Synopsis
This first instalment in the Teaching AI Ethics 2025 series revisits the theme of bias in generative AI. It explains how data bias, model bias and human bias interact to produce skewed or discriminatory outputs in large-language and image-generation systems, illustrates those problems with up-to-date research and examples, critiques the limitations of current "guard-rail" fixes, and closes with practical ways teachers can embed critical discussions of AI bias across English, Mathematics, Civics, Visual Arts and other subjects.
Key Topics
- How data bias, model bias, and human bias interact
- Up-to-date research and examples of discriminatory AI outputs
- Limitations of current "guard-rail" fixes
- Practical teaching approaches across multiple subject areas
Key Resources
- Bender et al., "On the Dangers of Stochastic Parrots" — ACM Conference on Fairness, Accountability, and Transparency
- OpenAI's documentation on ChatGPT bias
- Anthropic's published system prompts
- Research on AI image generators and gender/racial bias
- Studies on how guardrails and system prompts attempt (and fail) to eliminate bias