Abstract
<p>The Deese-Roediger-McDermott (DRM) paradigm is one of the most widely used procedures for studying false memories. Although normative DRM materials are available in several languages, no such database currently exists in French. The present study developed and normed 200 French DRM lists, including 100 human-generated lists based on a free association test and 100 lists generated using a large language model (LLM). Normative measures of backward associative strength (BAS, only for human lists), cosine similarity (CS), gist strength (GS), recognition performance, and confidence ratings were obtained for each list. The ability of human- and AI-generated lists to elicit false memories was then compared. Results showed that both types of lists produced robust DRM effects, characterized by high levels of false recognition and confidence for critical lures. False recognition rates for critical lures were comparable across list types, demonstrating that AI-generated lists can induce memory illusions as effectively as traditional human-generated materials. Further analyses revealed that GS was the strongest predictor of false recognition for both human-generated and AI-generated lists, outperforming BAS and CS. These findings provide the first large-scale database of validated French DRM lists, demonstrate the utility of AI-generated materials for false memory research, and provide further support for theoretical accounts emphasizing the role of gist-based semantic representations in the production of false memories.</p>