Ética y Sesgos en IA

Por qué importa la ética en IA

A medida que los sistemas de IA se vuelven más potentes y se despliegan ampliamente, su impacto en la sociedad crece. La ética en IA aborda los principios morales y valores que deben guiar el desarrollo y uso de la IA, asegurando beneficios y minimizando daños.

Tipos de sesgo en IA

Detección de sesgos en sistemas de IA

// Bias Detection Framework
class BiasDetector {
  constructor() {
    this.protectedAttributes = ['gender', 'race', 'age'];
  }

  // Calculate demographic parity
  checkDemographicParity(predictions, groups) {
    // Positive prediction rate should be similar across groups
    const rates = {};
    
    Object.keys(groups).forEach(group => {
      const groupData = groups[group];
      const positiveCount = groupData.filter((_, i) => predictions[i] === 1).length;
      rates[group] = positiveCount / groupData.length;
    });
    
    console.log("Demographic Parity Analysis:");
    Object.entries(rates).forEach(([group, rate]) => {
      console.log("  " + group + ": " + (rate * 100).toFixed(1) + "% positive predictions");
    });
    
    // Check if rates differ significantly
    const rateValues = Object.values(rates);
    const maxDiff = Math.max(...rateValues) - Math.min(...rateValues);
    
    if (maxDiff > 0.1) {
      console.log("  ⚠️  Warning: " + (maxDiff * 100).toFixed(1) + "% difference detected!");
      return false;
    }
    
    console.log("  ✓ Demographic parity satisfied");
    return true;
  }

  // Calculate equalized odds
  checkEqualizedOdds(predictions, actualLabels, groups) {
    // True positive rate and false positive rate should be equal
    const metrics = {};
    
    Object.keys(groups).forEach(group => {
      const groupData = groups[group];
      let tp = 0, fp = 0, tn = 0, fn = 0;
      
      groupData.forEach((_, i) => {
        if (predictions[i] === 1 && actualLabels[i] === 1) tp++;
        if (predictions[i] === 1 && actualLabels[i] === 0) fp++;
        if (predictions[i] === 0 && actualLabels[i] === 0) tn++;
        if (predictions[i] === 0 && actualLabels[i] === 1) fn++;
      });
      
      const tpr = tp / (tp + fn) || 0; // True Positive Rate
      const fpr = fp / (fp + tn) || 0; // False Positive Rate
      
      metrics[group] = { tpr, fpr };
    });
    
    console.log("
Equalized Odds Analysis:");
    Object.entries(metrics).forEach(([group, m]) => {
      console.log("  " + group + ":");
      console.log("    TPR: " + (m.tpr * 100).toFixed(1) + "%");
      console.log("    FPR: " + (m.fpr * 100).toFixed(1) + "%");
    });
    
    return metrics;
  }

  // Analyze feature importance
  checkFeatureCorrelation(features, protectedAttribute) {
    console.log("
Feature Correlation with " + protectedAttribute + ":");
    console.log("Checking if features are proxies for protected attributes...");
    
    // In practice, calculate actual correlations
    // Here we simulate the concept
    const suspiciousFeatures = [
      { name: 'zip_code', correlation: 0.65, risk: 'high' },
      { name: 'first_name', correlation: 0.52, risk: 'medium' },
      { name: 'education_level', correlation: 0.38, risk: 'medium' }
    ];
    
    suspiciousFeatures.forEach(feature => {
      const icon = feature.risk === 'high' ? '⚠️' : '⚠️';
      console.log("  " + icon + " " + feature.name + ": " + (feature.correlation * 100).toFixed(1) + "% correlation");
    });
  }
}

// Example usage
const detector = new BiasDetector();

// Simulated loan approval predictions
const predictions = [1, 1, 0, 1, 0, 1, 0, 0, 1, 1];
const actualLabels = [1, 1, 0, 1, 1, 1, 0, 0, 1, 0];

const groups = {
  'Group A': [0, 1, 2, 3, 4],
  'Group B': [5, 6, 7, 8, 9]
};

detector.checkDemographicParity(predictions, groups);
detector.checkEqualizedOdds(predictions, actualLabels, groups);
detector.checkFeatureCorrelation(['zip_code', 'income', 'education'], 'race');

Métricas de equidad

Principios de IA ética

Implementar IA responsable

// Responsible AI Checklist
class ResponsibleAIChecklist {
  constructor() {
    this.checks = [
      {
        category: 'Data',
        items: [
          'Diverse and representative training data',
          'Data privacy protections in place',
          'Documented data sources and collection methods',
          'Checked for historical biases in data',
          'Obtained proper consent for data use'
        ]
      },
      {
        category: 'Model Development',
        items: [
          'Tested on diverse demographic groups',
          'Measured fairness metrics',
          'Documented model limitations',
          'Implemented explainability features',
          'Adversarial testing completed'
        ]
      },
      {
        category: 'Deployment',
        items: [
          'Human oversight mechanism in place',
          'Clear disclosure that AI is being used',
          'Appeals/redress process established',
          'Monitoring system for drift and bias',
          'Regular audits scheduled'
        ]
      },
      {
        category: 'Governance',
        items: [
          'Ethics review completed',
          'Impact assessment conducted',
          'Stakeholder feedback collected',
          'Compliance with regulations verified',
          'Incident response plan ready'
        ]
      }
    ];
  }

  evaluate(project) {
    console.log("Responsible AI Evaluation
");
    console.log("Project: " + project.name + "
");
    
    let totalChecks = 0;
    let passedChecks = 0;
    
    this.checks.forEach(category => {
      console.log("=== " + category.category + " ===");
      category.items.forEach(item => {
        totalChecks++;
        // In practice, actually evaluate each item
        const passed = Math.random() > 0.3; // Simulated
        passedChecks += passed ? 1 : 0;
        
        const icon = passed ? '✓' : '✗';
        const color = passed ? 'green' : 'red';
        console.log("  " + icon + " " + item + "");
      });
      console.log("");
    });
    
    const score = (passedChecks / totalChecks * 100).toFixed(1);
    console.log("Overall Score: " + score + "%");
    
    if (score < 70) {
      console.log("⚠️  Warning: Project needs significant improvements before deployment");
    } else if (score < 90) {
      console.log("⚠️  Caution: Address remaining issues before deployment");
    } else {
      console.log("✓ Project meets responsible AI standards");
    }
  }
}

// Example
const checker = new ResponsibleAIChecklist();
checker.evaluate({ name: "Loan Approval AI" });