① Regular human rights training shall be provided to ensure that activists and staff are able to recognize biased or discriminatory expressions in generative AI outputs. [Alternatively, a designated reviewer for AI-generated outputs may be appointed.] ② If potentially problematic expressions are identified during the use of generative AI, use of the output shall be halted immediately and the issue reported to the [designated reviewer]. ③ The generative AI system should be instructed to revise the content in a non-discriminatory manner, and the revised output should be reviewed again. ④ Issues identified should be reported to the company or service provider operating the generative AI system. ⑤ If a generative AI system repeatedly produces discriminatory or hateful content, its use shall be discontinued. ⑥ Rather than relying solely on generative AI, users should consider gathering information and perspectives through alternative sources and channels. Generative AI is trained on vast amounts of data collected from the internet. This data often directly reflects discriminatory language, gender-, race-, and region-based biases, social hierarchies, and stereotypes. For example, if many people commonly use the term “illegal immigrant” rather than “undocumented migrant,” generative AI is likely to reproduce that terminology. In this way, there is a high risk that stigma, discrimination, stereotypes, and hateful expressions targeting marginalized or vulnerable groups will be amplified and reproduced. If such outputs are used without critical 70 71