Bias Detection

ATHENA detects bias across demographic subgroups, satisfying FDA AI/ML Guidelines Section IV.B and EU AI Act Article 10 requirements.

What is Bias?

Bias occurs when AI system performance varies systematically across protected groups:

Bias Type	Description	Example
Accuracy Disparity	AI is less accurate for some groups	85% accuracy for men, 62% for women
Representation Disparity	Some groups underrepresented in training	90% of training data from ages 25-45
Treatment Disparity	Different recommendations for similar cases	Recommending different treatments by race

Supported Subgroups

ATHENA supports both standard and custom demographic attributes:

Standard Attributes

Attribute	Values
Gender	male, female, non-binary, other
Age Group	18-24, 25-34, 35-44, 45-54, 55-64, 65+
Ethnicity	Configurable per customer
Region	Configurable per customer

Custom Attributes

Define any subgroup relevant to your domain:

{
  "subgroups": {
    "department": "Emergency",
    "shift": "Night",
    "experience_level": "Junior"
  }
}

Detection Methods

1. Statistical Parity

Compare rates across groups:

Approval rate (Group A): 72%
Approval rate (Group B): 51%
Disparity: 21 percentage points → BIAS DETECTED

2. Accuracy Disparity

Compare AI accuracy across groups:

AI Accuracy (Age 25-45): 89%
AI Accuracy (Age 65+): 64%
Disparity: 25 percentage points → BIAS DETECTED

3. Four-Fifths Rule (EEOC Standard)

If any group's rate is less than 80% of the highest group:

Highest rate: 72%
Threshold: 72% × 0.8 = 57.6%
Group B rate: 51%
51% < 57.6% → BIAS DETECTED

Severity Levels

Severity	Criteria	Action
High	Disparity >20% OR affects >100 users	Immediate intervention required
Medium	Disparity 10-20% OR affects 50-100 users	Review within 7 days
Low	Disparity <10% OR affects <50 users	Monitor and track

API Example

Detect Bias in Real-Time

const result = await athena.bias.detect({
  userId: 'user_123',
  decision: {
    actionTaken: 'approved',
    aiRecommended: 'approved',
    subgroups: {
      gender: 'female',
      ageGroup: '65+',
      department: 'Emergency'
    }
  }
});

if (result.biasDetected) {
  console.log(result.biasType);        // "accuracy_disparity"
  console.log(result.affectedGroups);  // ["age_group:65+"]
  console.log(result.severity);        // "high"
  console.log(result.recommendation);  // "Immediate review required"
}

Get Subgroup Performance

const subgroups = await athena.bias.subgroups({
  customerId: 'org_123',
  days: 30
});

for (const group of subgroups.data) {
  console.log(`${group.subgroup}: ${group.aiAccuracy}% accuracy`);
  if (group.biasDetected) {
    console.log(`  ⚠️ Bias detected: ${group.severity}`);
  }
}

Bias Alert Feed

Real-time feed of bias alerts:

const alerts = await athena.bias.feed({
  severity: 'high',
  limit: 50
});

for (const alert of alerts.data) {
  console.log(`${alert.timestamp}: ${alert.biasType} in ${alert.affectedSubgroup}`);
  console.log(`  Affected users: ${alert.affectedUsers}`);
  console.log(`  Recommendation: ${alert.recommendation}`);
}

Compliance Mapping

Regulation	Requirement	ATHENA Solution
EU AI Act Art 10	Training data bias	Representation disparity detection
FDA AI/ML IV.B	Demographic performance	Accuracy disparity by subgroup
Texas TRAIGA	Bias detection	Real-time bias feed
Colorado AI Act	Impact assessment	Subgroup performance reports

Webhooks

Set up real-time alerts for bias events:

await athena.webhooks.create({
  url: 'https://your-app.com/webhook',
  events: ['bias.detected'],
  filters: {
    severity: ['high', 'medium']
  }
});

Webhook Payload:

{
  "event": "bias.detected",
  "timestamp": "2025-12-25T10:00:00Z",
  "data": {
    "bias_type": "accuracy_disparity",
    "affected_subgroup": "age_group:65+",
    "severity": "high",
    "affected_users": 89,
    "disparity_percentage": 25,
    "recommendation": "Immediate review required (FDA AI/ML Section IV)"
  }
}

---

Next: Compliance Mapping