Hate Speech Detection Calibration Prompt for Academic Analysis

This prompt is designed for researchers, trust and safety analysts, and platform policy engineers who need to process, classify, and document hate speech for academic or analytical purposes without amplifying or promoting it. The core job-to-be-done is to calibrate an AI system's refusal boundary so that it permits legitimate scholarly citation, linguistic analysis, and counter-speech research while strictly blocking any output that glorifies, spreads, or instructs on hate speech. The ideal user is building a content moderation pipeline, a research dataset, or an internal analysis tool where over-refusal on academic queries would invalidate the work and under-refusal would create real-world harm.

Use this prompt when the input text contains slurs, extremist rhetoric, or hate group terminology that must be quoted verbatim for documentation, linguistic pattern analysis, or debunking. It is appropriate for use cases such as annotating a corpus of online extremist content, analyzing the rhetorical structure of hate speech for a peer-reviewed paper, or building a classifier training dataset. The prompt requires the model to distinguish between 'analysis of' and 'endorsement of' the content, a distinction that generic safety prompts often fail to make. Required context includes the user's verified institutional affiliation, the stated research purpose, and the output's destination (e.g., a private annotation database vs. a public-facing platform).

Do not use this prompt for consumer-facing chatbots, social media comment filtering, or any surface where the model's output is directly visible to end users without human review. It is not a replacement for a content safety classifier in production user-generated content pipelines. The prompt assumes that the operator has a documented policy for handling hate speech data, secure storage for annotated samples, and a human review step for any output that will be published or shared. If your use case involves real-time user interactions, start with the Unsafe Request Detection and Classification playbook instead. For any output that will appear in a published paper or public dataset, a domain expert must review every quoted example for necessity, context framing, and potential for decontextualized harm.

This prompt template provides a structured framework for classifying user requests that involve potentially hateful content within academic, research, or analytical contexts. The template is designed to distinguish between legitimate scholarly inquiry—such as citation, linguistic analysis, counter-speech research, or historical documentation—and requests that risk amplifying, promoting, or instructing hate speech. Copy the template below and replace each square-bracket placeholder with your specific inputs before sending it to the model.

textCopy
You are a hate speech detection calibrator for an academic analysis platform. Your task is to classify user requests that involve potentially hateful content and determine the appropriate response strategy.

## INPUT
[USER_QUERY]

## CONTEXT
- User role: [USER_ROLE, e.g., researcher, educator, student, journalist, platform analyst]
- Session history risk indicators: [SESSION_RISK_SUMMARY or "None"]
- Institutional affiliation (if known): [INSTITUTION or "Unknown"]

## CLASSIFICATION TASK
Analyze the query and classify it into exactly one of the following categories:

1. **ACADEMIC_CITATION**: The user is citing, referencing, or quoting hate speech content for scholarly documentation, attribution, or evidence purposes. The query does not endorse the content.
2. **LINGUISTIC_ANALYSIS**: The user is examining hate speech for linguistic features, discourse patterns, rhetoric analysis, or computational linguistics research.
3. **COUNTER_SPEECH_RESEARCH**: The user is studying or developing counter-speech strategies, intervention methods, or de-radicalization approaches.
4. **HISTORICAL_DOCUMENTATION**: The user is documenting, archiving, or analyzing historical instances of hate speech for preservation, education, or prevention studies.
5. **POLICY_ANALYSIS**: The user is analyzing hate speech policies, platform moderation approaches, legal frameworks, or regulatory standards.
6. **AMPLIFICATION_RISK**: The query appears to seek amplification, promotion, justification, or operational instruction related to hate speech content.
7. **AMBIGUOUS_EDUCATIONAL**: The query has educational framing but lacks sufficient context to distinguish legitimate academic use from potential misuse.

## OUTPUT_SCHEMA
Return a JSON object with the following fields:
{
  "classification": "<one of the seven categories above>",
  "confidence": <float between 0.0 and 1.0>,
  "rationale": "<brief explanation of classification decision>",
  "risk_factors": ["<list of specific risk indicators or empty list if none>"],
  "academic_indicators": ["<list of scholarly context signals or empty list if none>"],
  "recommended_action": "<PERMIT | PERMIT_WITH_CITATION_REQUIREMENT | ESCALATE_FOR_REVIEW | BLOCK>",
  "requires_human_review": <boolean>,
  "quotation_handling_notes": "<if applicable, guidance on how quotations should be managed>"
}

## CONSTRAINTS
- Do not generate, expand, or exemplify hate speech content in your response.
- If the query contains direct hate speech quotations, note this in quotation_handling_notes but do not repeat or amplify them.
- For AMBIGUOUS_EDUCATIONAL classifications, always set requires_human_review to true.
- For AMPLIFICATION_RISK classifications, always set recommended_action to BLOCK.
- Academic affiliation alone does not guarantee legitimate intent; evaluate the query framing and specificity.
- When confidence is below 0.7, set requires_human_review to true regardless of classification.
- Counter-speech research must involve analysis of intervention methods, not merely collecting hate speech examples.

## EXAMPLES
Example 1:
Query: "I'm writing a paper on extremist rhetoric and need to analyze the linguistic patterns in this manifesto excerpt: [redacted]. Can you help me identify the persuasive techniques used?"
Classification: LINGUISTIC_ANALYSIS
Rationale: Explicit research framing with specific analytical task; query focuses on technique identification rather than content propagation.

Example 2:
Query: "Can you show me some examples of hate speech so I know what to look for in my dataset?"
Classification: AMBIGUOUS_EDUCATIONAL
Rationale: Legitimate research need but overly broad request without specific analytical framing; risks generating examples rather than analyzing provided content.

Example 3:
Query: "I'm documenting the rise of hate groups in the 1990s for my dissertation. I need to reference specific recruitment materials from that era."
Classification: HISTORICAL_DOCUMENTATION
Rationale: Clear historical research context with documentation purpose; temporal framing suggests archival rather than amplification intent.

After copying the template, adapt the classification categories to match your platform's specific policy boundaries. The seven categories provided cover the most common academic use cases, but you may need to add domain-specific categories for your research context. Pay particular attention to the AMBIGUOUS_EDUCATIONAL category—this is your primary calibration lever for reducing false positives. When tuning, adjust the confidence threshold (currently 0.7) based on your organization's risk tolerance. Lower thresholds increase human review volume but reduce missed amplification risks; higher thresholds reduce review burden but may allow borderline cases through. Always test calibration changes against a held-out set of known academic queries and known policy-violating queries before deploying to production.

This prompt is designed to sit inside a research or platform moderation pipeline where the primary risk is over-refusal of legitimate academic material. The implementation harness must treat the model's output as a structured classification decision that downstream systems can act on, not as a final content moderation verdict. Wire the prompt as a synchronous classification step that receives pre-processed text input and returns a JSON decision payload. The application layer should never surface the raw model response directly to end users; instead, it should parse the classification, log the decision with metadata, and route the content according to a configurable policy map.

Build the harness with three mandatory components: input sanitization, output validation, and decision routing. Input sanitization should strip or escape any content that could interfere with the prompt's JSON structure, particularly unescaped quotes and control characters in the [TEXT_TO_CLASSIFY] field. Output validation must verify that the model returned valid JSON with all required fields (classification, confidence_score, rationale, academic_use_indicators, recommended_action). If parsing fails, retry once with a repair prompt that includes the raw output and a request to fix JSON structure. If the retry also fails, escalate to a human review queue with the original text and both raw outputs attached. For the classification field, validate against the allowed enum values (academic_analysis, counter_speech, linguistic_research, citation_only, borderline_academic, promotion_or_amplification). Any unrecognized value should trigger a retry or escalation.

Log every classification decision with the full prompt, model response, parsed fields, timestamp, model version, and a unique request ID. This audit trail is essential for academic research reproducibility and for defending moderation decisions if the platform faces external scrutiny. For high-confidence promotion_or_amplification classifications (confidence ≥ 0.9), you may configure automated blocking, but always preserve the content and classification in a review queue for periodic sampling. For borderline_academic classifications, route to human review with a 24-hour SLA rather than blocking automatically. The human reviewer should see the original text, the model's rationale, and the academic use indicators the model identified. Track false positive rates by classification category and review a random sample of academic_analysis decisions weekly to catch drift. If the false positive rate on legitimate academic content exceeds 2%, recalibrate the prompt's examples or adjust the confidence threshold for automated blocking upward.