Alison Darcy

2021 - Centre Stage HM1 8091 (51648554472)

Alison Darcy is a research psychologist and technologist. She is the Founder and President of Woebot Health, a company which provides digital therapeutics and behavioural health products

Quotes

• Initially it was myself and Pierre building web prototypes and games. We tried to make video games for CBT for 9 months, things like interactive fiction prototypes for dynamics of engagement. They all had so much dialogue it made sense for us to move to chatbots. Pamela Fox stepped in as our head of engineering and helped rebuild the stack.
  • Team — how did your team come together, and what are the roles?
• My background is in clinical psychology research and I reached out to the best in the field. We all agree that our work doesn’t scale. Athena Robinson, a former Stanford psychiatry professor just joined as our Chief Clinical Officer. Other than that, some of my other colleagues or friends have joined too — people who really care about mental health issues. It makes for a really great workplace — this idea is bigger than any of us. When you look at the data and what people share, it’s so personal and you don’t even hear things like this in human therapy.
  • Team — how did your team come together, and what are the roles?

• At the end, we always ask: How are you feeling, better, the same, or worse? We look at the failed transcripts and try to troubleshoot on what went wrong — we try to minimize people feeling worse. It’s hard to figure out what we can do for those feeling the same. When you get it wrong people are not shy about telling you — when a button doesn’t cover everything users want to say, users will tell you.
  • How do you measure success; what are your metrics? What are successful interactions? What are failed interactions?
• We also use standard search metrics, like precision and recall. For example, to detect people in crisis, we prefer recall search methods over precision as we want to be overinclusive and to identify anyone who could have a problem so we can refer them to human counselors or a help hotline.
  • How do you measure success; what are your metrics? What are successful interactions? What are failed interactions?

• We learn all the time and have learned to keep track of our learning. If you have two buttons, they should represent genuinely different pathways. If there is no natural response or utterance, we use an emoji as an easy button filler. We’ve learned a lot about images. We used to have a black and white image of a bomb and we found it was triggering for certain people, so we had to remove it. Generally I like the black and white images. Many people really dislike minions. There’s a lot due to personal taste here — some people are into videos, others aren’t into it. We have veered away from videos, but we may make some to help teach people difficult techniques. When people are upset, they can only process a little — so our language and scripts have to be really short. We find it easier to have many chat bubbles with 1 or 2 lines — we try to keep it lean and bouncy. There’s almost a rhythm to it — there are a few flows where we can capture help.
  • Editorial and scripting — what have you learn from flows so far? What have you learned from building rails and making a fun, engaging experiences (with images, gifs, emojis, etc)?

• Some users want to be served something else. We always issue an invitation to engage in a conversation. There’s no assumption people want help. Some people just want to chat, they don’t want help. So we try not to explicitly ask people what they are looking for. There really isn’t a lot of open chat away from buttons in Woebot. People may want a past video or lesson.
  • What are the most common things users ask outside of the main function — do they ask for jokes or other advice? That is, when they stray, where do they go?

• The only strategy that we’ve done is press — that has been reasonably successful. Interestingly, we got far more conversations as a result of blog article reviews than mainstream press. Bloggers just seem to find us. The tongue and cheek name helped and launching with data and the study helped. Some of the other mental health bots are less than useful and possibly dangerous — so having actual data from a study was important.
  • User acquisition strategy — how do people hear about your bot and start using it?

• The first is to just ask people if they want to converse. I’m not sure if we’ve nailed the invitation. Some of our push navigations were like “beep boop” — we are trying to understand the re-triggering schedule. [Note: Checking in to see what someone’s mental health is is a great reason to re-engage — Woebot naturally does this.]
  • Re-triggering and re-use strategy — how do you get initial users to engage again?

• It was a nice to have a paywall at first to get validation that people would pay for it. We had a decent conversion rate, but I wanted to gather data. Direct to consumer will be the longer term model. People were emailing us saying this is less expensive than our therapist; they were the ones who valued us the most. Convenience was a huge value proposition.
  • Monetization? How did you decide to go from charging $39 a month to free?

• It’s all completely de-identified and anonymized data — nobody in the company can ever see anyone’s Facebook profile. It was to protect the user and to protect us. The guiding principle is transparency. There is no open ended generative conversation — everything is scripted. We want to launch an app to have more privacy.
  • How do you think about the ownership of data and the privacy issues? It sounds like you have it set up so no clinician ever sees a user’s name?

• We use AWS Lambda and NodeJS. Our app will have an animated Woebot tied to NLP that can respond to verbal language with animation. We do our analytics in house as we move to HIPAA compliance for our apps. We don’t want a 3rd party looking at data. We built a dashboard over the summer. It’s hard to find really good data scientists and AI people, and Android engineers.
  • What can you tell us about your tech stack? Do you do NLP in-house, and what external services do you like?

• We are planning an iOS and Android app so people have extra privacy and to be HIPAA-compliant. There’s a good reason to keep it with text and to NOT do it in voice. You cannot see a negative thought in voice, but in a text chatbot you write it out and this externalization helps you overcome it. Messenger was a great launch platform — they make it easy to create and launch a bot. Users loved our prototypes there and we decided to launch there as it was easy.
  • Thoughts on the different platforms? FB Messenger vs iOS / Android apps? Do you plan on making a voice version?

• I really like ParentSpark — it helps parents with parenting. Their storytelling is great. It’s a mom coaching her kids and they help you as a parent. I also like Jeyant for medical screening; they have over a million users and screen for Zika and other diseases. PullString has done some nice character-based chatbots like Dr. Who. The Mabu chatbot, with eyes that follow you, is neat and I think IDEO designed it.
  • What other bots have you looked to for inspiration — what other chatbots made you say “WOW”? Are there other use cases you’ve thought were simply brilliant?