How the Trevor Project Is Using AI to Train Crisis Counselors For Teens

For the Trevor Project, a nonprofit dedicated to preventing suicides among LGBTQ+ youth, efficiently preparing volunteer counselors for the unpredictable twists and turns they might encounter in the course of their field work can be a constant challenge. That's why the group has turned to one of the most sophisticated AI language generators in the world to create a hotline worker training simulator versatile enough to mimic the experience of texting with a teen in crisis. The result is a chatbot called the Crisis Contact Simulator, or "Riley," a version of research group OpenAI's GPT-2 system that was trained on mock transcripts of conversations between counselors and at-risk teens. The Trevor Project developed the program in partnership with Google's charitable arm, Google.org. The nonprofit has begun using Riley as a way for trainees joining its TrevorText and TrevorChat text-based hotlines to perform role-play exercises without the scripted feel of more cut-and-dry training activities or the stakes of a real-life emergency situation. Riley is advanced enough to hold a coherent and far-ranging dialogue in the speaking style of a real-life teen with enough randomness of output to keep trainees on their toes. "Our Crisis Contact Simulator can engage in a prolonged back-and-forth dialogue with trainees and can use language in the same way people do, including language LGBTQ+ youth often use to describe their experiences and emotions," said Dan Fichter, head of AI and engineering at the Trevor Project. "The simulator maintains a consistent emotional and experiential narrative in talking about real-life feelings and situations." Unpredictability as a feature, not a bug While GPT-2—and its even more powerful successor, GPT-3—are able to generate text so passably realistic they have stirred fears around fake news production, they are still considered too unpredictable in content for many large-scale commercial applications, like customer-facing chatbots. Their utility so far has been largely confined to novelty uses, like text adventure games, writing aids and gag Twitter accounts. But the Trevor Project's chatbot demonstrates a novel case in which GPT-2's unpredictability actually works in the application's favor. "In trainings, it's really important that basically every path [of dialogue] leads somewhere," Fichter said. "And so an AI-based model that is able to roll with whatever questions that counselor asks as a trainee is really important to us." Fichter said the group delved into a wealth of research being conducted on AI in mental health capacities throughout the development process and tested several different natural language systems before deciding on GPT-2 for its versatility. The team also drew on the technical resources of Google.org's fellowship program, which places Google staff volunteers at nonprofits to guide tech-based projects. Based on that groundwork research, Fichter said he is fairly confident the project is the first instance of an AI being used this way in the nonprofit mental-health space at this scale. "To our knowledge, this is the world's first instance and a project that would not have been necessarily technologically achievable until very recently," Fichter said. Scaling up operations The persona of Riley is an archetype the Trevor Project has long used in its training exercises—a fictional LGBTQ+ young person "from a particular part of the country presented with a particular set of emotional circumstances and a particular reason for reaching out," said Fichter. Therefore, the team had abundant material on hand with which to train the AI to its particular purposes. In addition to bringing these materials to life, the Crisis Contact Simulator also allows the organization to train more volunteers at a time on their own schedules. Trainees simply conduct exercises within the bot and receive feedback on how they handled each interaction through a transcript that is generated from each session. This efficiency will be important as the Trevor Project aims to triple the size of its volunteer responder base this year and hold around 5,000 hours of virtual trainings in the process. Fichter is also in the process of examining other ways in which the group can use AI to streamline that ambitious effort. "Unfortunately, just an enormous number of young people need our services, and therefore, our crisis team is growing very rapidly," Fichter said. "There are probably a lot of ways over the next few months that we will start to apply different AI techniques to improve the program."