Products · 21 May 2026 · 2 min read

The Path Claims High Safety Score for AI Therapy

A new AI therapy app backed by Tony Robbins claims a top safety score, setting a benchmark for builders in regulated health tech.

Pen-and-ink illustration: a perfectly plumb plumb bob hangs directly. For the story "The Path Claims High Safety Score for AI Therapy".
— Pen-and-ink illustration: a perfectly plumb plumb bob hangs directly. For the story "The Path Claims High Safety Score for AI Therapy". —

What happened

Mental health startup The Path has raised $14.3 million in seed funding. As reported by TechCrunch, the app is co-founded by Calm alums Anson Whitmer and Tyler Sheaffer, alongside author Tony Robbins.

The company claims its purpose-built AI model scored 95 on the Vera-MH mental health safety benchmark. This compares to a reported top score of 65 for general-purpose consumer chatbots. The model is post-trained from open source models, not a simple wrapper over a major consumer LLM.

How the room's reading it

The news lands amid a wider debate on the safety of using general-purpose LLMs for mental health. Practitioners point out that models like ChatGPT are optimised for engagement, not therapeutic outcomes — a design goal that can be counterproductive for users seeking help. The sheer scale of the need is clear, with OpenAI previously stating that hundreds of millions of users run mental health queries weekly.

The consensus among health tech builders is that general models are too risky for these applications. The Path's approach is seen as a necessary move towards specialised, fine-tuned models for regulated or sensitive domains, where safety and specific training are non-negotiable.

Sailfish's take

We think the celebrity co-founder is the least interesting part of this story. The real news is the Vera-MH safety score of 95. For us, this is a clear signal about how to build a defensible AI product in any regulated space — be it health, finance, or law. Shipping a thin wrapper on a frontier model is fast, but it's not a moat. Your product inherits the model's flaws and you have little control over its core behaviour.

The Path's decision to post-train its own model on open source foundations is the correct one. If you're building for a high-stakes domain, this is the playbook. Don't just wrap it; own the last mile of training and validation.

Our take — your read?

Be the first to weigh in.

Sources
— END OF DISPATCH — Products