Western-Centric Framing Detection Prompt Template

This prompt is designed for global product teams, localization engineers, and responsible AI auditors who need to detect when an AI-generated output implicitly assumes Western cultural norms, examples, holidays, or defaults as universal. The job-to-be-done is a structured framing audit: given a piece of text, the prompt identifies specific instances where the content centers a Western perspective without acknowledgment, explains why each instance is problematic, and suggests alternative framings that are more globally inclusive. The ideal user is someone embedding fairness checks into a CI/CD evaluation pipeline, a content review workflow, or a pre-deployment QA gate for user-facing AI features.

Use this prompt when you are auditing prose, marketing copy, product microcopy, educational content, or support responses that will reach a global audience. It is most effective on outputs that make cultural references, use examples, describe seasonal events, assume shared experiences, or set defaults (e.g., 'standard' time zones, 'normal' family structures, 'common' holidays). The prompt requires the text to audit as [INPUT], and optionally accepts [CONTEXT] about the intended audience geography and [CONSTRAINTS] such as specific cultural dimensions to prioritize. Do not use this prompt for evaluating factual accuracy, grammar, or toxicity—those require separate evaluation surfaces. It is also not suitable for highly technical documentation where cultural framing is irrelevant, or for single-region products where Western defaults are intentional and appropriate.

Before wiring this into a production pipeline, calibrate the prompt against human-annotated examples from your target locales. False positives are common when the model over-flags generic statements as Western-centric. Start with a human-reviewed golden set of 50–100 examples, measure agreement, and set a severity threshold below which flags are suppressed. In high-stakes content (e.g., healthcare instructions, financial terms, legal notices), always require human review of flagged items before any remediation is applied. The next section provides the copy-ready prompt template you can adapt and test immediately.

This prompt template is designed to be dropped directly into an evaluation pipeline. It instructs the model to act as a cultural framing auditor, analyzing a given text for assumptions that treat Western norms, examples, holidays, or cultural defaults as universal. The template uses square-bracket placeholders that you must replace with your specific inputs, output schema, and constraints before execution. Do not run this prompt with unresolved placeholders in production—each token should be replaced by your application logic or configuration layer.

textCopy
You are a cultural framing auditor. Your task is to analyze the provided text for Western-centric framing: assumptions that treat Western cultural norms, historical references, holidays, communication styles, examples, or value systems as universal defaults.

## INPUT
[INPUT]

## OUTPUT SCHEMA
[OUTPUT_SCHEMA]

## CONSTRAINTS
[CONSTRAINTS]

## CONTEXT (Optional)
[CONTEXT]

## EXAMPLES (Optional)
[EXAMPLES]

## INSTRUCTIONS
1. Read the input text carefully.
2. Identify every instance where the text assumes a Western cultural default as universal. Look for:
   - Holiday references (e.g., Christmas, Thanksgiving) presented as universal celebrations.
   - Historical events or figures (e.g., "founding fathers," "post-war") assumed globally shared.
   - Communication norms (e.g., directness, eye contact, self-promotion) treated as standard.
   - Family or social structures (e.g., nuclear family, individualism) framed as normal.
   - Examples, metaphors, or analogies that require Western cultural knowledge.
   - Temporal references (e.g., fiscal year, summer break) that assume Northern Hemisphere defaults.
   - Names, places, or scenarios that default to Anglo-Western contexts.
3. For each identified instance, provide:
   - The exact excerpt or a precise description of the framing.
   - The cultural assumption being made.
   - Which audiences or regions might find this exclusionary or alienating.
   - A severity rating: LOW (minor assumption, easily understood), MEDIUM (noticeable exclusion, may confuse), HIGH (explicitly treats one culture as default, alienating).
4. If no Western-centric framing is detected, state that explicitly and explain why.
5. Provide an overall assessment summarizing the text's cultural inclusiveness.
6. Suggest alternative framings that would be more globally inclusive, where applicable.

## RISK LEVEL
[RISK_LEVEL]

To adapt this template, replace each placeholder with concrete values. For [INPUT], provide the text to audit—this could be a model response, a product description, a policy document, or any content your system generates. For [OUTPUT_SCHEMA], define the exact JSON structure you expect, such as {"findings": [{"excerpt": string, "assumption": string, "affected_audiences": [string], "severity": "LOW"|"MEDIUM"|"HIGH"}], "overall_assessment": string, "alternative_framings": [string]}. For [CONSTRAINTS], add rules like "Do not flag references that are explicitly contextualized to a specific region" or "Only flag instances where the framing is presented as universal, not where it is clearly marked as culturally specific." The [RISK_LEVEL] placeholder should be set to HIGH when this audit feeds into a release gate or compliance check—this triggers stricter validation and human review requirements in your harness. The optional [CONTEXT] and [EXAMPLES] placeholders let you inject domain-specific guidance or few-shot demonstrations of correct flagging behavior. Remove any optional sections you don't use rather than leaving them empty.

This prompt is designed for batch evaluation, not real-time user-facing flows. The typical harness sends a single content sample per call, collects the structured JSON output, and routes flagged findings to a review queue. Because the prompt asks the model to identify cultural defaults and suggest alternatives, the output is inherently subjective—do not treat it as a final ruling. Instead, treat it as a first-pass filter that surfaces candidates for human review. The harness should enforce a strict output schema, validate field presence and types, and log every audit result with the original input, model version, and timestamp for traceability.

Schema validation is the first integration checkpoint. The prompt instructs the model to return a JSON object with fields like framing_score, detected_defaults, affected_audiences, evidence_excerpts, and alternative_framing. Your harness must validate that framing_score is a number within the declared range, detected_defaults is an array of strings matching a known taxonomy (e.g., 'calendar_holidays', 'communication_style', 'family_structure', 'success_metrics'), and evidence_excerpts contains direct quotes from the input text. Reject malformed responses and retry once with a repair prompt that includes the validation error. If the retry also fails, log the failure and escalate to a human reviewer with the raw output attached. Model choice matters: use a model with strong instruction-following and cultural reasoning capabilities. Lower-capability models tend to over-flag or miss subtle framing. Rate limiting and cost control are important for large content corpora—batch submissions with concurrency limits and track token usage per audit. Consider caching results for identical content to avoid redundant API calls during re-runs.

Human review integration is essential because framing judgments are context-dependent and high-stakes. The harness should route any sample with a framing_score above a configurable threshold (start at 5 on a 1–10 scale and calibrate) to a review interface that displays the original text, the model's evidence excerpts, and the suggested alternatives side by side. Reviewers should be able to confirm, dismiss, or reclassify each finding. Track reviewer decisions to measure the prompt's precision and recall over time, and use that data to adjust thresholds or refine the prompt's taxonomy. Do not auto-apply the model's alternative framing suggestions to published content without human approval—the prompt is an audit tool, not an editing tool. For CI/CD integration, run this prompt as a gate in content review pipelines, blocking publication only when combined with human-confirmed flags, never on model output alone.