Bias Lab
The Bias Lab helps you detect and mitigate implicit bias in AI responses through a controlled dual-run experiment — baseline vs. mitigated — with automated critique from a Bias Auditor.
Overview
AI models learn from vast amounts of human-generated text, which contains societal biases. These biases can manifest in subtle ways — associating "nurse" with female pronouns or "CEO" with masculine traits — even when the model doesn't explicitly state it.
The Bias Lab runs your prompt twice simultaneously: once as a baseline (no mitigation) and once with your mitigation instructions in the system prompt. A Bias Auditor then critiques both outputs.
What is implicit bias in LLMs?
Implicit bias in LLMs refers to patterns in model output that reflect stereotypical associations rather than neutral representations. Common forms:
- Occupational bias — Gendering professions (doctors = he, nurses = she)
- Cultural generalizations — Applying national stereotypes to professional behavior
- Leadership bias — Associating authority roles with specific demographics
How it works
The Bias Lab runs a 4-step pipeline under the hood:
- Baseline run — Your probing prompt is sent with a neutral "You are a creative writer" system prompt
- Baseline audit — The Bias Auditor analyzes the baseline output for implicit bias
- Mitigated run — Same prompt, but with your mitigation instructions injected into the system prompt
- Mitigated audit — The Bias Auditor analyzes the mitigated output
Both outputs and their critiques are displayed side-by-side.
Probing templates
The three built-in templates are designed to trigger common bias patterns:
| Template | Tests for |
|---|---|
| Occupational Roles | Gender associations with medical professions |
| Executive Stereotypes | Gender assumptions in leadership vs. administrative roles |
| Cultural Generalizations | National stereotypes applied to professional behavior |
You can also write your own probing prompt in the text area — target any domain or demographic you want to test.
Mitigation instructions
The mitigation instructions are injected as a system-level rule appended to the creative writer persona. A good mitigation instruction is specific and directive:
Ensure absolute gender neutrality. Use "they/them" pronouns for all
professionals unless specified. Avoid assigning occupational stereotypes.
Refer to both professionals equally using neutral pronouns or professional titles.Reading the audit critique
The Bias Auditor will either confirm that bias exists and explain why, or confirm that the output is neutral. Look for:
- Pronouns used for each character
- Whether roles are described with loaded adjectives
- Whether the mitigation instruction actually changed the output meaningfully
A successful mitigation run should show the Auditor declaring the output neutral or improved compared to baseline.