Research

Machine Learning-based Crisis Triage

Just presenting resources is the first step. Additional work is required to overcome other structural or attitudinal barriers to reaching out. We have developed methods to increase conversion to crisis resources by systematically helping users overcome common barriers and misconceptions. In a randomized controlled trial with Harvard, we showed our methods increased conversion rates by 23%.

Linguistic correlates of effective online peer support

Koko has millions of posts and responses from its peer support network. We analyzed linguistic patterns that predicted favorable outcomes.

Can we build an AI therapist? Can we build empathy from algorithms?

We created an artificially empathic chatbot using Koko data. We find that, while we can simulate empathic expression to a degree, there may be an upper bound machines can never cross. Unlike humans, they don't have lived experiences and so their efforts to empathize might always ring hollow.

Helping others on Koko predicts decreased symptoms of depression

Most looking to get help on Koko go on to help multiple other people. We make it simple and easy. We also find that helping others on Koko has a significant effect on mood regulation and depression symptoms.

Randomized controlled trial of the Koko platform

We randomly assigned people to use Koko (then called 'Panoply') or a matched control task (expressive writing). We found improvements for both, but the differences were significantly greater for the Koko platform.