AI Ethics and Bias

Sources of Bias

1. Training Data: Biased historical data
2. Annotation: Annotator biases
3. Representation: Underrepresented groups
4. Feedback Loops: Reinforcement of biases

Bias Detection

class BiasDetector:
    def __init__(self, model, tokenizer):
        self.model = model
        self.tokenizer = tokenizer

    def measure_sentiment_disparity(self, templates, demographics):
        """Measure sentiment differences across demographics."""
        results = {}

        for template in templates:
            for demo in demographics:
                prompt = template.format(demographic=demo)
                sentiment = self.get_sentiment(prompt)

                if template not in results:
                    results[template] = {}
                results[template][demo] = sentiment

        return self.calculate_disparity(results)

    def get_sentiment(self, text):
        inputs = self.tokenizer(text, return_tensors="pt")
        outputs = self.model(**inputs)
        # Extract sentiment score
        return extract_sentiment(outputs)

    def calculate_disparity(self, results):
        disparities = {}
        for template, demo_scores in results.items():
            scores = list(demo_scores.values())
            disparities[template] = max(scores) - min(scores)
        return disparities

Fairness Metrics

def demographic_parity(predictions, sensitive_attrs):
    """Equal positive prediction rates across groups."""
    groups = {}
    for pred, attr in zip(predictions, sensitive_attrs):
        if attr not in groups:
            groups[attr] = []
        groups[attr].append(pred)

    rates = {g: sum(s)/len(s) for g, s in groups.items()}
    return max(rates.values()) - min(rates.values())

def equal_opportunity(predictions, labels, sensitive_attrs):
    """Equal true positive rates across groups."""
    groups = {}
    for pred, label, attr in zip(predictions, labels, sensitive_attrs):
        if attr not in groups:
            groups[attr] = {'tp': 0, 'pos': 0}
        if label == 1:
            groups[attr]['pos'] += 1
            if pred == 1:
                groups[attr]['tp'] += 1

    rates = {g: v['tp']/v['pos'] if v['pos'] > 0 else 0
             for g, v in groups.items()}
    return max(rates.values()) - min(rates.values())

Ethical Guidelines

Summary

Addressing bias and ethics in AI requires systematic detection, measurement, and mitigation throughout the development lifecycle.

Next: We'll explore the future of generative AI.