Uncertainty Language Insertion Prompt

This prompt is for product teams who need to calibrate the tone of an AI-generated output when the underlying confidence is moderate. The core job-to-be-done is rewriting a high-confidence-sounding draft to include appropriate uncertainty language—hedging, qualifiers, and probability markers—without undermining the factual content or making the output sound useless. The ideal user is a developer or product manager integrating an AI feature where overconfidence erodes user trust, such as a diagnostic assistant, a financial commentary tool, or a medical information summarizer. You need a pre-existing draft text and a known confidence score or level to use this prompt effectively.

This prompt is not a confidence estimator. It assumes you already have a confidence signal from an upstream classifier, a model's log probabilities, or a separate calibration step. Do not use it to generate the initial answer or to assess confidence; it only rewrites for tone. It is also not a fact-checker—it will not correct hallucinations or add missing evidence. If the draft contains fabricated details, inserting uncertainty language will make it sound responsibly uncertain but will not fix the underlying inaccuracy. For high-stakes domains like healthcare or legal, always pair this with source-grounding and human review, and never rely on tone adjustment alone to manage risk.

Before using this prompt, ensure you have a clear definition of what 'moderate confidence' means in your system, a draft output that is factually as accurate as you can make it, and a target tone that preserves usefulness. The prompt works best when you can provide concrete examples of acceptable and unacceptable uncertainty language for your domain. Avoid using it when confidence is very low—in those cases, an abstention or clarification request prompt is more appropriate than a hedged answer. The next step is to copy the prompt template, adapt the placeholders to your domain's terminology, and run it against a golden set of drafts with known confidence levels to calibrate the output tone.

This prompt template is designed to be dropped into a production pipeline after a model has generated a draft answer and a confidence score has been assigned. It takes the original draft, the confidence score, and the source context, then rewrites the output to include appropriate hedging, qualifiers, or probability language. The goal is to preserve the useful content while ensuring the tone matches the actual certainty level—preventing overconfident statements that erode user trust or create compliance risk.

textCopy
You are a calibrated uncertainty editor. Your task is to rewrite a draft response so that its language accurately reflects the provided confidence score. Do not remove factual content unless it is unsupported by the source context. Do not add new facts. Your only job is to adjust the tone and certainty language.

[CONSTRAINTS]
- If confidence is HIGH (>=0.9): Keep the original tone. Minor hedging is acceptable only where the source context is ambiguous.
- If confidence is MEDIUM (0.7-0.89): Insert appropriate hedging language (e.g., "likely," "suggests," "may indicate," "appears to") for claims that are not directly and unambiguously supported by the source context.
- If confidence is LOW (0.5-0.69): Add explicit uncertainty markers (e.g., "it is possible that," "limited evidence suggests," "we cannot confirm") and flag the output for human review.
- If confidence is VERY LOW (<0.5): Do not rewrite. Instead, output the abstention template: "I don't have enough information to answer this confidently."
- Preserve all citations and source references exactly as they appear in the draft.
- Do not change the structure, bullet points, or numbered lists.
- Maintain the original language unless inserting uncertainty markers.

[INPUT]
Confidence Score: [CONFIDENCE_SCORE]
Draft Response: [DRAFT_RESPONSE]
Source Context: [SOURCE_CONTEXT]

[OUTPUT_SCHEMA]
{
  "rewritten_response": "string (the full rewritten response with uncertainty language applied)",
  "uncertainty_markers_added": ["string (list of specific phrases or sentences where hedging was inserted)"],
  "requires_human_review": boolean,
  "abstention_triggered": boolean
}

[EXAMPLES]
Example 1 (Medium Confidence):
Confidence Score: 0.78
Draft Response: "The error is caused by a misconfigured load balancer. Restart the service to resolve the issue."
Source Context: "Logs show a spike in 503 errors after the load balancer config was updated. The service restart temporarily resolved similar issues in the past."
Rewritten Response: "The error is likely caused by a misconfigured load balancer, based on the timing of the config change and the 503 error spike. Restarting the service may resolve the issue, as this has helped in similar past incidents."

Example 2 (Low Confidence):
Confidence Score: 0.62
Draft Response: "The customer is eligible for a full refund under our policy."
Source Context: "Policy section 4.2 states refunds are available for 'defective products reported within 30 days.' The customer reported the issue on day 31, but the ticket notes the defect was present on day 28."
Rewritten Response: "The customer may be eligible for a refund, but this is uncertain. Our policy covers defects reported within 30 days, and the report came on day 31. However, the defect was documented on day 28, which could support an exception. Human review is recommended."

Adapt this template by adjusting the confidence thresholds in [CONSTRAINTS] to match your organization's risk tolerance. For regulated domains, lower the MEDIUM threshold to 0.8 and require human review for anything below it. Replace the abstention template with your own fallback language. The [OUTPUT_SCHEMA] is designed for programmatic consumption—wire the requires_human_review boolean into your escalation pipeline and log the uncertainty_markers_added array for audit trails. Before deploying, run this prompt against a golden dataset of 50-100 draft-response pairs with known confidence scores and have a domain expert evaluate whether the rewritten tone matches the actual certainty level. Common failure modes include over-hedging on high-confidence outputs (making the system sound unsure when it shouldn't) and under-hedging on medium-confidence outputs (preserving overconfident language).

The Uncertainty Language Insertion Prompt is designed to sit between a content generation step and the final user-facing output. In a typical pipeline, a primary model or RAG system produces a high-confidence draft, and this prompt rewrites it to include calibrated uncertainty language when the system's confidence score falls below a configurable threshold. The harness must pass three inputs: the original draft text, a structured confidence payload (containing the overall score and any per-claim or per-source confidence breakdowns), and a set of tone and accuracy preservation constraints. The output is a rewritten version of the draft with inserted hedging, qualifiers, or explicit uncertainty statements, plus a diff-like change log for auditability.

Wire this prompt into your application as a post-processing step with a strict validation layer. After receiving the rewritten output, run a schema validator to confirm the response includes both the rewritten_text and change_log fields. Then apply a content-preservation check: use a sentence-transformers similarity score or an LLM-as-judge eval to verify that the rewritten text does not contradict the original draft on factual claims. If the similarity score drops below 0.92 or the judge flags a factual deviation, discard the rewrite and fall back to the original draft with a prepended uncertainty disclaimer. For high-risk domains such as healthcare or legal, route any output where the confidence score is below 0.7 to a human review queue before publication. Log every insertion decision—original confidence, threshold used, changes made, and whether human review was triggered—for audit and prompt iteration.

Model choice matters here. Use a model with strong instruction-following and tone control, such as GPT-4o or Claude 3.5 Sonnet, because the task requires precise lexical edits without semantic drift. Avoid smaller or older models that may over-insert hedging (turning every sentence into a disclaimer) or strip out useful content entirely. Set temperature to 0 or a very low value (0.1–0.2) to keep edits consistent. Implement a single retry with a stricter prompt variant if the first attempt fails validation; if the retry also fails, fall back to the original draft with an uncertainty banner. Do not loop beyond one retry—this is a tone-calibration task, not a reasoning task, and repeated attempts rarely improve output quality. For RAG systems, pass per-source confidence scores into the prompt's [CONFIDENCE_DATA] placeholder to enable claim-level hedging rather than blanket uncertainty across the entire response.