Lessons From Writing for a Mental Health Chatbot

Lessons From Writing for a Mental Health Chatbot

Insight / Cancer

Surviving cancer is an experience chock full of “supposed tos.” After treatment, if doctors tell you that you have no evidence of disease, you’re supposed to feel a great sense of relief and get back to your normal life. If you’re burdened with chronic cancer, you’re supposed to keep living as best as you can for as long as you can as if everything were normal. But once you’ve had cancer, you know the old normal is gone. And your new normal can be unsettling and unfamiliar.

Meet Vivibot

There’s a chatbot for young people who feel overwhelmed by all the unsettling, unfamiliar emotions those “supposed tos” bring up. Her name is Vivibot, and she’d reassure you that most of those supposed tos are an unhelpful source of stress. Actually, she’d probably call them “sucktastic.”

Vivibot is a chatbot with a big personality. In more technical terms, Vivibot is a conversational user interface (CUI) that engages and supports adolescents and young adults (AYA) ages 15 to 29 who are dealing with life beyond cancer. The goal of Vivibot is to share techniques based on science that people can use to deal with all the feelings cancer brings up—or as doctors would put it, positive psychology coping skills.

 

Hopelab has long focused on the challenges faced by young people who have had cancer.

In May 2016, Hopelab kicked off a cancer mindshift project that ultimately evolved into Vivibot. Two years of research and development went into Vivibot before I joined the team as a user experience (UX) writer in the winter of 2018. As a consultant, I’d worked on many projects in communications strategy, identity, and copywriting, including several youth health interventions. I was also, briefly, a scientist. This all made me a good match for the Vivibot project.

When I came on board, Vivibot was already a robot prototype talking to people on Facebook Messenger. Beta testing was beginning to reveal that some of Vivibot’s intended charm needed fine tuning. The Hopelab team knew it had to make small and large changes to Vivibot’s personality, language, and the flow of conversation.

Testing the Opening Message

One of the first changes we tested was small, but addressed one of the most important elements of the chatbot script. We suspected Vivibot’s first few text messages could make or break the conversation days or even weeks later. Even robots have to make a positive first impression. Vivibot’s second message had to do most of the heavy lifting in the introductory conversation. It needed to establish a clear value proposition that explained why someone would want to talk to Vivibot, introduce her personality, and set an expectation that would entice the user to keep coming back to chat. All in a single succinct text message.

Hopelab asked me to write several variations of the message, each with a distinctly different value proposition. Here’s the first test we ran comparing the original message to three alternate versions:

When we saw the +5.3% retention boost we got by changing up just this one message, we knew we had to adjust more content and run more experiments. Over the next seven months, we wound up reevaluating every word, emoji, and image Vivibot messaged to users. I created about a half dozen versions of what we call the “Day 1” conversation. I pored over thousands of tweets to absorb the weird and wonderful ways people manipulate language to express their emotions. I scoured the internet for GIFs and memes, and created new ones when needed.

 

 

As far as I could tell, there was no editorial or writing style guide about creating engaging content specifically for young people trying to navigate a post-treatment life . . . So I decided to make one.

As I wrote and rewrote more of Vivibot’s supportive chatter, I wanted my writing to be informed by as much expertise as possible. I dove down the research rabbit hole. Cancer Research UK and Cancer Institute NSW (Australia) have both shared guidelines for writing clinical patient information in a way that’s supportive and clear. There is some existing academic research on language that is supportive of cancer survivorship resilience. The Conscious Style Guide site has a library of resources about language that is empowering, respectful, and inclusive, but little specific to cancer. There’s a hodgepodge of useful if anecdotal essays, think-pieces, and even comics about the mindset of living life after cancer.

But, as far as I could tell, there was no editorial or writing style guide about creating engaging content specifically for young people trying to navigate a post-treatment life, much less about creating a chatbot for them.

So I decided to make one.

What ended up being titled “The Vivibot Product Design Guidelines” eventually became much more than a writing style guide. It documents decisions we made about everything from word choice to user-response buttons to emoji skin tone to respecting gender identities to the perils of putting ellipses (…) in text messages.

While much of the Vivibot guidelines are specific to, well, Vivibot, we learned a lot that’s worth sharing with others who support AYA cancer survivors, as well as with chatbot and conversational user interface designers.

Refining the chatbot personality

In the earliest phase of product development, Hopelab created a persona for Vivibot based on characteristics of real AYAs dealing with life beyond cancer. This persona, along with user input, guided all product decisions. Then, as prototyping and testing revealed what users liked and didn’t like, Hopelab evolved its understanding of what people needed Vivibot’s personality to be.

One thing that stood out was how strongly users reacted when Vivibot’s personality drifted from its core. Just like a real person, Vivibot’s personality is multi-faceted and nuanced, but that does not mean Vivibot’s personality could be inconsistent. Beyond what Vivibot’s users were telling us, research has shown that people prefer consistent interactive characters, and that consistent characters have a greater influence over user behavior.

By the end of beta, we’d honed Vivibot’s voice to communicate a unique and strong personality, and made thousands of tweaks to the script to make her more consistent.

Core Character Personality Traits

  • Sassy but not rude
  • Loyal but not fawning
  • Honest but not judgemental
  • Rebellious but not disgruntled
  • Curious but not pushy
  • Upbeat but not unrealistic
  • Sincere but not boring
  • Compassionate but not maternal
  • Youthful but not childish
  • Smart but not clinical
  • Informative but not formal

Voice.

Since Vivibot is a digital chatbot, users experience her personality almost entirely through her written texts. GIFs, memes, and emoji can add flair, but her words had to deliver meaning and be recognizably Vivibot.

Users told us they’d heard enough clinical speak during cancer treatment. They wanted Vivibot to mimic how they would interact in online conversations with a slightly older, or at least a more experienced, AYA friend. Vivibot is not a therapist or a doctor, nor a replacement for one. She never sounds like a parent, teacher, coworker, or a boss.

In every interaction, Vivibot’s shows she cares. She proves she is empathetic and understanding by being curious, acknowledging pain (but not overdoing it), and by not constantly trying to make users see the bright side. Though Vivibot is a chatbot, her voice is not robotic. She speaks in a natural and friendly way, filled with emotional cues and interjections.

Tone.

In the course of even one day of conversation, Vivibot’s personality and voice has to stay consistent. However, her tone can and should shift—though not suddenly or too frequently—depending on the content she’s sharing and the user’s likely emotional state. She has to meet users where they are emotionally.

For example, when Vivibot asks the user about a potentially stressful or important personal experience, Vivibot’s tone is caring and sincere. Toward the end of the day’s chat, Vivibot could shift to an appropriately sassier, humorous tone to lighten the mood and match the emotional lift the user might be feeling because of the conversation.

Diction & Language.

Users responded positively when Vivibot used language that young people use to talk to each other—but only up to a point. Being clear, direct, and casual was good. But when Vivibot slung too much slang, users would tell us she was trying too hard.

Vivibot gets her point across clearly by choosing language that is accurate and specific, but also familiar and interesting. Vivibot chooses everyday, natural language that someone who isn’t a scientist, doctor, or psychologist would use—also called plain language. But, she doesn’t say things so simply that she comes across as condescending. Using plain language makes it easier for people to read, understand, and make informed decisions about dealing with life beyond cancer.

Humor

When I joined the project, the team was of two minds about whether Vivibot should be funny. Not everyone finds humor helpful, and not everyone laughs at the same things. Humor during emotional moments and especially during cancer conversations requires a high level of awareness and sensitivity. Doing humor the wrong way could be worse than no humor at all.

I looked to the evidence. Limited but growing research has shown that people dealing with cancer treatment, past or present, often choose humor and laughter. They’ll turn to it as a coping mechanism, for relaxation, or to improve their mood. Humor has also been shown to reduce isolation, anxiety, depression, and stress, and improve overall health and wellness.

Evidence in hand, I brought more humor to Vivibot with a few caveats: Vivibot is never funny for the sake of funny. Vivibot uses humor at the right time in the right way. Vivibot never ever makes jokes about the user. Vivibot also does not tell jokes that make her seem uncaring, dismissive, or judgemental, or that trivialize a person’s feelings or experiences.

 

The language of cancer

Many adolescents and young adults do not like to be labeled at all, much less labeled by an illness or diagnosis. No matter their age, cancer does not define anyone.

Person-first language is a way to choose more inclusive alternatives that describe people’s situation, rather than impose an identity. Shifting Vivibot to person-first language had a major effect on how she talks about all things cancer-related.

The trouble with “survivor.”

The biggest challenge: not everyone thinks of themselves as a “survivor.” It may be something others have called them, but not a label with which they personally identify. Some may dislike the term, or simply prefer a different term.

Survivor also has no universal meaning. Some define it as a person who has finished cancer treatment. Others, including the National Cancer Institute and the American Cancer Society, define it as a person from the time of diagnosis through the rest of their life. Complicating things further, the National Coalition for Cancer Survivorship “extended” its definition to identify “family, friends and caregivers” as survivors as well.

The language used to describe people living with and beyond cancer is fraught with pitfalls, disagreement, and cliches. To be clear, I’m not arguing for the eradication of the term “cancer survivor”—there’s some research that indicates forming empowering identities like “survivor” can be helpful to post-cancer adjustment in young people. But the cloudy meaning didn’t work for us.

We decided Vivibot would fully embrace person-first language to avoid labeling anyone by their illnesses or disabilities. Instead of “cancer survivor,” Vivibot would usually (though not always) say something like “a person who has had cancer,” or “a person dealing with life after cancer.” Or Vivibot could simply say “you.”

 

Insights for intervention designers and UX writers

The process of developing Vivibot challenged some of our assumptions and affirmed others. Here are just a few lessons learned that others can apply to creating conversational interfaces and other supportive interventions for AYAs.

Create alongside users.

Create alongside users. One non-negotiable for Hopelab, always, is to co-design with the young people the intervention aims to support. For the Vivibot project, that meant collaborating with AYAs from day one as advisors, focus group participants, and beta testers. All the pieces that became part of Vivibot started with our users.

Test test test.

User testing and data analysis is another non-negotiable in any Hopelab project. This gives Hopelab abundant opportunities to skip the guesswork and move quickly toward verifiable results. Every test we performed gave us an answer we needed, and kicked off a new set of ideas for ways to improve the Vivibot user experience.

Give users control.

Over and over, users told us they wanted more control over the conversation—or at least the feeling of having more control. In the chatbot format, this came down to giving users more choices. Sometimes it was about having more varied response buttons in the conversation so users could pick one that matched their feelings. Other times it meant letting users skip parts of the conversation if they weren’t feeling up to it.

Personalize it.

Users reacted negatively to any repetition that made talking with Vivibot feel like a chore. Sure, reusing content would have been a time saver and much easier to code, but it would come at the cost of user engagement. Even though users knew they were talking to a pre-programmed chatbot, we still had to make it feel like an authentic one-on-one conversation.

Be a fierce advocate for users.

A cohort of AYA cancer survivors and psychology advisors helped shape and sometimes even write a prototype version of Vivibot. As beta testing began, Hopelab realized it needed a professional UX writer who could transform the voice and tone and add subtle humor, all while making the language more consistent, inclusive, and welcoming—even if that writer (me) wasn’t the same age as target users.

To experience Vivibot yourself, check it out. If you’d like to discuss Vivibot or other innovative health interventions, get in touch with Hopelab at hi@hopelab.org.

Lauren Girardin is a communications consultant and freelance writer based in San Francisco. She helps organizations engage their communities and tell their stories. Her website is laurengirardin.com and you can connect with her on Twitter at @girardinl.


Related Content

View all Insights
Young, Wild, and Quarantined
Mental Health

While technology has often been criticized for provoking our separation and loneliness, it may be a critical part of the solution

Read More
Making a Mental Health Startup? How a Professional Can Help
Mental Health

Hopelab joined HackMentalHealth, Vinaj Ventures, and Second Muse to host a mentoring event for early-stage startups in the digital mental health space.

Read More
JMIR Randomized Controlled Feasibility Trial on Vivibot
Chatbots

This pilot randomized controlled trial examined the feasibility of delivering positive psychology skills via the Vivibot chatbot and its effects on key psychosocial well-being outcomes in young adults treated for cancer.

Read More
Reverse Engineering Health
Behavior Change

From video games to apps to chatbots, Hopelab has created a broad range of science-based interactive technologies by reverse engineering health. We’re moving the needle on health behavior and moving the needle on outcomes.

Read More