DeepfakeWatch

Deepfake technology is a threat to information integrity across social media platforms. Understanding how the public views and talks about these manipulations produced by AI is critical in the development of methods for countering them. However, manually analysing thousands of comments is labour-intensive and requires special knowledge. This dataset allows one to analyze the public stance and discourse patterns around deepfake content. I have extracted 2351 English comments from 41 YouTube videos that are about deepfakes between July 2025 and January 2026. The classification of each comment was done using BART-large-mnli for the detection of the stance and identification of the claims. Comments are assigned to four different types of stance: believesreal (where the user had concerns about threats from deepfake use), believesfake (where the user is skeptical), uncertain (where the user is doubtful), and metadiscussion (the user's general commentary about the situation). Six types of claims were identified: detectionliteracy (spotting techniques), voicefraud (audio cloning), warpropaganda (military context), celebrity (impersonation), scamfinance (financial fraud), and electionpolitics (political manipulation). Topic modeling using NMF identified 8 discussion themes (0.53 coherence), much better than LDA alternatives. We were able to track temporal changes over seven months using the divergence measure called Jensen-Shannon divergence, and we found empirical evidence of significant change points for discourse patterns. Privacy is ensured with SHA-256 hashing of any and all identifiers. This dataset serves these research areas of computational social science, misinformation detection, and content moderation. Researchers can use it for training models for the detection of stance or analyzing the evolution of discourse or the public awareness of AI-generated content. 1. Primary Dataset: Records: 2,351 comments Format: CSV Columns: 18 (video_id, comment_id_hash, month, like_count, spam_score, stance_label, stance_conf, claim_type, claim_conf, topic_id_nmf, topic_p0-p7_nmf) 2. Stance Distribution: meta_discussion: 1,264 (53.8%) uncertain: 484 (20.6%) believes_fake: 360 (15.3%) believes_real: 243 (10.3%) 3. Claim Type Distribution: detection_literacy: 748 (31.8%) voice_fraud: 565 (24.0%) war_propaganda: 416 (17.7%) celebrity: 373 (15.9%) scam_finance: 137 (5.8%) election_politics: 112 (4.8%) 4. Topic Model: Algorithm: NMF (k=8) Coherence: 0.5310 Features: TF-IDF vectors Topics: AI-generated content, fake news debates, war propaganda, military authenticity, social discourse, Ukraine narrative, bot accusations, South Asian discussions 5. Quality Metrics: Duplicates: 0 Missing values: 0 Spam flagged: 9.45% (retained with scores) Video coverage: 32.8% (41/125 videos) 6. Documentation: README.md: Complete methodology and usage guide CODEBOOK.md: All variable definitions and distributions深度伪造（Deepfake）技术对全球社交媒体平台的信息完整性构成严重威胁。明晰公众对这类人工智能生成的篡改内容的看法与讨论，对于研发对抗此类威胁的方法至关重要。然而，人工分析海量评论不仅耗时费力，还需具备专业领域知识。本数据集可用于分析公众围绕深度伪造内容所持的立场及其话语模式。 本研究从2025年7月至2026年1月期间发布的41个关于深度伪造的YouTube视频中，提取了2351条英文评论。使用BART-large-mnli模型完成每条评论的分类，用于立场检测与主张识别。评论被划分为四类立场：believesreal（即用户担忧深度伪造应用带来的威胁）、believesfake（即用户持怀疑态度）、uncertain（即用户存疑）、metadiscussion（即用户就该议题展开的一般性评论）。共识别出六类主张：detectionliteracy（识别伪造技术）、voicefraud（音频克隆诈骗）、warpropaganda（军事相关宣传）、celebrity（名人模仿）、scamfinance（金融诈骗）、electionpolitics（政治操纵）。 通过非负矩阵分解（NMF）进行主题建模，共识别出8个讨论主题（一致性得分为0.53），其效果优于潜在狄利克雷分配（LDA）模型。本研究借助詹森-香农散度（Jensen-Shannon divergence）这一离散度指标，追踪了七个月内的话语时序变化，并为话语模式存在显著变化点提供了实证依据。所有标识符均通过SHA-256哈希（SHA-256 hashing）进行处理，以保障用户隐私。 本数据集适用于计算社会科学、虚假信息检测与内容审核等研究领域。研究人员可利用该数据集训练立场检测模型，或分析话语演化进程与公众对人工智能生成内容的认知水平。 1. 基础数据集 记录量：2351条评论 格式：CSV格式 字段数：共18项，包括video_id（视频ID）、comment_id_hash（评论ID哈希值）、month（发布月份）、like_count（点赞数）、spam_score（垃圾信息评分）、stance_label（立场标签）、stance_conf（立场置信度）、claim_type（主张类型）、claim_conf（主张置信度）、topic_id_nmf（NMF主题ID）、topic_p0-p7_nmf（NMF各主题概率值） 2. 立场分布 metadiscussion：1264条（占比53.8%） uncertain：484条（占比20.6%） believes_fake：360条（占比15.3%） believes_real：243条（占比10.3%） 3. 主张类型分布 detection_literacy：748条（占比31.8%） voice_fraud：565条（占比24.0%） war_propaganda：416条（占比17.7%） celebrity：373条（占比15.9%） scam_finance：137条（占比5.8%） election_politics：112条（占比4.8%） 4. 主题模型 算法：NMF（主题数k=8） 一致性得分：0.5310 特征提取方式：TF-IDF向量 主题包括：人工智能生成内容、虚假新闻辩论、战争宣传、军事真实性辨析、社会话语讨论、乌克兰叙事、机器人账号指控、南亚地区讨论 5. 质量指标 重复数据：0条 缺失值：0条 垃圾信息标记率：9.45%（保留所有评分数据） 视频覆盖范围：32.8%（共纳入125个视频中的41个） 6. 文档说明 README.md：包含完整的研究方法与使用指南 CODEBOOK.md：包含所有变量定义与分布情况