Dataset : When the Machine Is Not Enough: An Autoethnography of AI-Assisted Amateur Research and the Boundaries of Epistemic Authority

========================================ZENODO RECORD UPDATE======================================== Title:When the Machine Is Not Enough: An Autoethnography of AI-Assisted Amateur Research and the Boundaries of Epistemic Authority Description:Replication dataset for a manuscript submitted to AI & Society. This deposit contains the analysis code, coded datasets, and tone classification results supporting an autoethnographic study of AI-assisted amateur research across three disciplinary domains. An emergency medicine physician used OpenAI's ChatGPT-4o as an epistemic collaborator for research projects in mathematics (Collatz conjecture), cosmology (Information-Topological Cosmology), and science and technology studies (Creative Singularity Triangle). All three resulting manuscripts were submitted to peer-reviewed journals and rejected. The study analyzes these failures through the lens of Collins and Evans's interactional expertise, Gieryn's boundary work, and Polanyi's tacit knowledge. The core dataset comprises 2,566 unique assistant messages across 47 research conversations, obtained after deduplication (removing 1,123 duplicates from overlapping ChatGPT exports) and filtering of non-research conversations. All messages were classified at the message level into four tone categories—encouragement-dominant (E), criticism-dominant (C), mixed (M), and neutral/informational (N)—via full census using Anthropic's Claude Opus 4 as the initial coder, with all classifications subject to author review. Key findings: The overall encouragement-to-criticism ratio is 3.6:1, varying with disciplinary verifiability—1.2:1 in mathematics (where claims are formally decidable), 5.3:1 in speculative cosmology (where falsification within a conversation is difficult), and 3.3:1 in STS (intermediate). Among evaluative messages (n = 760), 63.7% were encouragement-dominant, 18.6% mixed, and 17.8% criticism-dominant. File descriptions: 1. research_messages_deduped.csv (8.2 MB) - Deduplicated corpus of 2,566 assistant messages (plus corresponding user messages; 4,808 rows total) across 47 research conversations - Columns: assigned_project, conv_title, msg_id, role, datetime_kst, date, hour, model, text, text_length, word_count - Projects: Collatz (6 conversations, 402 messages), ITC (33 conversations, 1,746 messages), CST (8 conversations, 418 messages) 2. census_all_tones.json - Complete tone classification results: dictionary mapping each msg_id to its tone label (E/C/M/N) - 2,566 entries covering all assistant messages 3. census_review.csv - Author review file with msg_id, project, conversation title, assigned tone, and 200-character text preview - Designed for validation and reproducibility checking 4. full_census/batch_000.json through batch_051.json (52 files) - Input batches for the classification pipeline, each containing up to 50 messages with full text, project label, and conversation title 5. full_census/result_000.json through result_051.json (52 files) - Raw classification outputs from Claude Opus 4, each containing msg_id and tone label pairs Note: Raw ChatGPT conversation logs (JSON exports) are available from the corresponding author upon request. The logs contain the complete interaction history but are not deposited due to their size (>50 MB) and to avoid potential privacy concerns from incidental personal information in conversation text. Keywords:artificial intelligence, large language models, expertise, autoethnography, tone classification, peer review, RLHF, sycophancy, encouragement trap, boundary work License: CC BY 4.0 ========================================FILES TO UPLOAD (replacing old deposit)======================================== Keep:- (none from old deposit; all superseded) Upload new:1. tone_classification/research_messages_deduped.csv2. tone_classification/full_census/census_all_tones.json3. tone_classification/full_census/census_review.csv4. tone_classification/full_census/batch_000.json ~ batch_051.json (52 files)5. tone_classification/full_census/result_000.json ~ result_051.json (52 files)========================================ZENODO数据集更新======================================== 数据集标题：当机器力有不逮：AI辅助业余研究的自我民族志研究与认知权威的边界 数据集说明：本数据集为投稿至《AI & Society》期刊的学术论文的复现数据集。本存档包含支撑三项跨学科领域AI辅助业余研究的自我民族志研究的分析代码、编码数据集与语气分类结果。 一位急诊医学医师将OpenAI的ChatGPT-4o作为认知合作者，开展了数学领域（考拉兹猜想）、宇宙学领域（信息拓扑宇宙学）以及科学与技术研究领域（创造性奇点三角）的三项研究项目。三篇最终手稿均提交至同行评审期刊并被拒稿。本研究借助柯林斯与埃文斯的互动专长理论、吉尔林的边界工作理论以及波兰尼的隐性知识理论，对这些失败案例展开系统性分析。 核心数据集涵盖47项研究对话中的2566条唯一助手消息，该数据集经过去重（从重叠的ChatGPT导出结果中移除1123条重复内容）与非研究对话过滤后得到。所有消息均通过以Anthropic的Claude Opus 4作为初始编码者的全普查方式，在单条消息层面被划分为四类语气：主导型鼓励（E）、主导型批评（C）、混合型（M）以及中立/信息型（N），所有分类结果均经过作者人工审核。 核心研究结果显示：整体鼓励与批评的比例为3.6:1，该比例随学科可验证性程度呈现显著差异——在主张可形式判定的数学领域为1.2:1，在对话内难以证伪的思辨性宇宙学领域为5.3:1，在科学与技术研究（STS）领域为3.3:1（介于两者之间）。在760条评价性消息中，63.7%为主导型鼓励，18.6%为混合型，17.8%为主导型批评。 文件说明： 1. research_messages_deduped.csv（8.2 MB）：包含47项研究对话中去重后的2566条助手消息（及对应的用户消息；总计4808行）的语料库。字段包括：assigned_project、conv_title、msg_id、role、datetime_kst、date、hour、model、text、text_length、word_count。项目分类：考拉兹（6项对话，402条消息）、ITC（信息拓扑宇宙学，33项对话，1746条消息）、CST（创造性奇点三角，8项对话，418条消息） 2. census_all_tones.json：完整的语气分类结果：以每个msg_id为键映射其语气标签（E/C/M/N）的字典，包含全部2566条助手消息的条目。 3. census_review.csv：作者审核文件，包含msg_id、项目、对话标题、分配的语气标签以及200字符的文本预览，用于验证与可复现性检查。 4. full_census/batch_000.json ~ batch_051.json（共52个文件）：分类流水线的输入批次文件，每个文件最多包含50条消息，包含完整文本、项目标签与对话标题。 5. full_census/result_000.json ~ result_051.json（共52个文件）：Claude Opus 4生成的原始分类输出，每个文件包含msg_id与语气标签的配对数据。 注：原始ChatGPT对话日志（JSON导出文件）可向通讯作者申请获取。由于日志文件体积超过50 MB，且为避免对话文本中附带的偶然个人信息引发隐私问题，本存档未包含该日志。 关键词：人工智能、大语言模型（Large Language Model）、专长、自我民族志、语气分类、同行评审、人类反馈强化学习（RLHF）、谄媚倾向、鼓励陷阱、边界工作 许可证：CC BY 4.0 ========================================待上传文件（替换旧存档）======================================== 保留内容：（旧存档无保留内容；所有内容均已更新） 上传新内容： 1. tone_classification/research_messages_deduped.csv 2. tone_classification/full_census/census_all_tones.json 3. tone_classification/full_census/census_review.csv 4. tone_classification/full_census/batch_000.json ~ batch_051.json（共52个文件） 5. tone_classification/full_census/result_000.json ~ result_051.json（共52个文件）