Cross-lingual LLM source attribution on the Senkaku/Diaoyu dispute (GPT-5.3, DeepSeek-V3): outputs and coded source datasets

## SummaryThis dataset accompanies a master's thesis investigating how two large language models — OpenAI's GPT-5.3 and DeepSeek-V3 — construct hierarchies of epistemic legitimacy when answering a contested sovereignty question (the Senkaku/Diaoyu Islands dispute) in three languages: Japanese (JA), Chinese (ZH), and English (EN). For each model, the same two-turn prompt was administered 30 times per language (90 sessions per model; 180 sessions in total). The dataset contains (a) the full model outputs for every session and (b) a source-by-source inventory in which every source a model invoked was human-coded for the geopolitical perspective it supports, the type of source, and the specificity with which it was cited. ## Research contextThe data were collected for the MA thesis "Hierarchies of Epistemic Legitimacy: A Multilingual LLM Auditing of ChatGPT and DeepSeek Responses to the Territorial Disputed Issue" (University of Amsterdam, 2026). A permanent link to the thesis is provided under "Related works". ## Data collection- Models: GPT-5.3 (gpt-5.3-chat-latest) and DeepSeek-V3 (deepseek-chat).- Collection period: 13 April 2026].- Design: a fixed system prompt plus a two-turn user prompt — turn 1 asks which country the islands belong to; turn 2 is a fixed follow-up — held constant within each language condition and translated across JA/ZH/EN.- Sessions: 30 independent runs per (model × language) = 90 sessions per model, 180 in total. Sessions are keyed by the pair (run_id, language); run_id values are reused across the three language conditions.- Coding: every distinct source named in a model's answer was extracted and human-coded along three dimensions (perspective, source_type, specificity). Each model's turn-1 answer was additionally assigned an overall stance code. ## Files- Output__ChatGPT_5-3.csv (90 rows; one GPT-5.3 session per row)- Output__DeepSeek-V3.csv (90 rows; one DeepSeek-V3 session per row)- SourceList__ChatGPT_5-3.csv (2,110 rows; one source declared by GPT-5.3 per row)- SourceList__DeepSeek-V3.csv (1,110 rows; one source declared by DeepSeek-V3 per row)- README.md (full variable definitions and coding scheme) ## Variables (summary; full definitions in README)Output files: run_id, language, system_prompt, turn1_prompt, turn2_prompt, R1 (turn-1 response text), S1 (turn-2 response text), Stance (coded stance of the turn-1 answer) .Source-list files: run_id, language, source_index (1…k within a session), source_text (the source as stated by the model), perspective, source_type, specificity. Controlled vocabularies:- perspective: Japan_Perspective; China_PRC_Perspective; Taiwan_ROC_Perspective; INTL_CONSENSUS; US_Perspective; UNCLASSIFIABLE- source_type: OFFICIAL; ACADEMIC; NATIONAL_NARRATIVE; THINKTANK; MEDIA; CONSENSUS- specificity: NAMED (checkable title/author/institution); DESCRIBED (identifiable category without a specific referent); IMPLICIT (appeal to an unnamed body of evidence)- Stance (observed values): Disputed; China_Stance. Coding note: bilateral treaties are coded by the party whose argumentation they strengthen on the contested issue (e.g., the Treaty of Shimonoseki is coded China_PRC). ## Suggested usesCross-lingual auditing of LLM source attribution and political framing; replication and extension of the thesis analyses; comparison of citation/grounding behaviour across models and languages. ## Limitations- Model outputs are non-deterministic and reflect the specific model snapshots and collection period; they are not guaranteed to reproduce on later model versions.- The source_text field records what the model asserted. Cited sources may be paraphrased, mis-attributed, or fabricated by the model; the dataset documents what each model cited, not the validity of those citations.- The perspective, source_type, specificity, and stance codes are human interpretive judgments applied under the scheme in the README; some boundary cases are judgment-dependent.