When AI invents evidence: evaluating citation accuracy in generative AI-generated academic writing

Objective Generative AI increasingly supports academic writing, literature synthesis, and knowledge generation, but inaccurate, incomplete, or fabricated citations may introduce unsupported evidence. This study evaluated the accuracy and reliability of citations generated by four contemporary models across computer science, health and medicine, education, and social sciences. A controlled cross-model experiment used 20 standardized academic prompts submitted independently to GPT-5.6, Gemini 3.1 Pro, Claude Opus 4.6, and Grok 4.5; generated references were independently checked against bibliographic and publisher records using predefined verification criteria. Results description The prompts generated 80 responses: 42 (52.5%) were satisfactory, 32 (40.0%) contained citation-related problems, and 6 (7.5%) produced no usable references. Satisfactory-response proportions were highest for GPT-5.6 (65.0%), followed by Claude Opus 4.6 (60.0%), Grok 4.5 (55.0%), and Gemini 3.1 Pro (30.0%). Problems included incomplete references, invalid or mismatched digital object identifiers, bibliographic inconsistencies, inaccessible links, and topic mismatches. Differences between models were not statistically significant (χ²(6) = 7.68, p = 0.263). Problems occurred across all models, indicating that fluent citation generation does not ensure reliable referencing and that AI-generated references require independent verification before academic use.