- Issue
- Journal of Siberian Federal University. Humanities & Social Sciences. 2026 19 (11)
- Authors
- Gutovskaya, Marina S.; Skarabahaty, Vadzim V.
- Contact information
- Gutovskaya, Marina S. : Belarusian State University of Foreign Languages (Minsk, Republic of Belarus); marina‑; ORCID: 0000‑0002‑2091‑7815; Skarabahaty, Vadzim V.: Belarusian State University (Minsk, Republic of Belarus); ORCID: 0009‑0007‑1707‑5403
- Keywords
- bibliographic confabulation; ghost citation; large language models; linguistic mechanisms of generation; citation reliability rate; chimeric reference; academic discourse; thematic canon density
- Abstract
This study reconstructs the linguistic mechanisms of bibliographic confabulation using a corpus of 102 bibliographic references extracted from texts generated by five large language models (Alice AI LLM, GigaChat Ultra, ChatGPT‑5.3 Instant, Gemini 3 Flash, and DeepSeek V3) based on two human‑authored texts, a review article and a research article. The analysis shows that the weighted average citation accuracy coefficient reaches 74.5 %, with a cross‑model variance of 50.9 percentage points. The study identifies four mechanisms: analogical completion, syncretic combination of real components (five sub‑mechanisms), propositional blending, and structural‑format completion (three sub‑ mechanisms). Confabulation intensity depends on how well the relevant thematic domain is represented in a model’s training corpus; consequently, the probability of a mechanism’s realization increases as its architectural complexity decreases
- Pages
- 2320–2330
- EDN
- CGMSXR
- Paper at repository of SibFU
- https://elib.sfu-kras.ru/handle/2311/160216
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).