# AI Bias 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's 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; 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 the [[AI Agent Routing|routing]] favors certain agents, certain perspectives dominate - **Panel composition bias**: which agents sit on an [[AI Agent Panels|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 ([[AI Skill Testing]]) - Include explicit "consider the opposite" or "what am I missing" instructions in skills - Rotate examples in few-shot prompts to avoid anchoring - Use [[AI Guardrails]] to flag one-sided outputs - See [[How to limit AI Bias]] for practical techniques ## References - ## Related - [[How to limit AI Bias]] - [[AI Safety]] - [[AI Alignment]] - [[AI Guardrails]] - [[AI Hallucination]] - [[AI Sycophancy]] - [[Responsible AI]] - [[AI Agent Identity]] - [[AI Agent Panels]] - [[AI Skill Testing]] - [[Context Poisoning]]