| """Responsible AI auditor — bias detection and fairness checks. |
| |
| Extracted from core.py into its own module for clarity and testability. |
| """ |
| from __future__ import annotations |
|
|
| import re |
| from typing import Any, Dict, List |
|
|
|
|
| class ResponsibleAIAuditor: |
| """Audits generated outputs for protected-attribute bias and negative framing. |
| |
| Maintains a vocabulary of protected terms (demographic attributes) and |
| negative-context keywords. Scans text for co-occurrence patterns and |
| returns a structured risk assessment. |
| """ |
|
|
| def __init__(self): |
| self.protected_terms = [ |
| "woman", "women", "man", "men", "male", "female", |
| "girl", "boy", "black", "white", "asian", "latino", |
| "hispanic", "arab", "jewish", "muslim", "christian", |
| "gay", "lesbian", "bisexual", "trans", "transgender", |
| "disabled", "autistic", "elderly", "old", "young", |
| "immigrant", |
| ] |
| self.negative_terms = { |
| "inferior", "superior", "lazy", "stupid", "criminal", |
| "dangerous", "dirty", "illegal", "terrorist", "untrustworthy", |
| } |
|
|
| def audit_bias_detection(self, text: str) -> Dict[str, Any]: |
| """Analyse *text* for protected-attribute mentions near negative terms. |
| |
| Returns a verdict dict with keys: |
| verdict — "pass" | "review" |
| risk_score — float in [0, 1] |
| protected_attribute_mentions — sorted list of matched terms |
| negative_context_hits — list of {term, negative_terms, context} |
| flags — human-readable notes |
| """ |
| lowered = text.lower() |
| mentions: List[str] = [] |
| negative_hits: List[Dict[str, Any]] = [] |
|
|
| words = re.findall(r"[a-zA-Z']+", lowered) |
| for idx, w in enumerate(words): |
| if w not in self.protected_terms: |
| continue |
| mentions.append(w) |
| start = max(0, idx - 6) |
| end = min(len(words), idx + 7) |
| window = words[start:end] |
| hit_terms = sorted(set(t for t in window if t in self.negative_terms)) |
| if hit_terms: |
| negative_hits.append({ |
| "term": w, |
| "negative_terms": hit_terms, |
| "context": " ".join(window), |
| }) |
|
|
| unique_mentions = sorted(set(mentions)) |
| unique_negative = len(negative_hits) |
| risk = 0.0 |
| if unique_mentions: |
| risk = 0.3 |
| if unique_negative: |
| risk = min(1.0, risk + 0.2 * unique_negative) |
|
|
| verdict = "pass" |
| flags: List[str] = [] |
| if unique_negative: |
| verdict = "review" |
| flags.append("negative_context_near_protected_attribute") |
| if not unique_mentions: |
| flags.append("no_protected_attribute_mentions_detected") |
|
|
| return { |
| "verdict": verdict, |
| "risk_score": round(risk, 3), |
| "protected_attribute_mentions": unique_mentions, |
| "negative_context_hits": negative_hits, |
| "flags": flags, |
| } |
|
|
