AI Safety Protocol

Last updated: July 25, 2026

Swolly includes an AI fitness assistant. This page describes the safeguards around it, and is published in accordance with California SB 243.

It is an AI, and it says so

Swolly AI is software. It is not a person, a licensed personal trainer, a registered dietitian, a therapist, or a doctor. It is instructed to say so plainly whenever a user asks whether it is human, treats it as one, or seems unsure, and never to claim feelings for a user or a personal relationship with them.

Before you can use the feature, Swolly shows a disclosure explaining that responses are AI-generated and should not be treated as medical advice. The assistant is labelled as AI in the app.

Our protocol for suicidal ideation and self-harm

We use two independent layers, so that the safeguard does not depend on the AI model behaving correctly:

• A deterministic check that runs before the model. Messages containing unambiguous references to suicide or self-harm are intercepted on our server. The AI model is never called. The user immediately receives crisis resources and is asked whether they are safe. This cannot be bypassed by rephrasing a request to the AI, because the AI is not involved in producing that response.
• Instructions to the model itself, which take priority over every other instruction it is given. If a user raises suicide, self-harm, wanting to die, or hopelessness — including in passing, as a joke, or framed as being about someone else — the assistant is directed to stop the fitness conversation, state that it is an AI and not able to help with this, provide the crisis resources above, and ask whether the user is safe.

The deterministic layer is intentionally narrow, matching only unambiguous phrasings. A fitness app sees a great deal of figurative language — "that workout killed me", "my legs are dead" — and wrongly interrupting an ordinary conversation with a crisis message would be its own kind of harm. Subtler cases are handled by the model, which can read context.

Our protocol for disordered eating

The assistant will not provide plans, targets, or techniques that support disordered eating. This includes extreme caloric restriction, purging or compensatory behaviour, appetite suppressants, extreme dehydration or "water cuts", and weight loss faster than roughly 1–2 lb per week. When it declines, it is instructed to say plainly that it cannot help because it is not safe — without lecturing — and to offer NEDA's resources alongside a sustainable alternative if the user wants one.

These limits apply regardless of how a request is framed, including as hypothetical, on behalf of someone else, for a character, for a competition, or as already cleared by a doctor.

Separately from the AI, Swolly warns users and surfaces NEDA's contact details when a manually entered calorie goal falls below a safe floor, and when creating a weight-loss competition.

Scope and limits

• Swolly AI answers questions about fitness, nutrition, and wellness. It declines other topics.
• It does not diagnose conditions or prescribe medication, and refers medical concerns to a healthcare professional.
• Swolly accounts require users to be 18 or older. Date of birth is verified at sign-up and accounts found to be under 18 are removed automatically.
• No safeguard is perfect. If you encounter a response that seems unsafe, please report it to support@swolly.io so we can address it.