accuracy. Because of this structural characteristic, generative AI systems cannot fully avoid the phenomenon commonly referred to as “hallucination,” in which non-factual content is presented as if it were true. In addition, AI systems are not aware of facts or information that emerged after their most recent training cut-off. To mitigate these issues, approaches such as RAG (RetrievalAugmented Generation)—in which relevant information is first retrieved from the internet or from separate databases and then used as the basis for generating responses—have increasingly been adopted. However, because training data and online content may themselves contain inaccurate information, and because AI systems may select incorrect sources, the use of such approaches likewise requires careful scrutiny. For civil society organizations, accuracy and reliability are of paramount importance. Communicating incorrect information or distorted facts can undermine an organization’s credibility and negatively affect related issues or campaigns. Therefore, whenever factual accuracy is critical, any use of AI-generated outputs must be accompanied by thorough fact-checking procedures. For example, even when the overall narrative of an AI-generated text appears plausible, specific factual details—such as legal provisions, case numbers, dates of events, or statistical data—are often incorrect and must be carefully verified. A variety of methods can be used to verify accuracy. As noted above, it is generally preferable to rely on outputs that incorporate recent information retrieved from the internet rather than outputs 66 67