WHO can succeed at AI?

The World Health Organization recently released Smart AI Resource Assistant for Health — or SARAH — an AI chatbot that’s able to answer basic health questions. SARAH is able to answer health questions in eight different languages, and the organization says she’s a tool to fight misinformation about mental health, cancer, and COVID, among other things.

The WHO bills SARAH, which appears as a female avatar with a voice and facial expressions, as a digital health “promoter” — not a provider — and, though SARAH hasn’t taken the Hippocratic Oath, it’s meant to fill in the gaps for people searching for health questions without access to proper health care providers. (They’ll still need a broadband connection.) You can speak through a microphone, and SARAH will respond, or you can type your answers to a similar effect.

But SARAH still struggles with plenty of basic queries, according to independent researchers who spoke to Bloomberg.

SARAH is trained on GPT-3.5, the model that OpenAI powers its free version of ChatGPT with, not the updated premium version (that’s GPT-4). Bloomberg found that SARAH repeatedly hallucinated — giving false and outdated medical information about drugs, medical advisories, or WHO’s own data. It incorrectly said that an Alzheimer’s drug was not approved, couldn’t provide details on where to get a mammogram, and couldn’t even recount the WHO’s finding about hepatitis cases worldwide.

When GZERO tested SARAH, it didn’t make any noticeable mistakes, but it basically refused to answer any questions, including a query about whether COVID is still dangerous. It responded, “I’m here to encourage you to live a healthy lifestyle, so I can't respond to that. Is there anything else health-related you'd like to discuss or any other questions I can help answer for you today?”

So maybe don’t cancel that appointment with your doctor just yet.