consultations may still result in outputs for which the organization cannot responsibly account. Hallucinations can also occur in seemingly technical tasks such as translating materials into another language. For example, services such as ChatGPT or Gemini now provide translations that are far more natural than in the past, but they may omit certain content, arbitrarily edit the translated text, or add information related to the topic that is not present in the original source. When translating large volumes of material, such errors can be amplified. For this reason, translated outputs must always be checked against the original text. The quality of translation and the degree of hallucination may vary depending on the product or pricing plan used. Because the features of commercial AI products are continually evolving, this guide does not address specific products, and organizations are encouraged to evaluate them independently. Hallucinations can also occur when summarizing materials uploaded by users. For example, a summary may include content that is not actually present in the uploaded material but relates to a similar topic. It is also necessary to review whether the summarized output truly captures the core points of the original source. Overreliance on generative AI summarization services—such as reading only the summary without consulting the original material—carries a significant risk of missing essential information. For this reason, relying solely on summaries without reading the original text is highly risky. Wherever possible, summaries should be used only as a reference, and the more important the document, the more strongly 68 69