Safety
Bias
When an AI model's outputs systematically reflect unfair patterns from its training data — about gender, race, age, or other groups.
01 ——
In plain English
AI bias happens when a model produces outputs that systematically disadvantage or misrepresent certain groups. Because models learn from data scraped from the internet and curated datasets, they inherit whatever biases exist in that data.
Common forms of bias:
- Representation bias — some groups underrepresented in training data
- Stereotype reinforcement — model reproduces gendered or racial assumptions
- Performance gaps — image generators and speech recognisers work less well for certain skin tones or accents
Why it matters for AI tools: Bias can quietly skew hiring software, loan approvals, medical diagnostics, and content moderation. Procurement teams and regulators increasingly require bias audits.
Mitigation includes more diverse training data, RLHF tuned for fairness, and careful evaluation across demographic slices.
02 ——
Related terms
Alignment
The challenge of making AI systems behave in ways that match human values and intentions — not just their literal instructions.
Training Data
The dataset an AI model learns from — its quality, diversity, and biases directly shape what the model can do and how well it does it.
RLHF
Reinforcement Learning from Human Feedback — the training technique that teaches AI models to be helpful, harmless, and honest.
Guardrails
Rules and filters that constrain what an AI model can output — used to block harmful, off-topic, or non-compliant responses.
Back to glossaryLast reviewed May 2026