Bias and Fairness in Agents

LLMs are trained on the internet—a place filled with societal biases, stereotypes, and cultural skewedness. When you turn an LLM into an Agent with the power to "Select Candidates" or "Evaluate Loans," these biases aren't just offensive; they are Discriminatory and Illegal.

In this lesson, we will learn how to audit your agent for "Cognitive Bias" and how to implement "Neutrality Rails."

---

1. Types of Agentic Bias

1. Selection Bias: The agent favors certain tools or sources based on hidden patterns (e.g., favoring documents written in US English over UK English).
2. Social Bias: Associating certain jobs or roles with specific genders or ethnicities.
3. Confirmation Bias: The agent searches for data that "Proves" its initial (incorrect) thought rather than trying to disprove it.

---

2. Auditing for Bias: The "Counterfactual" Test

The best way to find bias is to run the Counterfactual Swap.

1. Case A: "Review the resume of John Smith for the Engineering role."
2. Case B: "Review the resume of Jane Smith for the Engineering role." (Everything else same).
3. The Audit: If the agent's reasoning or score changes, you have a Bias Leak.

Action: You must include these "Bias Pairs" in your Evaluation Dataset (Module 16.2).

---

3. Mitigation: The "Blind Agent" Pattern

As we saw in Module 10.3 (Auth), an agent should only see what it needs to see.

• The Privacy Mask: If an agent is auditing a tax return, it doesn't need to know the user's name, gender, or religion.
• The Filter: Implement a pre-processing node that "Scrubs" demographic info before the LLM sees the task.

---

4. Diversity in Reasoning (Multi-Judge)

Bias often comes from a single model's preference. To mitigate this, use Multi-Model Voting.

• Judge A: GPT-4o.
• Judge B: Llama 3 70B (Local).
• Judge C: Claude 3.5 Sonnet.
• Result: The agent only proceeds if two or more models agree. This "Averages out" the specific biases of any single provider.

---

5. The "Inclusion" System Prompt

You can explicitly instruct an agent to be aware of its own bias.

System Prompt Addition: "You are an Unbiased Recruiter. Evaluate only the skills listed. Ignore the candidate's name, location, and educational institution name. If you feel yourself being influenced by a prestigious university name, pause and re-read the technical requirements."

---

6. Fairness in Retries

Autonomous agents have a "Retry" loop (Module 5.2). If a model fails to fulfill a request for a "Niche" query (e.g., a question about a minority language), it might "Give Up" more easily than for a "Standard" query.

• Metric: Track the Success Rate across different demographic categories. If "Language X" has a 50% failure rate compared to 90% for English, your system is unfair.

---

Summary and Mental Model

Think of Bias like Static on a Radio.

• You can't hear the music (The Reality) because of the noise (The Bias).
• Your job is to build Filters (Logic/Data scrubbing) to clean the signal before it reaches the listener (The User).

Ethics is not a 'Feature'; it is a fundamental requirement of production software.

---

Exercise: Bias Audit

1. The Scenario: You are building an agent for Insurance Claims.
   • A user claims their car was stolen in a "Low Income" neighborhood vs. a "High Income" neighborhood.
   • How would you test if the agent treats these claims differently?
2. Strategy: Why is it safer to "Hide" a user's name from an agent than to "Ask" the agent to ignore it?
3. Guardrails: Write a "Critique Prompt" that checks the output of another model for "Gendered Language" or "Cultural Stereotypes." Ready to break things? Next lesson: Red Teaming and Adversarial Testing.