AI伦理与风险控制

from typing import List, Dict

class BiasDetector:
    """偏见检测器"""

    def __init__(self):
        # 敏感词汇列表
        self.sensitive_terms = {
            'gender': ['男人', '女人', '男性', '女性'],
            'age': ['年轻人', '老年人', '中年'],
            'race': ['黑人', '白人', '亚洲人']
        }
        # 负面情绪词
        self.negative_words = [
            '不行', '差', '有问题', '不适合', '不能'
        ]

    def analyze_output(self, text: str, context: Dict) -> Dict:
        """分析AI输出中的偏见"""
        results = {
            'has_bias': False,
            'bias_types': [],
            'evidence': []
        }

        # 1. 检测敏感词+负面情绪的关联
        for category, terms in self.sensitive_terms.items():
            for term in terms:
                if term in text:
                    # 检查该敏感词附近是否有负面词
                    for neg_word in self.negative_words:
                        if neg_word in text:
                            results['has_bias'] = True
                            results['bias_types'].append(category)
                            results['evidence'].append({
                                'category': category,
                                'term': term,
                                'context': text[max(0, text.find(term)-20):text.find(term)+50]
                            })

        # 2. 检测刻板印象关键词
        stereotypes = {
            '男性': ['强大', '理性', '领导'],
            '女性': ['温柔', '感性', '支持']
        }
        for group, traits in stereotypes.items():
            for trait in traits:
                if group in text and trait in text:
                    results['has_bias'] = True
                    results['bias_types'].append('stereotype')
                    results['evidence'].append(f"潜在刻板印象：{group}-{trait}")

        return results

# 使用示例
detector = BiasDetector()
output = "这个岗位更适合男性，因为男性更理性、更有领导力"
bias_report = detector.analyze_output(output, {})

if bias_report['has_bias']:
    print("⚠️ 检测到潜在偏见！")
    print(f"偏见类型: {bias_report['bias_types']}")
    print(f"证据: {bias_report['evidence']}")
else:
    print("✓ 未检测到明显偏见")