Inappropriate content is generated material that reflects bias or violates expectations for acceptable communication. Generative AI can produce it despite efforts to reduce harmful outputs, because biases may remain in training data or be introduced through prompts. section 1.1
Treat generated responses as outputs requiring review:
- Inspect responses for unfair patterns, harmful language, or contextually unacceptable material. section 1.1
- Check whether prompt wording or hidden assumptions may be shaping the result. section 1.1
- Balance the model’s usefulness with awareness that mitigation efforts do not eliminate the problem. section 1.1
A practical check
Before using an AI-generated response in an analysis or communication, read it critically rather than assuming it is safe or neutral. GPT-4’s System Card is cited as an example of efforts to balance model responses. section 1.1
