VivaMed Refusal Gate

The safety problem

Health misinformation in short video causes real harm: people skip treatment, take unsafe "detoxes" or fear safe habits. The user is anyone who scrolls health Shorts. The failure mode of AI health content is worse: a fluent model will confidently say things no study supports.

What it does

VivaMed's Refusal Gate is a safety layer for AI-written health content: no source, no claim.

• Before a script is drafted, a relevance gate checks that at least two real PubMed abstracts actually study the question. If not, it refuses to write anything.
• Every claim must cite one of those abstracts; citations to papers not in the evidence are rejected automatically.
• A second model checks each claim against its source; a final reviewer approves or rejects. Unsafe, overstated or unsourced scripts are rejected with a reason, and that record is kept.
• Every published Short links the real studies, and says "educational only, not medical advice".

What existed before, and what is new in this event

• Before: the PubMed-grounded writer and renderer.
• New during SafeSpace Hacks: the two-tier review (cheap models check; a stronger model makes the final call) and a full audit trail of every approval and rejection.

Trust and responsible design

No user data is collected. Uncertainty is stated, not hidden: when the evidence is mixed, the script says so. Honest limits: the reviewers are models plus a medical student, not a practising clinician; a bad paper can still pass the gate.

How we built it

Python, NCBI E-utilities for PubMed, free language models through a local router, local text-to-speech and rendering.

AI disclosure

Scripts are drafted and checked by AI models from real abstracts; the code was written with an AI coding assistant.

Built With

• automation
• llm
• misinformation
• pubmed
• python
• safety