Evaluating AI Safety
in Digital Mental Health
12 November 2026 · 17:00 CET
Everyone says their AI is safe. Almost nobody can show you how they know. This session digs into what real evaluation looks like before an AI mental health tool reaches a real user — what to test, what red-teaming actually catches, and where founders get it wrong. Bring your eval framework, your blind spots, your disagreements.
Hosted and facilitated by the nest team. Agenda below is a draft, shaped around what members are actually stuck on.
- What "safe" actually means — moving past the word to specific, testable claims.
- Red-teaming, for real — what it catches, what it misses, and how to start if you've never done it.
- Where evaluation breaks down — the gap between a clean eval set and a distressed real user.
- Founder blind spots — patterns that show up repeatedly across the products we see.
- Open conversation — bring a specific tool or decision you're stuck on.
Places are limited to keep the conversation genuinely useful. If you are building, funding or evaluating an AI tool that touches someone in a vulnerable moment, this is the room.
You're in.
Evaluating AI Safety in Digital Mental Health — 12 November, 17:00 CET, online.
A calendar invite with the link is on its way to your inbox.
This one looks at what real AI safety evaluation looks like before a tool reaches someone in a vulnerable moment — red-teaming, where founders get it wrong, and the gap between a clean eval set and a real user. Hosted by nest, 60 to 90 minutes, small group, cameras on if you're up for it. No pitching, no slides.
Bring a specific tool or decision you're stuck on.
One more step for nest notes — confirm your subscription here, it only takes a few seconds.
Know someone who'd want to be in this room? Forward this.
Want to co-host?
We're looking for someone with real AI safety, red-teaming or evaluation experience in digital mental health to help shape and run this one with us. Apply here.
Thank you.
We'll be in touch if it looks like a good fit for this session.