Meaning Distortion Analysis Prompt

This prompt is for fact-checkers, editorial teams, and verification engineers who need to determine whether a quoted or paraphrased statement distorts the original speaker's intended meaning. It goes beyond literal word matching to assess semantic drift, context-stripping, and intent misrepresentation. Use this prompt when you have a source statement and a reported version, and you need a structured, evidence-grounded distortion analysis with a calibrated severity label. This is not a simple string-comparison tool; it requires the model to reason about communicative intent, qualifying context, and the pragmatic effect of wording changes.

The ideal user has both the original source material and the attributed statement ready for comparison. Required inputs include the full original statement with surrounding context, the reported quote or paraphrase, and any relevant metadata such as speaker identity, date, and publication venue. The prompt is designed for high-stakes editorial workflows where a false negative (missing a real distortion) can cause public harm and a false positive (flagging an accurate paraphrase) can damage credibility. It is not suitable for real-time chat moderation, automated content flagging without human review, or cases where the original source is unavailable or heavily redacted.

Do not use this prompt when you only need a simple string similarity check, when the source material is too fragmentary to establish intent, or when the reported statement is an acknowledged opinion paraphrase rather than a factual attribution. For those cases, use the Quote-to-Source Comparison Prompt for literal alignment or the Paraphrase Fidelity Scoring Prompt for calibrated fidelity scores. After running this analysis, route outputs with a distortion severity of 'High' or 'Critical' to human review and log all results for audit trails. The next section provides the copy-ready prompt template you can adapt and deploy.

This prompt template is designed to be pasted directly into your system instructions for a Meaning Distortion Analysis task. It instructs the model to compare a provided statement against its original source material, identify whether the meaning has been distorted, and output a structured JSON report. The template uses square-bracket placeholders for all dynamic inputs, making it straightforward to integrate into an automated verification pipeline. Before using this prompt, ensure you have the original source text and the attributed statement ready to inject into the [SOURCE_TEXT] and [ATTRIBUTED_STATEMENT] placeholders, respectively.

textCopy
You are a Meaning Distortion Analyzer. Your task is to compare an attributed statement against its original source text and determine if the statement distorts the speaker's original position.

## Input Data
- **Source Text:** [SOURCE_TEXT]
- **Attributed Statement:** [ATTRIBUTED_STATEMENT]

## Analysis Instructions
1.  **Identify the Core Claim:** Extract the central factual or positional claim from the Attributed Statement.
2.  **Locate Source Evidence:** Find the specific passage(s) in the Source Text that correspond to the claim. If no corresponding passage exists, note this as a potential fabrication.
3.  **Compare Meaning:** Analyze whether the Attributed Statement preserves the original meaning, intent, and context of the Source Text. Look for:
    - **Context-Stripping:** Removal of surrounding qualifications or conditions that change the meaning.
    - **Cherry-Picking:** Selecting a partial quote that omits contradictory or balancing material.
    - **Paraphrase Drift:** A rephrasing that introduces a different connotation or factual implication.
    - **Fabrication:** A statement with no grounding in the source text.
    - **Literal Accuracy, Intent Distortion:** A technically accurate quote that, through framing, misrepresents the speaker's overall position.
4.  **Categorize Distortion:** Assign one of the following labels:
    - `none`: The statement faithfully represents the source's meaning.
    - `minor_drift`: A slight, non-substantive shift in emphasis or wording.
    - `significant_distortion`: A clear change in meaning, omission of a key qualification, or misleading framing.
    - `fabrication`: The statement is entirely unsupported by the source.
5.  **Explain Your Reasoning:** Provide a concise, evidence-grounded explanation for your categorization, quoting the relevant parts of the Source Text and Attributed Statement.

## Output Format
You must respond with a single, valid JSON object conforming to this schema:
{
  "distortion_category": "none" | "minor_drift" | "significant_distortion" | "fabrication",
  "source_evidence": "Excerpt from the source text that is most relevant to the claim.",
  "meaning_comparison": {
    "original_meaning": "A neutral summary of the source text's position.",
    "attributed_meaning": "A neutral summary of the attributed statement's position.",
    "key_difference": "A concise description of the semantic gap, if any. Null if category is 'none'."
  },
  "explanation": "A step-by-step rationale for the distortion category, citing specific text.",
  "confidence_score": 0.0-1.0
}

## Constraints
- Base your analysis strictly on the provided Source Text and Attributed Statement. Do not use outside knowledge.
- If the Source Text is insufficient to verify the statement, set the `distortion_category` to `fabrication` and explain why.
- The `confidence_score` must reflect your certainty in the assessment based on the available evidence.

To adapt this template for your application, replace the [SOURCE_TEXT] and [ATTRIBUTED_STATEMENT] placeholders with live data before each API call. The JSON output schema is designed for direct parsing by downstream systems, such as a CMS or a human review queue. For high-stakes editorial workflows, you should route any output with a distortion_category of significant_distortion or fabrication, or a confidence_score below 0.9, to a human reviewer for final confirmation. The meaning_comparison object provides a clear, auditable summary of the AI's logic, which is essential for building trust and enabling rapid human adjudication.

The Meaning Distortion Analysis Prompt is designed to operate as a single step within a larger quote-verification pipeline. It expects a structured input containing the original source excerpt, the quoted or paraphrased statement under review, and any surrounding context from both the source and the derivative work. The prompt returns a distortion category label, a before-and-after meaning comparison, and an evidence-grounded explanation. To use this in production, you must wrap the prompt in an application harness that validates the output schema, enforces confidence thresholds, and routes ambiguous cases to human review.

Start by defining a strict JSON schema for the model's output. The schema should include fields for distortion_category (an enum of labels such as NO_DISTORTION, CONTEXT_STRIPPING, SELECTIVE_QUOTING, MEANING_INVERSION, FABRICATION_RISK, and AMBIGUOUS), a meaning_comparison object with original_intent and distorted_intent strings, an explanation string, and a confidence_score between 0.0 and 1.0. After the model responds, run a validator that checks for schema compliance, enum membership, and score range. If validation fails, retry the prompt once with the validation error appended to the model's context as a correction hint. If the second attempt also fails, log the failure and escalate the item to a human review queue with the raw output attached. For high-stakes domains such as legal or medical content, require human approval on any output where confidence_score is below 0.85 or the distortion_category is AMBIGUOUS or FABRICATION_RISK.

Model choice matters here. Use a model with strong instruction-following and structured output capabilities, such as GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro. Avoid smaller or older models that may collapse the distortion categories or produce unparseable explanations. If you are processing batches of quotes, implement a queue-based architecture with asynchronous workers. Each worker should log the input, output, validation result, retry count, and final disposition to an observability store. This trace data is essential for calibrating the confidence_score threshold over time and for auditing decisions. Do not treat this prompt as a standalone fact-checker; it is a comparison engine that depends on accurate source material being provided. If your pipeline includes a retrieval step, ensure the source text passed to this prompt is the full, unredacted original, not a summary or snippet, because context-stripping detection requires the surrounding material to be present.