AI Bias
Systematic errors in AI outputs caused by biases in training data, model architecture, prompts, or agent configurations that must be continuously monitored and mitigated.
Also known as: Algorithmic Bias, AI Discrimination, Machine Learning Bias
Category: AI
Tags: ai, ethics, biases, fairness
Explanation
AI bias refers to systematic errors in AI output caused by biases in training data, model architecture, prompts, skills, or agent configurations. Bias is not a bug that gets fixed once; it is a property of every AI system that must be continuously monitored and mitigated.
## Where bias lives
### In the models
- **Training data bias**: models learn from internet text, which reflects historical biases in race, gender, culture, profession, and worldview
- **Representation gaps**: underrepresented languages, cultures, and perspectives get worse performance
- **Recency bias**: training data cutoffs mean models favor older knowledge while recent developments are missing
- **Western/English-centric**: most training data is English and Western, skewing outputs toward those perspectives
### In the prompts and skills
- **Framing bias**: how you phrase a prompt shapes the answer. "What are the risks?" produces different output than "What are the opportunities?"
- **Anchoring bias**: examples given in few-shot prompts anchor the model toward those patterns
- **Selection bias**: which context you load (and which you leave out) biases the output toward that subset of knowledge
- **Confirmation bias in skills**: skills that look for specific things tend to find them, ignoring contradicting evidence
### In the agents
- **Identity bias**: an agent defined as an "expert in X" will overweight X's importance and underweight alternatives
- **Memory bias**: what an agent remembers shapes future interactions. Accumulated memories can drift toward particular viewpoints
- **Routing bias**: if routing favors certain agents, certain perspectives dominate
- **Panel composition bias**: which agents sit on an evaluation panel determines which perspectives are represented
## Mitigation
- Use diverse agent panels to get multiple perspectives
- Audit prompts and skills for framing and anchoring effects
- Include explicit "consider the opposite" or "what am I missing" instructions
- Rotate examples in few-shot prompts to avoid anchoring
- Use AI guardrails to flag one-sided outputs
- Conduct regular bias audits on model outputs across demographic groups
Related Concepts
← Back to all concepts