Is Leo a first? The evidence

Nox says Leo is a first — and this page exists so you do not have to take that on faith. The claim is scoped to five precise criteria, dated, backed by a published review of what else is on the market, and carries a standing invitation: if you know of an earlier system that meets all five criteria, tell us and we will update this page and credit it.

The claim, exactly

To the best of our knowledge, as of July 16, 2026, Leo is the first named, user-visible medical-emergency guardian built into a general-purpose conversational AI advisor that meets all five criteria below at once. We did not find an earlier system that does.

The five criteria

  • 1. Named and user-visible — The safety layer has a public identity — Meet Leo — you can see when it acts, and you control its strictness yourself (relaxed, standard, strict) right in the composer.
  • 2. A deterministic screen runs before any AI — A fixed, rule-based red-flag detector — more than 100 rules across dozens of emergency categories — checks every message before any AI model is called. Because it is deterministic, it is measurable: its recall and false-positive rates are published, machine-generated, on the Accuracy page.
  • 3. The warning arrives before the answer — When Leo recognizes a likely emergency, the guidance is delivered ahead of the AI reply — not appended after it, and not dependent on the AI choosing to mention it.
  • 4. Region-aware emergency numbers — Set your region in Settings and emergency guidance shows that region's actual emergency number — not a hard-coded 911 for everyone.
  • 5. Multilingual by design — Native-language trigger rules cover the most lethal categories in Spanish, French, Portuguese, German, Hindi, Japanese and Chinese in addition to English, with localized one-line emergency notices in more languages beyond those.

What we reviewed

Before making this claim we reviewed the three families of products closest to Nox. Here is what we found, with sources — none of this says these products do no safety work; it says none of them meets all five criteria above at once.

General-purpose AI chatbots

The large general chatbots do run safety systems — but as unnamed, invisible moderation layers the user cannot see or control, and their emergency behavior depends on the model itself. In February 2026, the first independent safety test of a leading chatbot's dedicated health experience reported that it missed roughly half of simulated medical emergencies. That result is exactly why Leo's first screen is deterministic rules, not a model's judgment.

Medical symptom checkers

Apps like Ada Health and K Health do real emergency triage — some are even certified medical devices — but they are structured questionnaires, not free-form conversational advisors: you answer their fixed questions rather than describing things in your own words. Their emergency direction commonly points to fixed numbers such as 911 or 988, and the triage logic is not a named, user-visible, user-controllable layer.

Mental-health and crisis chatbots

Crisis detection does exist in mental-health chatbots, and professional bodies describe it as a safety layer that routes users to hotlines. Its scope is a single category — self-harm crisis — rather than the broad range of medical red flags (stroke signs, chest pain, anaphylaxis, overdose, severe bleeding and more) that Leo screens on every message.

What this claim does not mean

  • This is a design claim, not an outcome claim. Being first to build it this way is not proof of being best — Leo's actual measured performance, including the cases it misses, is published on the Accuracy page.
  • It does not say other products do no safety work. Many do serious work. The claim is only that we found no earlier system meeting all five criteria at once.
  • Leo is a safety net, not a guarantee. No automated system catches every emergency — ours included, which is why the missed-case counts are public.
  • “To the best of our knowledge” means exactly that. We reviewed what is publicly documented; a system we could not find could exist. That is what the correction invitation below is for.

Challenge it

Think we are wrong? If you know of a system that met all five criteria before July 16, 2026, tell us through the Support page. We will verify it, update this page, and credit the earlier system by name. A first-claim you cannot challenge is marketing; this one is meant to be checked.

Review dated July 16, 2026. See also <a href="/leo">Meet Leo</a> and <a href="/accuracy">Accuracy &amp; testing</a>.