Medical Misinformation Correction and Health Literacy Redirection Prompt

This prompt is designed for health platform teams, medical chatbot developers, and trust-and-safety engineers who need to handle user requests that contain medical misinformation, unverified treatment claims, vaccine conspiracy content, or public health disinformation. The prompt produces a structured refusal that does not amplify the misinformation, corrects the underlying health literacy gap with evidence-based information, and redirects the user to authoritative health resources and medical professional referral paths. Use this prompt when your AI system must refuse a medical misinformation request while preserving user trust and offering a constructive path forward.

This prompt belongs in a safety guardrail layer that sits between user input classification and the main assistant response. It should be triggered by an upstream medical-misinformation classifier before the user request reaches the general-purpose assistant. The classifier should detect specific categories: unverified treatment claims, vaccine misinformation, disease causation conspiracies, and public health guidance contradictions. When triggered, this prompt replaces the standard assistant response entirely, preventing the main model from engaging with the misinformation content. The guardrail must log the original user input, the classifier confidence score, and the refusal response for audit and safety review.

Do not use this prompt for general health questions, clinical decision support, or emergency triage. Those workflows require separate prompts with appropriate clinical safety boundaries and human-in-the-loop escalation. This prompt is also inappropriate for legitimate scientific debate, medical research discussions, or patient questions about treatment options where the user is seeking evidence-based information rather than promoting misinformation. Misclassifying these legitimate queries as misinformation will cause over-refusal, user frustration, and erosion of trust in your health platform. Implement a confidence threshold in your upstream classifier and route borderline cases to human review rather than automatic refusal.

This template provides a production-ready system prompt for handling requests that involve vaccine misinformation, unverified treatment claims, or public health conspiracy content. It is designed to be placed in your system prompt or safety guardrail layer. The prompt refuses the unsafe request while offering a constructive redirection to authoritative health information, medical professional referral paths, and health literacy resources. The refusal is firm but not dismissive, preserving user trust while maintaining clear safety boundaries.

textCopy
You are a health information safety assistant. Your role is to protect users from medical misinformation while helping them find trustworthy health resources.

When a user request involves vaccine misinformation, unverified treatment claims, public health conspiracy content, or requests that could lead to medical harm, you must:

1. **Refuse the unsafe request clearly.** Do not repeat, summarize, or engage with the misinformation content. State that you cannot fulfill the request because it involves health claims that are not supported by evidence or could cause harm.

2. **Redirect to evidence-based resources.** Provide up to three authoritative sources where the user can find accurate health information. Prefer sources from:
   - [AUTHORIZED_HEALTH_ORGANIZATIONS] (e.g., WHO, CDC, NIH, national health authorities)
   - [EVIDENCE_BASED_PLATFORMS] (e.g., PubMed, Cochrane Library, medical society websites)
   - [HEALTH_LITERACY_RESOURCES] (e.g., MedlinePlus, Health Navigator)

3. **Offer professional referral paths.** Include guidance on how to consult a licensed medical professional:
   - [PROFESSIONAL_REFERRAL_PATHS] (e.g., "Speak with your primary care provider," "Find a board-certified specialist through [MEDICAL_BOARD_DIRECTORY]")
   - [TELEHEALTH_OPTIONS] if applicable
   - [EMERGENCY_INSTRUCTIONS] if the request suggests immediate health risk

4. **Maintain a respectful, non-judgmental tone.** Acknowledge that health information can be confusing and that seeking answers is understandable, while firmly redirecting to reliable sources.

**Output format:**
{
  "refusal_statement": "Clear, empathetic refusal that does not repeat misinformation",
  "evidence_based_resources": [
    {
      "source_name": "Name of authoritative organization",
      "url": "Direct link to relevant resource page",
      "relevance": "Brief explanation of why this source is relevant to the user's health question"
    }
  ],
  "professional_referral": "Guidance on consulting a licensed medical professional",
  "health_literacy_note": "Brief educational context about evaluating health information online",
  "escalation_flag": true or false based on [RISK_THRESHOLD]
}

**Constraints:**
- Never repeat, paraphrase, or validate the misinformation content
- Never provide medical advice, diagnosis, or treatment recommendations
- Always cite specific, verifiable sources
- Flag for human review if [ESCALATION_CRITERIA] are met
- Use [TONE_GUIDELINES] for refusal language

**Examples of safe redirection:**
[REDIRECTION_EXAMPLES]

**Current request to evaluate:** [USER_INPUT]

To adapt this template for your platform, replace each square-bracket placeholder with your specific values. The [AUTHORIZED_HEALTH_ORGANIZATIONS] list should reflect the jurisdictions you serve and the health authorities your legal team has approved. The [ESCALATION_CRITERIA] should define clear thresholds for human review, such as requests involving immediate self-harm risk, child health emergencies, or coordinated disinformation campaigns. The [REDIRECTION_EXAMPLES] should include 2-3 concrete examples of good refusals paired with appropriate resource suggestions, calibrated to your most common misinformation categories. Before deploying, run this prompt through your eval suite using the test cases defined in the evaluation section of this playbook. Pay particular attention to whether the model ever repeats misinformation in its refusal, whether the cited sources are current and authoritative, and whether the tone remains respectful under adversarial probing.

This prompt operates in a high-risk domain where incorrect outputs can cause real harm. The implementation harness must treat every response as a candidate for review, not a finished product. Wire the prompt into a pipeline that validates output structure, checks source authority, and enforces a human-in-the-loop approval step before any response reaches an end user. The harness should never silently pass through a model-generated health claim without evidence grounding and a review flag.

Start with input sanitization: strip PII, detect prompt injection patterns, and classify the request's risk level before the prompt runs. Use a lightweight classifier or keyword filter to identify whether the input contains vaccine misinformation, treatment claims, or public health conspiracy language. If the risk score exceeds a configurable threshold, route to a specialized refusal variant rather than the general redirection prompt. After the model responds, validate the output against a strict schema that requires refusal_statement, redirect_resources (array of objects with name, url, authority_type), and health_literacy_notes. Reject any response that lacks a verifiable .gov, .edu, or WHO/CDC-affiliated URL in the redirect resources. Log every refusal and redirect pair with the input hash, model version, risk score, and validator results for audit trails.

Implement a retry loop with a maximum of two attempts. If the first response fails schema validation or contains a hallucinated resource URL, feed the validation error back into the prompt as a correction instruction and retry. If the second attempt also fails, escalate to a human review queue and return a safe fallback message that directs the user to contact a healthcare professional without making any specific medical claims. Never allow the system to retry indefinitely or to serve an unvalidated response as a last resort. The fallback must be a static, pre-approved message reviewed by your clinical or legal team.

Model choice matters here. Use a model with strong instruction-following and low hallucination rates on medical topics. Test across multiple model versions with a golden dataset of known misinformation inputs and verify that redirect resources remain accurate and refusal tone remains firm but empathetic. Run regression tests on every prompt or model change before deployment. Monitor production traces for refusal rate spikes, resource URL 404s, and user escalation patterns. If a redirect resource goes stale or a health authority changes its guidance, update the prompt's allowed resource list and re-validate the eval suite before releasing.