatbots a shot for mental health support and be more persistent in using them for a longer time. Okay. So what are the options? I came across this post from Healthline that features an interesting test run of four mental health chatbots including Woebot, Wysa, Joyable, and Talkspace. The feedbacks are provided by a clinical psychologist and two people who’ve tried in-person therapy. I leave it to you to study this post and make your own decision on which one you would like to try. I myself picked the first one Woebot.
Get Arezou Soltani’s stories in your inbox
Join Medium for free to get updates from this writer.
Woebot knows its place: Wobot is developed by researchers from Standford University and falls under the category of Internet-based Cognitive Behavioural Therapy (CBT), a psycho-social intervention aimed at changing negative thought patterns and developing coping strategies to solve problems. I installed it from the AppStore and experimented with it for a week. This is how Woebot introduced itself:
Woebot’s conversations were almost entirely scripted. You can either select from predefined options such as ‘Help me feel better’ and be guided for trying different techniques such as “Challenge Negativity”, “Challenge Stress”, “Just Breath”, “Gratitude Journal”, or use freestyle typing. When I strayed away from the script, Woebot didn’t get tripped up too much. But I wouldn’t recommend asking complex empathy questions or doing any sort of Turing test on the Woebot. The creators have explicitly mentioned that their app is not a substitute for therapy and there is no wild exaggeration about Woebot’s capabilities. Below you can see some unsuccessful examples. Woebot also offers ways of reaching someone if your situation is serious.
Overall, I liked Woebot and I believed it is a good medium for practicing good thinking hygiene. Another interesting aspect for me was Woebot responses. From the listener's point of view, it may have happened to you that someone you know or care about reached out to share his grief or anxiety about something, but you may not have enough life experience or capacity to support that person. Instead of saying nonsense or crude responses, sometimes it is best just to say “I hear you” like Woebot. Or instead of showing toxic positivity or other distorted thinking patterns such as all-or-nothing thinking, we can help people to get down to a healthy and adaptive level of negative emotions that make sense to hold on as a human. Having said that, towards the end of the trial week, I could easily guess Woebot responses and didn’t feel the need to keep up our conversation.
Can I also talk to Meena? Well, I couldn’t. Meena is the new Google AI chatbot that has been trained on 341 gigabytes of public social media conversations[3]. Despite the four aforementioned chatbots, Meena is an open-domain virtual assistant meaning you can have any conversation about any topic with it and it is claimed to be the best chatbot ever created. It doesn’t sound to be specialized in mental health, but since it is beating the state-of-the-art evaluation metrics, sounds very tempting to converse with (at least as a companion).
Final note
The other day, I came across the following tweet and couldn't stop laughing.
If we formulate human therapy as a problem we want to solve with a machine, I don’t think any chatbot with state-of-the-art performance will be the far-reaching solution. Remember, we are talking about a complex problem that can drive individuals upwards in the Big data Maslow’s Pyramid if it can be solved. We may need to redefine the problem at the hand and think of a chatbot as a lower-intensity option compared to a human-therapy. There is no need to compete with the empathy that defines what is unique about us being human. We can just accept, we are talking to a machine that has observed many uncomfortable situations that other people have gone through and have a conversation.
[1] Lovejoy, C.A., 2019. Technology and mental health: the role of artificial intelligence. European Psychiatry, 55, pp.1–3.
[2] Turkle, S., 2017. Alone together: Why we expect more from technology and less from each other. Hachette UK.
[3] https://developers.google.com/machine-learning/crash-course
[4] https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.html
