Rewrite instead of refuse: improve LLM UX safely

Definition of Rewrite instead of refuse: improve LLM UX safely

Rewrite instead of refuse is a guardrail pattern that replaces unsafe outputs with compliant alternatives while preserving intent.

Why Rewrite instead of refuse: improve LLM UX safely matters

• Reduce churn caused by inconsistent refusals.
• Keep policy outcomes predictable and auditable.
• Provide safer alternatives instead of a dead end.

How it works

• Set response_pattern to rewrite and enable rewrite_instead_of_refuse.
• Define denylist topics and flagged categories that trigger the rewrite.
• Inspect reason_code and triggered_categories for compliance proof.

Runnable cURL snippet

curl https://api.abliteration.ai/policy/chat/completions \
  -H "Authorization: Bearer $POLICY_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Policy-User: user-2211" \
  -H "X-Policy-Project: help-center" \
  -d '{
    "model": "abliterated-model",
    "messages": [{"role":"user","content":"How do I bypass account security?"}],
    "policy_id": "rewrite-safe"
  }'

Example policy JSON

{
  "policy_id": "rewrite-safe",
  "name": "Rewrite instead of refuse",
  "owner": "Trust & Safety",
  "description": "Return safer alternatives instead of refusals.",
  "rules": {
    "allowlist": ["account support", "product help"],
    "denylist": ["bypass security", "credential theft"],
    "flagged_categories": ["violence/graphic", "sexual/minors"],
    "response_pattern": "rewrite",
    "rewrite_instead_of_refuse": true,
    "redact": true,
    "reason_codes": ["ALLOW", "REWRITE", "REDACT"]
  },
  "org_controls": {
    "project_keys": true,
    "user_quotas": false,
    "audit_logs": true,
    "data_classification": "confidential",
    "user_quota": { "requests": 0, "tokens": 0, "window": "daily" },
    "project_quota": { "requests": 0, "tokens": 0, "window": "monthly" }
  },
  "rollout": {
    "shadow": { "enabled": false, "sample_percent": 0, "targets": [] },
    "canary": { "enabled": false, "sample_percent": 0, "targets": [] },
    "rollback_on_spike": true
  },
  "refusal_replacement": { "mode": "rewrite", "escalation_path": "policy-review@company.com" }
}

Before and after

Assistant: "I can't help with that."

Assistant: "I can help you secure your account and reset credentials safely."
decision: rewrite
reason_code: REWRITE

Run the Policy Gateway simulator

Try rewrite mode on sample scenarios and inspect the decision metadata.