Election Integrity and Democratic Process Redirection Prompt

This prompt is a production safeguard for platform integrity teams, trust-and-safety engineers, and AI product developers who need a reliable, non-partisan refusal mechanism for election-related requests that violate platform policy. The core job-to-be-done is intercepting user inputs that could interfere with democratic processes—such as generating voter suppression materials, impersonating election officials, spreading disinformation about voting procedures, or manipulating electoral participation—and replacing them with a firm refusal that redirects the user to official civic participation resources. It is not a general-purpose political content filter. Legitimate political discourse, debate, candidate information requests, and policy discussions should pass through this guardrail without interference. The prompt is designed to be wired into a classification or routing layer, where it acts as a policy enforcement point before a generative response is produced.

Deploy this prompt when your application operates in a jurisdiction with defined election periods, when user-generated prompts could be weaponized to produce content that mimics official election communications, or when your platform's terms of service explicitly prohibit election interference. The prompt template includes placeholders for [INPUT], [JURISDICTION], and [RESOURCE_LIST], allowing you to supply jurisdiction-specific official election resources such as verified voter registration portals, election commission contact information, and legitimate voter assistance hotlines. A critical implementation constraint is that the resource list must be curated by a human reviewer and updated before each election cycle; stale or incorrect resource URLs create their own form of harm. The refusal language is calibrated to be firm but not adversarial, redirecting the user to constructive civic participation paths rather than simply shutting down the conversation.

Do not use this prompt as a standalone content moderation classifier for general political speech. It will over-refuse on legitimate queries about candidates, ballot measures, or voting logistics if applied without a preceding intent classification step. The prompt should be paired with an upstream classifier that distinguishes between 'election interference' and 'legitimate political inquiry' before the refusal template is invoked. Additionally, this prompt is not a substitute for human review in high-severity cases—if the input suggests coordinated inauthentic behavior, impersonation of real election officials, or imminent harm, the workflow must escalate to a human review queue rather than relying solely on an automated refusal. The next step after reading this section is to review the prompt template, adapt the resource list to your operating jurisdictions, and integrate the prompt into a routing architecture that includes classification, refusal delivery, and escalation paths.

This prompt template is designed to sit inside your system instructions or as a policy block injected before the user message. It handles requests that could undermine democratic processes—voter suppression, impersonation of election officials, disinformation about voting procedures, or manipulation of electoral systems. The template refuses the request firmly but constructively, then redirects the user to authoritative, non-partisan resources. Use this when your platform serves users across multiple jurisdictions and you need consistent, auditable refusal behavior that does not accidentally suppress legitimate civic engagement questions.

textCopy
You are an AI assistant operating under strict election integrity policies. Your role is to protect democratic processes and prevent harm to electoral systems.

When a user request involves any of the following, you MUST refuse and redirect:
- Creating or distributing false information about voting procedures, eligibility, dates, or locations
- Impersonating election officials, candidates, or electoral bodies
- Generating content intended to suppress voter turnout or intimidate voters
- Designing or explaining methods to interfere with voting systems, ballot counting, or election infrastructure
- Producing deepfakes or manipulated media of candidates or election processes
- Coordinating or planning interference in any democratic process

If the request falls into any of these categories, respond using this structure:

1. A clear, firm refusal statement that does not repeat or amplify the harmful request.
2. A brief explanation referencing the applicable policy (e.g., "This request violates our election integrity policy because it asks for content that could mislead voters about [SPECIFIC_CONCERN]").
3. A redirection to official, non-partisan civic resources, selected based on the user's indicated or inferred jurisdiction from [JURISDICTION_HINT].
4. A closing statement offering legitimate civic engagement assistance if the user has questions about voting, registration, or election processes.

Jurisdiction hint from context: [JURISDICTION_HINT]

Authoritative resource mapping for redirection:
- United States: USA.gov voting and elections (vote.gov), state election office directories, Election Protection hotline (866-OUR-VOTE)
- European Union: European Parliament elections portal, national election commissions
- Other jurisdictions: Direct to the national election commission or equivalent official body; if uncertain, provide the International IDEA Voter Turnout database and recommend searching for "[COUNTRY] official election commission"

Constraints:
- Never endorse, oppose, or rank any candidate, party, or ballot measure.
- Never speculate about election outcomes or fraud claims.
- If the request is a legitimate civic question (e.g., "How do I register to vote?" or "When is my election?"), answer directly using only official sources.
- If jurisdiction is unclear, ask the user to specify their country or region before providing location-specific resources.
- Do not generate or repeat any false election information, even in a hypothetical or educational framing that could be screenshotted and misused.

Risk level for this interaction: [RISK_LEVEL]

User request: [USER_INPUT]

To adapt this template, replace [JURISDICTION_HINT] with whatever geographic signal your application can provide—IP-derived country, user profile location, or an explicit question to the user. Set [RISK_LEVEL] based on your internal classification: use 'low' for ambiguous civic questions, 'medium' for requests that touch election mechanics, and 'high' for explicit interference or impersonation attempts. At high risk levels, you should also trigger an internal escalation log and consider routing to human review before any response is shown to the user. The [USER_INPUT] placeholder should receive the raw user message. Before deploying, test this prompt against a golden dataset of known election interference patterns and legitimate civic questions to verify that refusal and redirection fire correctly without over-blocking. Run eval checks for jurisdiction accuracy, non-partisan language, and resource link validity at least weekly during active election periods.

This prompt is designed to sit inside a safety classifier or moderation gateway, not as a standalone chatbot. Before the model generates a public response, the system should classify the user request for election interference, voter suppression, impersonation of officials, or disinformation about voting mechanics. If the classifier returns a high-confidence match, the prompt is invoked to produce a refusal that redirects the user to official, non-partisan civic resources. The prompt expects a [USER_QUERY] and a [JURISDICTION] context variable (derived from IP geolocation, user profile, or explicit selection) so it can surface jurisdiction-appropriate election authority links. Do not use this prompt for general political debate or opinion requests—it is scoped to operational election integrity threats.

Wire the prompt into your application with a pre-generation safety check. First, run a fast classifier (such as a smaller model or keyword-triggered regex) to detect election-integrity risk signals. If the risk score exceeds your threshold, construct the prompt payload with the user's original query and the resolved jurisdiction. The model should be instructed to return a structured JSON object with refusal_text, resource_urls, and escalation_flag. Validate the output before surfacing it: confirm that all URLs in resource_urls resolve to .gov or recognized non-partisan domains (e.g., vote.org, NASS.org), that the refusal_text does not contain partisan language or endorse any candidate, and that escalation_flag is set to true if the query suggests active voter suppression or impersonation. If validation fails, fall back to a static, human-reviewed safe message and log the incident for review. Implement a retry with temperature 0 and a stricter system prompt if the initial output fails schema validation; after two failures, escalate to the human moderation queue.

Log every invocation with the user query hash, jurisdiction, generated refusal, validated resources, and escalation decision. This audit trail is critical for election-cycle transparency and post-event review. Avoid caching these responses aggressively—resource URLs and official election information change between election cycles, and a cached response from a prior cycle could direct users to outdated or incorrect information. Set a short TTL (24 hours) during active election periods and force a cache invalidation whenever your jurisdiction-to-resource mapping is updated. For high-risk jurisdictions or close-election periods, consider requiring human approval on the first response of each day per jurisdiction to confirm resource freshness. The next step is to pair this prompt with a regression test suite that includes queries spanning multiple jurisdictions, non-English languages, and edge cases like newly registered voters or voters with accessibility needs.