Commercial Bias Detection Prompt for Sources

This prompt is designed for trust-and-safety engineers and evidence pipeline builders who need to automatically flag commercially compromised sources before they enter a RAG system, research brief, or decision-support workflow. It identifies sponsored content, native advertising, affiliate-driven reviews, and vendor-influenced research by analyzing funding disclosures, promotional language patterns, and relationship signals in the source text and its metadata. The ideal user is someone integrating this check into an automated ingestion pipeline where sources arrive from mixed-authority domains—places like industry blogs, review sites, trade publications, and research portals where paid placement and editorial content coexist without clear labeling.

Use this prompt when your pipeline ingests content from domains where commercial influence is possible but not always disclosed. Common triggers include: processing product reviews before they enter a market intelligence brief, vetting research reports before citing them in a decision memo, or filtering search results in a domain where affiliate marketing is prevalent. The prompt works best when you can provide both the source text and available metadata—author affiliations, funding statements, publication venue, and any disclosure language. It is not designed to assess technical accuracy, factual correctness, or scientific validity. A source can be commercially biased yet factually accurate, or unbiased yet wrong. This prompt evaluates commercial influence signals only.

Do not use this prompt as a standalone fact-checker, a replacement for editorial judgment, or the sole gatekeeper for high-stakes decisions. In regulated domains such as financial analysis, medical evidence review, or legal research, the output must feed into a broader review pipeline that includes human oversight. The prompt can produce false positives—flagging legitimate product comparisons as biased—and false negatives—missing sophisticated native advertising that mimics editorial style. Always pair this check with source authority assessment, factual verification, and cross-source corroboration before blocking or downgrading content. The next section provides the copy-ready template you can adapt for your pipeline.

This prompt template is designed to be copied directly into your AI harness and adapted with your specific source content, metadata, and output requirements. It instructs the model to act as a commercial bias auditor, systematically examining a source for funding disclosures, affiliate relationships, promotional language, and other signals that compromise objectivity. The prompt forces the model to cite specific evidence for each finding rather than returning vague impressions, making the output auditable and defensible in trust-and-safety workflows.

textCopy
You are a commercial bias auditor evaluating source content for trust-and-safety review. Your task is to analyze the provided source and return a structured JSON assessment of commercial bias indicators.

## INPUT
Source Content:
[SOURCE_TEXT]

Source Metadata:
[SOURCE_METADATA]

## INSTRUCTIONS
Examine the source for the following commercial bias signals. For each signal detected, provide specific evidence from the source text or metadata. If a signal is not detected, set its `detected` field to `false` and provide an empty `evidence` array.

### Required Signals to Check
1. **Funding Disclosure**: Does the source disclose who funded the content, research, or publication?
2. **Affiliate Relationships**: Does the source contain affiliate links, commission arrangements, or referral incentives?
3. **Sponsored Content Markers**: Is the content labeled as sponsored, promoted, paid, or native advertising?
4. **Promotional Language Patterns**: Does the text use marketing language, superlatives, urgency cues, or one-sided product advocacy?
5. **Vendor or Product Affiliation**: Is the author, publisher, or organization financially connected to products, services, or vendors mentioned in the content?
6. **Selective Citation**: Does the source cite only favorable evidence while ignoring contradictory research or competing products?
7. **Undisclosed Conflicts**: Are there indications of financial relationships, employment ties, or ownership interests that are not explicitly disclosed?

## OUTPUT SCHEMA
Return ONLY valid JSON matching this structure:
{
  "assessment_id": "string",
  "overall_bias_risk": "low" | "medium" | "high" | "critical",
  "overall_confidence": 0.0-1.0,
  "signals": [
    {
      "signal_name": "string",
      "detected": true | false,
      "confidence": 0.0-1.0,
      "evidence": [
        {
          "quote": "exact text from source",
          "location": "paragraph or section reference",
          "explanation": "why this quote indicates commercial bias"
        }
      ]
    }
  ],
  "summary": "concise explanation of findings and risk rationale",
  "requires_human_review": true | false
}

## CONSTRAINTS
- Quote source text exactly in evidence fields. Do not paraphrase.
- If multiple pieces of evidence exist for one signal, include all of them.
- Set `requires_human_review` to `true` if overall_bias_risk is `high` or `critical`, or if any signal has confidence below 0.7.
- Do not invent funding sources, affiliations, or relationships not present in the provided content.
- If the source is too short to assess (under 100 words), set overall_bias_risk to `low`, confidence to 0.3, and note the limitation in summary.

Adaptation guidance: Replace [SOURCE_TEXT] with the full text of the article, research paper, blog post, or document under review. Replace [SOURCE_METADATA] with available metadata such as author name, publisher domain, publication date, and any known funding or affiliation records. If your pipeline already extracts metadata separately, pass it as a structured JSON string inside the metadata placeholder. For batch processing, wrap this prompt in a loop that feeds one source per call and aggregates results. The output schema is designed for direct insertion into a trust-and-safety database or review queue. Before deploying, run this prompt against a golden dataset of known-biased and known-neutral sources to calibrate your risk thresholds and confidence expectations.

The Commercial Bias Detection Prompt is designed to operate as a pre-retrieval or post-retrieval filter in an evidence pipeline. It should run before source content is used for answer generation or evidence ranking. The prompt accepts source metadata (title, publisher, author affiliations, funding disclosures, and content excerpts) and returns a structured bias assessment. In production, this prompt is typically called via an API within a processing step that gates downstream consumption: sources flagged with high commercial bias risk are either demoted in ranking, excluded from evidence sets, or routed for human review depending on your risk tolerance.

Integration pattern: Wrap the prompt in a service function that accepts a SourceRecord object and returns a BiasAssessment schema. Validate the output against a strict JSON schema before accepting it. Common fields to enforce: bias_risk_level (enum: low, medium, high, critical), commercial_relationship_detected (boolean), funding_disclosures_found (boolean), promotional_language_score (float 0-1), affiliate_links_detected (boolean), and evidence_citations (array of strings quoting specific text spans that support the assessment). If the model returns malformed JSON or missing required fields, use a retry with schema reminder pattern: resend the same input with an added constraint like Your previous response was missing the required field 'evidence_citations'. Return ONLY valid JSON matching the output schema. After two retries, route the source to a human review queue rather than silently accepting a degraded assessment.

Model selection and latency: For high-throughput pipelines processing thousands of sources, use a fast classifier model (GPT-4o-mini, Claude Haiku, or a fine-tuned open-weight model) for the initial bias screen. Reserve a more capable model (GPT-4o, Claude Sonnet) for sources flagged as high or critical risk, or for cases where the initial assessment has low confidence. Log every assessment with the model version, prompt template hash, input source identifier, and output scores. This audit trail is essential for trust-and-safety teams that need to explain why a source was excluded or demoted. Tool use: If your pipeline has access to external databases (advertiser registries, media ownership databases, or funding disclosure repositories), provide these as tools the model can call to verify disclosures rather than relying solely on the source text. The prompt template's [TOOLS] placeholder should be populated with function definitions for any available verification endpoints.

Threshold tuning and evals: Do not deploy this prompt without a labeled evaluation set. Build a golden dataset of 200-500 sources with known commercial relationships (sponsored content, affiliate marketing, vendor-funded research, and independent sources as negative examples). Measure precision and recall at each risk threshold. Common failure modes to test: (1) academic papers with declared industry funding being misclassified as promotional content, (2) product review sites with standard affiliate disclaimers being flagged as high risk, (3) non-English sources where funding language follows different conventions. Run these evals on every prompt or model change before deployment. For high-stakes domains like healthcare or finance, require human review of all high and critical assessments before sources are excluded from evidence pipelines.