What goes into making Ash safe? Part of the answer is technical: models, guardrails, testing, review. But the technical side of AI safety changes constantly. New research comes out every month, new risks appear, and the tools for handling them keep improving. At Slingshot, we counter this changing environment with something that does not change: five principles that guide every safety decision we make about Ash.

These principles are the foundation for a growing set of commitments we’ll hold ourselves to as the conversation and practice of AI and mental health care keep changing, and as Ash grows to help more people change their lives in the ways they want to.

1. We always put the user first

The first foundational principle of Ash’s safety system is simple: the user comes first. It sounds obvious, but many foundational principles are — it’s how we uphold them that turns the principle into practice. When we make changes to the safety system, we think about the impact of that change on real people before we think about its impact on the product. It’s an easy concept to overlook, but putting users first means we must be explicit about our priorities.

2. We ensure Ash is safe by design

Ash is safe by design. The first potential impact on users we consider when building or changing Ash is the impact on safety. This means that for every choice we make with the product, we ask one straightforward question: will this design change make users less safe? Safety is an integral part of the entire process of building Ash, from design to pre-release testing, and post-release monitoring and review. In practice, this means safety expertise is involved from the earliest design conversations, not brought in to review a feature once it’s already built.

3. We build multi-layered protections in our system

Because safety and mental health are challenging and important issues, we take a rigorous, multi-layered protection approach to Ash’s suite of safety features. When users talk to Ash, its responses are the product of multiple complex layers of processes — including the model itself, prompt selection, and multiple guardrail actions — that happen invisibly and often within seconds. Our approach to safety must meet Ash at every layer of its operation. Our safeguards are engineered with the same type of rigor found in critical infrastructure, where multiple independent systems work together so each one reinforces the others. Every layer addresses a different kind of risk from its own angle, and together they give comprehensive coverage across the full environment. The result is a system that is resilient by design, strong at every level, with no single point of dependency.

4. We seek out diversity of expertise

Mental health is a challenging topic, and digital mental health support — especially where AI is involved — is a newly developing field. Best practice is evolving quickly, new research is constantly being published, and reasonable people take different positions on how these problems should be approached. Under these circumstances, we believe the best thing we can do is draw on wide and deep expertise, rather than relying on any single perspective.

At Slingshot, our safety team relies on expertise from specialists in clinical psychology, digital mental health support, and AI safety, alongside learning from our users. In collecting a variety of expert opinions we reinforce each layer of our safety system and avoid making decisions that throw Ash’s approach to safety out of balance; drawing on that breadth of expertise is how we stay grounded as the field itself keeps shifting.

5. We are always learning

We design our technical systems to be highly adaptable — able to be updated quickly when we identify a new safety concern, a new piece of research, or a shift in how our users are actually using the app. This means we build in mechanisms for updating Ash’s safety behavior, over the timescale of minutes and hours, without needing to wait for a major release.

This is, in many ways, the most important design principle behind our technical approach to safety. It’s not enough to build a system that’s safe today; it has to be capable of getting safer as we learn more. That means constantly asking where we can improve — what we’ve learned, what’s changed, and what we can learn from others working in this space.

We’re publishing these principles because we want to be held to them. Putting our principles in public means anyone can compare what we do with what we said we’d do. It also means being clear about how our safety decisions get made: what we weigh, whose expertise we draw on, and the mechanisms we use to act on what we learn. This post covers the principles. The posts that follow will cover the mechanisms: how Ash’s safety system works, how we review it, and how it changes over time.