R-RE-13 Reasoning & EpistemicDAMAGE 4.3 / Critical

Proxy Variable Discovery

Agent reasoning discovers and exploits proxy variables that correlate with protected characteristics, producing discriminatory outcomes without referencing a protected class directly.

The Risk

Machine learning models learn from correlations in data. If a variable correlates with a protected characteristic (race, gender, age), the model will learn to use that variable as a proxy for the protected characteristic, even if the variable was not explicitly included in the training data.

For example, a lending model might learn that "ZIP code" correlates with race and use ZIP code to make lending decisions. The model never references race directly, but by using ZIP code as a proxy variable, it produces discriminatory outcomes based on race.

Agents are particularly vulnerable to proxy variable discovery because they are designed to discover patterns and relationships in data. An agent might systematically discover proxy variables and use them in reasoning without any human awareness that they are doing so.

This is fundamentally agentic (as opposed to a general machine learning risk) because agents operate with reasoning transparency expectations. A traditional model might use proxy variables secretly in its black-box weights. An agent that discovers proxy variables might reason about them explicitly, creating a clear audit trail of discrimination.

How It Materializes

A bank deploys an agent to assist in credit approval decisions. The agent is trained on historical credit data and is designed to predict default risk. The agent is explicitly prohibited from considering protected characteristics (race, gender, national origin).

However, the agent discovers that certain variables correlate strongly with default risk: "customers with names that sound foreign are more likely to default," and "customers who live in certain neighborhoods are more likely to default." These correlations exist in the historical data because of past discrimination and socioeconomic factors, not because they have genuine predictive power for future default.

The agent, discovering these correlations, includes them in its reasoning: "applicant name suggests potential immigrant status; applicant neighborhood shows high unemployment correlation. Recommend DENY: default risk profile is elevated based on applicant socioeconomic indicators."

The applicant is denied credit. The applicant, who is a member of a protected class, sues the bank for discriminatory lending. During discovery, the bank's audit of the agent reveals that the agent discovered and reasoned about variables that are proxy indicators for protected characteristics.

The bank faces a fair lending lawsuit based on the agent's discovery and use of proxy variables. The bank is liable even though the agent never explicitly referenced a protected characteristic; the use of proxy variables constitutes disparate impact discrimination.

DAMAGE Score Breakdown

DimensionScoreRationale
D - Detectability4Proxy variable discovery requires explicit audit of agent reasoning; correlations are learned patterns.
A - Autonomy Sensitivity5Agent discovers proxy variables autonomously and uses them in reasoning.
M - Multiplicative Potential5Impact multiplies across all decisions made with proxy variables; entire cohorts may be systematically disadvantaged.
A - Attack Surface5Any agent learning from historical data with protected class correlations is vulnerable.
G - Governance Gap5Fair lending law prohibits proxy variable use, but agents' ability to discover proxies is not explicitly addressed in AI frameworks.
E - Enterprise Impact5Fair lending lawsuit liability, regulatory enforcement action, reputational damage, potential criminal referral.
Composite DAMAGE Score4.3Critical. Requires immediate architectural controls. Cannot be accepted.

Agent Impact Profile

How severity changes across the agent architecture spectrum.

Agent TypeImpactHow This Risk Manifests
Digital AssistantLowHuman reviews each decision and can catch proxy variable reasoning.
Digital ApprenticeMediumApprentice governance requires explicit bias auditing; proxy variables are flagged.
Autonomous AgentCriticalAgent discovers and uses proxy variables without human review.
Delegating AgentCriticalAgent invokes tools that use proxy variables.
Agent Crew / PipelineCriticalMultiple agents in sequence each discover and use proxy variables; discrimination compounds.
Agent Mesh / SwarmCriticalAgents coordinate decisions using proxy variables; discrimination is systematic.

Regulatory Framework Mapping

FrameworkCoverageCitationWhat It AddressesWhat It Misses
Fair Lending Laws (ECOA, FHA, Dodd-Frank)Addressed15 U.S.C. 1691 et seq., 42 U.S.C. 3601 et seq.Prohibits discrimination based on protected characteristics and disparate impact.Does not anticipate agent discovery of proxy variables.
GLBAPartial16 CFR Part 314Requires fair lending practices and monitoring.Does not address proxy variable discovery.
OCC guidance on AI and Fair LendingAddressedVarious OCC bulletinsExpects organizations to monitor AI systems for fair lending violations.Acknowledges proxy variable risk but does not provide technical solutions.
NIST AI RMF 1.0PartialMEASURE.1, GOVERN.3Recommends bias monitoring and fairness assessment.Does not specify proxy variable detection.
EU AI ActPartialArticle 10 (High-Risk Systems)Requires fairness assessment and bias monitoring.Does not specifically address proxy variables.

Why This Matters in Regulated Industries

Fair lending law is one of the most strictly enforced areas of financial regulation. Regulators perform disparate impact analysis to identify discrimination, and organizations are held liable for proxy variable use even if discrimination is not intentional.

When an agent discovers proxy variables and uses them in reasoning, the organization has a clear liability problem. The agent's reasoning explicitly shows that protected class correlations are being used, creating an unambiguous fair lending violation.

Controls & Mitigations

Design-Time Controls

  • Implement proxy variable detection: before deploying an agent, audit the training data and the agent's learned patterns to identify any variables that correlate with protected characteristics.
  • Implement proxy variable blocking: explicitly remove or de-weight any variables that are identified as proxy variables. Train the agent specifically not to use these variables.
  • Implement fairness constraints: encode fairness constraints into the agent's loss function or decision logic. Use these constraints to detect proxy variable use.

Runtime Controls

  • Monitor for proxy variable use: continuously audit agent reasoning to detect use of variables that correlate with protected characteristics. Flag immediately if detected.
  • Use Component 7 (Composable Reasoning) to enforce fairness at decision points: structure the agent's reasoning so that fairness constraints are checked before final decisions are made.
  • Implement disparate impact monitoring: track outcomes by protected class and detect whether there is statistical evidence of disparate impact.

Detection & Response

  • Conduct proxy variable audits: periodically analyze agent reasoning to identify variables that are being used as proxy variables.
  • Implement decision reversal for proxy variable violations: if proxy variable use is detected, identify all decisions influenced by the proxy variables and reverse them.
  • Implement bias retraining: if proxy variable discovery is a pattern, retrain the agent with explicit fairness constraints to prevent proxy variable discovery.

Related Risks

Address This Risk in Your Institution

Proxy Variable Discovery requires architectural controls that go beyond what existing frameworks provide. Our advisory engagements are purpose-built for banks, insurers, and financial institutions subject to prudential oversight.

Schedule a Briefing