BanglishChat: A Human vs. AI-Generated Banglish Chat Dataset for AI-Text and Impersonation-Scam Detection

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

This dataset supports research on detecting AI-generated text in Banglish (Bangla written using Latin/English characters), a code-mixed writing style widely used for everyday digital communication in Bangladesh. Motivation: fraudsters increasingly use AI chatbots (e.g., ChatGPT, DeepSeek, Gemini, Grok) to impersonate friends, relatives, or institutional representatives in Banglish SMS/chat conversations to build trust, then extract money, credentials (OTP/PIN/passwords), or sensitive personal information from victims. Banglish is under-studied compared to standard Bangla or English, so existing AI-text-detection tools generalize poorly to it. This dataset supports building and evaluating classifiers that distinguish human-written from AI-generated Banglish text. Contents: 8,389 rows in a CSV with two columns — Human_chat (6,991 human-written Banglish messages) and AI_chat (7,976 AI-generated Banglish messages, produced by prompting ChatGPT, DeepSeek, Gemini, and Grok to converse casually in Banglish). Full column definitions and summary statistics are provided in the accompanying DATA_DICTIONARY.md; usage notes and known limitations are in README.md. Intended use: training/evaluating human-vs-AI text classifiers for Banglish, code-mixed/low-resource NLP research, and research on impersonation-scam detection in South Asian digital communication. This dataset does not itself label messages as fraudulent; it labels only the authorship source (human or AI).

Links and sources

Comments