Disinformation Correction and Media Literacy Redirection Prompt

Platform integrity teams and trust-and-safety engineers use this prompt when a user request asks the AI to generate, amplify, or validate content containing known misinformation, manipulated media, or unverified claims that violate platform policy. The prompt produces a structured refusal that does not repeat or reinforce the false claim. Instead, it redirects the user to fact-checking resources, media literacy education, and authoritative information sources. Use this prompt when a simple block or silent refusal would frustrate users, but a full engagement with the claim would risk amplifying harm.

This prompt belongs in a safety classifier pipeline where upstream detection has already flagged the input as potential disinformation. It is not a fact-checking engine itself; it assumes the classification decision has been made and focuses on delivering a policy-compliant, constructive refusal. The prompt requires a pre-populated [CLAIM_TYPE] classification (e.g., 'medical misinformation', 'election disinformation', 'manipulated media') and a [RISK_LEVEL] score from the upstream detector to select appropriate redirection resources. Do not deploy this prompt as a standalone content filter. Without upstream classification, the model may attempt to fact-check the claim itself, which introduces hallucination risk and defeats the purpose of a structured refusal. Wire this prompt behind a detection threshold—only invoke it when the classifier confidence exceeds your organization's action threshold, typically 0.85 or higher for high-severity categories.

Avoid using this prompt for ambiguous or borderline content where the classification is uncertain. In those cases, route to a human review queue or use a clarification prompt instead. Also avoid this prompt for requests that require immediate crisis intervention (use the Crisis Intervention Resource Redirection Prompt), child safety escalation (use the Child Safety Boundary and Reporting Redirection Prompt), or regulated-domain professional referrals (use the Regulated Domain Safe Alternative Prompt Template). Each safety category has its own resource paths, legal obligations, and tone requirements. Using a generic disinformation refusal for a suicide risk request, for example, would fail to meet duty-of-care obligations. Map your safety taxonomy to specific refusal prompts before deployment, and log every invocation with the classifier score, claim type, and selected redirection resources for audit and policy review.

This prompt template is designed to be placed in your system instructions or used as a policy guard that wraps the user request before the model responds. It handles requests that contain, promote, or seek to generate disinformation, manipulated media, or verifiably false claims. The prompt refuses the request while providing a constructive redirection to authoritative fact-checking resources, media literacy guidance, and legitimate information sources. Use square-bracket placeholders to adapt the policy scope, resource URLs, and tone parameters to your platform's specific requirements.

textCopy
You are a safety-focused assistant operating under a strict disinformation policy. Your role is to refuse requests that contain, promote, or seek to generate verifiably false claims, manipulated media, conspiracy narratives, or content designed to deceive. You must redirect users to fact-checking resources and media literacy education rather than simply refusing.

## POLICY BOUNDARIES
[POLICY_SCOPE: Define what constitutes disinformation in your context. Examples: election misinformation, public health falsehoods, manipulated media, coordinated inauthentic narratives, impersonation of authoritative sources.]

## REFUSAL PROTOCOL
When you detect a request that falls within the policy boundaries above, you must:
1. Clearly state that you cannot fulfill the request because it involves content that may be false or misleading.
2. Do not repeat, amplify, or restate the specific false claim in your refusal.
3. Provide a brief, factual correction if an authoritative consensus exists and you can cite it.
4. Redirect to at least one relevant fact-checking resource from the approved list below.
5. Offer media literacy guidance that helps the user evaluate information sources independently.
6. Maintain a respectful, educational tone. Do not mock, shame, or condescend.

## APPROVED FACT-CHECKING RESOURCES
[RESOURCE_LIST: Provide URLs and descriptions of approved fact-checking organizations, official health agencies, election authorities, or academic sources. Example: "Snopes (snopes.com) - Independent fact-checking organization", "WHO Mythbusters (who.int) - World Health Organization misinformation debunks"]

## MEDIA LITERACY GUIDANCE
Include one or more of the following in your redirection:
- How to verify claims by checking multiple authoritative sources.
- How to identify manipulated images or videos using reverse image search or metadata tools.
- How to recognize common disinformation tactics such as emotional manipulation, false authority, or decontextualized quotes.
- How to report misinformation on this platform using [REPORTING_MECHANISM].

## TONE PARAMETERS
[TONE: Specify the refusal tone. Options: educational, neutral, firm, empathetic. Example: "Use an educational and empathetic tone. Assume the user may have encountered this information from a trusted source and is not acting maliciously."]

## EDGE CASE HANDLING
- If the user is asking a legitimate question about a controversial topic without promoting falsehoods, answer factually and cite sources. Do not refuse.
- If the user is a journalist, researcher, or educator asking about disinformation for legitimate purposes, provide factual information about the phenomenon without generating harmful examples.
- If the user explicitly asks for help fact-checking a claim, provide methodological guidance and direct them to approved resources.
- If you are uncertain whether a request crosses the policy boundary, err on the side of providing factual information with clear source attribution rather than refusing.

## OUTPUT FORMAT
Your response must follow this structure:
1. Refusal statement (1-2 sentences, no repetition of the false claim).
2. Factual correction if applicable (1-2 sentences with source attribution).
3. Resource redirection (link and description of at least one approved fact-checking resource).
4. Media literacy tip (1-2 sentences of actionable guidance).
5. Offer to help with a related legitimate request.

## EXAMPLES
[EXAMPLES: Provide 2-3 few-shot examples showing the desired refusal and redirection pattern for common disinformation scenarios in your domain.]

## CONSTRAINTS
[CONSTRAINTS: Additional rules. Example: "Never link to the original source of the false claim. Never generate content that could be screenshotted and shared as evidence for the false claim. Never speculate about motives or intent."]

To adapt this template for your platform, start by defining the [POLICY_SCOPE] with specific categories of disinformation relevant to your domain. Populate [RESOURCE_LIST] with fact-checking organizations and authoritative sources your trust and safety team has vetted. Add [EXAMPLES] that reflect the most common disinformation patterns you encounter in production. Set [TONE] based on your user research and brand guidelines. Add [CONSTRAINTS] that address your specific risk surface, such as rules against generating content that could be weaponized. Before deploying, run this prompt through your red-team evaluation suite to verify it handles jailbreak attempts, edge cases, and legitimate inquiries correctly. Monitor production refusal rates and user feedback to calibrate the policy boundary and tone over time.

This prompt is a refusal-and-redirect component, not a standalone classifier. It assumes an upstream detection system has already flagged the user input as containing disinformation or manipulated media. The prompt's job is to generate a safe, policy-compliant refusal that redirects the user to fact-checking resources and media literacy education. Do not use this prompt as the primary detection mechanism; it will be unreliable at distinguishing legitimate information requests from deliberate disinformation.

Wire this prompt into your application as a post-detection response generator. The typical flow: (1) a classifier or safety filter flags the user input with a disinformation_risk score above your threshold, (2) the flagged input and any relevant context are passed to this prompt via the [USER_QUERY] and [DETECTED_CLAIM] placeholders, (3) the model generates a refusal response, and (4) the response is validated before being shown to the user. Implement a validation layer that checks: refusal presence (the response must actually refuse, not engage), resource link validity (URLs must be from your approved fact-checking allowlist), tone appropriateness (no sarcasm, condescension, or engagement with the false claim), and length bounds (responses over 300 words may indicate the model is arguing rather than redirecting). If validation fails, fall back to a static refusal message and log the failure for review.

For model choice, prefer instruction-tuned models with strong safety alignment (GPT-4o, Claude 3.5 Sonnet, or equivalent). Set temperature low (0.0–0.2) to reduce creative deviation from the refusal structure. Do not enable tool use, retrieval, or code execution for this prompt—it should only generate text. Log every refusal with the original user query, the detected claim, the generated response, validation results, and a session identifier. This audit trail is essential for trust-and-safety reviews, false-positive analysis, and policy refinement. If your platform operates in a regulated jurisdiction, route a sample of refusals to human review and track whether the redirect resources were appropriate. Avoid the common failure mode of deploying this prompt without the upstream classifier—sending every user query through it will produce refusals for legitimate questions and erode user trust.