External Bias Integration

Version: 1.1.0 Status: Production Ready

Overview

ATHENA External Bias Integration enables you to ingest fairness metrics from industry standard bias detection tools and correlate them with ATHENA human trust calibration data. This combination reveals bias amplification, which occurs when human behavior reinforces algorithmic bias.

The Problem

Statistical fairness tools such as IBM AIF360, Fairlearn, and AWS Clarify detect model bias, identifying when an AI produces unfair outputs for specific demographic groups.

These tools miss a critical layer: how humans interact with biased output.

ATHENA detects two key behavioral patterns:

• Overtrust: Users following biased AI recommendations without appropriate scrutiny

• Undertrust: Users rejecting accurate AI recommendations for specific groups

When you combine external model bias data with ATHENA human trust patterns, you can identify bias amplification. This is the most dangerous scenario, where human behavior makes algorithmic bias worse.

Supported External Tools

Tool	Tool ID	Metrics Supported
IBM AI Fairness 360	ibm_aif360	Demographic Parity, Equalized Odds, Disparate Impact
Microsoft Fairlearn	fairlearn	Demographic Parity, Equalized Odds, Equal Opportunity
AWS SageMaker Clarify	aws_clarify	SHAP values, Disparate Impact, Statistical Parity
Google Vertex Explainability	google_vertex	Feature attributions
Custom	custom	Any fairness metric

Quick Start

Step 1: Generate Fairness Metrics

Run your bias detection tool. The following example uses IBM AIF360:

from aif360.metrics import BinaryLabelDatasetMetric

metric = BinaryLabelDatasetMetric(
    dataset, 
    unprivileged_groups=[{'gender': 0}],
    privileged_groups=[{'gender': 1}]
)

statistical_parity = metric.statistical_parity_difference()
disparate_impact = metric.disparate_impact()

Step 2: Send to ATHENA

curl -X POST https://api.athenatrust.ai/v1/model-fairness-signals \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "externalToolId": "ibm_aif360",
    "externalToolVersion": "0.6.1",
    "modelId": "loan_approval_v3",
    "metricName": "statistical_parity_difference",
    "metricValue": 0.15,
    "threshold": 0.1,
    "passesThreshold": false,
    "protectedAttribute": "gender",
    "privilegedGroup": "male",
    "unprivilegedGroup": "female",
    "sampleSize": 10000,
    "rawPayload": {
      "statistical_parity_difference": 0.15,
      "disparate_impact": 0.82
    },
    "signalTimestamp": "2025-12-27T10:30:00Z"
  }'

Step 3: Automatic Correlation

ATHENA performs the following actions automatically:

1. Stores the fairness signal

2. Correlates with human decisions for the same model

3. Calculates amplification risk

4. Generates alerts when amplification is detected

5. Fires bias.amplification webhook when severity is high or above

Amplification Risk Levels

Risk Level	Amplification Score	Description
none	0.0 to 0.09	No amplification detected
low	0.1 to 0.24	Minor amplification. Monitor the situation.
medium	0.25 to 0.49	Moderate amplification. Investigation recommended.
high	0.5 to 0.74	Significant amplification. Action required.
critical	0.75 to 1.0	Severe amplification. Immediate intervention required.

API Endpoints

Fairness Signals

Method	Endpoint	Description
POST	/model-fairness-signals	Ingest external fairness signal
GET	/model-fairness-signals	List signals with filters
GET	/model-fairness-signals/:id	Get single signal

Amplification Alerts

Method	Endpoint	Description
GET	/bias/amplification	List amplification alerts
GET	/bias/amplification/:id	Get single alert
PATCH	/bias/amplification/:id	Update alert status
POST	/bias/amplification/analyze/:signalId	Trigger analysis

Webhook Events

Two webhook events are available for external bias integration.

model.bias.ingested

Fired when a fairness signal is successfully ingested.

{
  "event": "model.bias.ingested",
  "timestamp": "2025-12-27T10:30:00Z",
  "data": {
    "signalId": "sig_abc123",
    "modelId": "loan_approval_v3",
    "externalToolId": "ibm_aif360",
    "metricName": "demographic_parity",
    "metricValue": 0.82,
    "passesThreshold": true
  }
}

bias.amplification

Fired when amplification risk reaches high or above.

{
  "event": "bias.amplification",
  "timestamp": "2025-12-27T10:35:00Z",
  "data": {
    "alertId": "alert_xyz789",
    "modelId": "loan_approval_v3",
    "severity": "high",
    "amplificationScore": 0.68,
    "protectedAttribute": "gender",
    "affectedSubgroup": "female",
    "overtrustRate": 0.42,
    "message": "42% overtrust rate detected for female applicants on biased model"
  }
}

SDK Support

JavaScript

import Athena from '@athena-ai/sdk';

const athena = new Athena({ apiKey: 'YOUR_API_KEY' });

// Submit fairness signal
const { signalId } = await athena.fairness.submitSignal({
  externalToolId: 'ibm_aif360',
  modelId: 'loan_approval_v3',
  metricName: 'demographic_parity',
  metricValue: 0.82,
  protectedAttribute: 'gender',
  rawPayload: { },
  signalTimestamp: new Date().toISOString()
});

// Get amplification alerts
const { alerts } = await athena.fairness.listAmplificationAlerts({
  severity: 'high',
  status: 'open'
});

Python

from athena_sdk import Athena

athena = Athena(api_key="YOUR_API_KEY")

# Submit fairness signal
signal = athena.fairness.submit_signal(
    external_tool_id="ibm_aif360",
    model_id="loan_approval_v3",
    metric_name="demographic_parity",
    metric_value=0.82,
    protected_attribute="gender",
    raw_payload={},
    signal_timestamp="2025-12-27T10:30:00Z"
)

# Get amplification alerts
alerts = athena.fairness.list_amplification_alerts(
    severity="high",
    status="open"
)

Regulatory Mapping

Regulation	Article	How External Bias Integration Helps
EU AI Act	Art. 10	Documents external bias audits
EU AI Act	Art. 14	Detects human oversight failures
Colorado AI Act	Section 1702	Subgroup impact analysis
Texas TRAIGA	Section 2(a)	Algorithmic accountability

Best Practices

1. Run external tools regularly. Execute bias analysis daily or after model retraining.

2. Include confidence intervals. This helps ATHENA assess statistical significance.

3. Use consistent model IDs. Match the identifiers in your ATHENA decision logs.

4. Set appropriate thresholds. Define acceptable bias levels for your industry.

5. Subscribe to webhooks. Enable real time amplification alerts.

Next Steps

• IBM AIF360 Integration

• Microsoft Fairlearn Integration

• AWS SageMaker Clarify Integration

• Google Vertex Integration