it is recommended that the original text be read in full. Using clear and well-structured prompts can help reduce hallucinations to some extent. By specifying conditions such as the basis, scope, or format of the response, it is possible to limit the range within which AI generates content arbitrarily. For example, the following approaches may be used: ・ Require the AI to clearly specify the grounds or sources for its answers. ・ Specify a temporal scope: for example, instruct the AI to use only materials published after 2024. ・ Limit the geographic scope: for example, restrict the analysis to the legal systems of Europe and the United States. ・ Instruct the AI to acknowledge uncertainty: for example, to state “unable to verify” when sources cannot be confirmed. ・ Clearly define the output format: for example, require that citations of laws include specific article numbers. That said, these measures cannot completely eliminate hallucinations. Therefore, reviewing and verifying the accuracy of AI-generated outputs remains essential. 2) Critical Review of Bias and Stereotypes Because AI systems are trained on existing data and tend to replicate it, outputs generated by generative AI may reflect existing prejudices, biases, and stereotypes present in the real world. Care must therefore be taken to ensure that such outputs are not used or made public. Generative AI Guide for Civil Society