RedNote-Vibe: A Longitudinal AIGT Dataset

• RedNote-Vibe is a comprehensive 5-year dataset from Xiaohongshu that captures detailed textual, temporal, and engagement signals for both human and AI-generated content.
• It organizes posts into human, AI-generated, and exploration splits, allowing controlled comparison of stylistic and engagement differences across time periods.
• The accompanying PLAD framework employs interpretable psycholinguistic features to detect AI-generated text, achieving competitive benchmark results against transformer models.

RedNote-Vibe is a longitudinal dataset for analyzing AI-generated text on social media, introduced as the first 5-year dataset for social media AIGT analysis and constructed from Xiaohongshu posts spanning January 2020 through July 2025 (Li et al., 26 Sep 2025). It is designed to capture not only textual content but also timestamps, domain labels, and user-engagement signals such as likes, comments, and collections, thereby enabling joint study of linguistic form, temporal dynamics, and interaction patterns. The same work also introduces the PsychoLinguistic AIGT Detection Framework (PLAD), an interpretable detection approach based on psycholinguistic features, and uses the dataset to examine how human- and AI-generated content differ in both style and engagement outcomes (Li et al., 26 Sep 2025).

1. Dataset definition and longitudinal design

RedNote-Vibe is sourced from Xiaohongshu (RedNote), described as a Chinese “lifestyle sharing” social-media platform with over 300 M monthly active users (Li et al., 26 Sep 2025). A web crawler retrieved 120 k notes using 10 topical categories and approximately 50 tags/domain, after which automatic topic filtering with Qwen-2.5-7B reduced the corpus to 98,714 notes. Each note includes title, body text, user-provided tags, publication timestamp, engagement counts for likes, comments, and collections, and a domain label drawn from ten categories (Li et al., 26 Sep 2025).

The dataset is explicitly longitudinal. Its time span begins in the pre-LLM period and extends into the period of widespread LLM deployment, allowing comparison across historical phases rather than only within a single static snapshot. This temporal structure is central to the dataset’s purpose: it supports analysis of the evolution of AIGT prevalence and the changing relationship between linguistic style and social-media engagement.

The corpus is organized into three splits. The Human-Written Set consists of all notes published before November 2022 and is treated as guaranteed human-authored. The AI-Generated Set is created by asking 17 modern LLMs, including OpenAI GPT-o3, Google Gemini-2.5, Anthropic Claude, and Alibaba’s Qwen3, to produce stylistic paraphrases of pre-LLM notes under a seed-based protocol that preserves length, tone, and colloquial errors while avoiding copying; each LLM produced at least 1,000 samples. The Exploration Set contains all post-LLM notes from 2023–2025 and remains unlabeled, serving temporal and engagement analyses rather than supervised ground truth (Li et al., 26 Sep 2025).

This split design distinguishes synthetic, controlled AI generation from real-world post-LLM social-media production. A plausible implication is that RedNote-Vibe is structured to separate detection benchmarking from ecological analysis: the human and AI sets provide ground truth, whereas the exploration set preserves uncertainty about authorship in actual platform use.

2. Corpus composition and engagement characteristics

RedNote-Vibe covers ten topical categories, including Career, Wellness, Travel, and others, with domain-level engagement statistics reported in the dataset summary (Li et al., 26 Sep 2025). For example, the Health domain contains 14.9 k posts with average length 398 characters and mean engagement values of 383.3 likes, 38.8 comments, and 322.0 collections, whereas Fashion contains 11.4 k posts with average length 273 characters and mean engagement values of 718.6 likes, 40.9 comments, and 265.9 collections. Food contains 7.9 k posts with average length 427 characters and mean engagement values of 49.0 likes, 11.2 comments, and 35.5 collections (Li et al., 26 Sep 2025).

The engagement distribution is heavy-tailed. Approximately 50% of posts have total engagement, defined as likes plus comments plus collections, less than or equal to 10, while a small tail of viral posts has very high engagement (Li et al., 26 Sep 2025). This matters methodologically because downstream analyses of social response cannot be interpreted under a Gaussian or homoscedastic assumption; the reported use of log-transformed boxplots for engagement comparisons is consistent with this distributional asymmetry.

A common misunderstanding in social-media AIGT research is to treat content prevalence and content impact as the same phenomenon. RedNote-Vibe separates them: prevalence is studied through temporal estimates of AI-classified content in the exploration set, whereas impact is studied through engagement metrics attached to each note. This separation is integral to the dataset’s design and underlies much of its analytical value (Li et al., 26 Sep 2025).

3. PLAD: psycholinguistic detection framework

The PsychoLinguistic AIGT Detection Framework (PLAD) is the companion detection method introduced alongside RedNote-Vibe (Li et al., 26 Sep 2025). PLAD uses 31 features organized into four psychological dimensions: Emotional & Social Grounding, Cognitive Architecture, Lexical Identity & Stylistic Signature, and Cohesion & Textual Flow. The examples given for these dimensions include emoji density and interactive/dialogic stance; perspectival complexity and axiological coherence; type-token ratio, prosodic rhythm consistency, and “Imperfection”; and phrasal repetition frequency and semantic progression (Li et al., 26 Sep 2025).

Feature extraction combines direct statistical measures with semantic or psychological ratings produced by a proxy LLM under chain-of-thought rubrics and then verified by a secondary chain-of-thought check (Li et al., 26 Sep 2025). Two explicit statistical definitions are provided. Emoji density is given as the ratio of the number of emoji to total tokens, and type-token ratio is defined as

$\frac{|V|}{N},$

where $|V|$ is the number of unique word types and $N$ is the total number of tokens (Li et al., 26 Sep 2025).

PLAD represents each input note as a 31-dimensional feature vector $\mathbf f(x)\in\mathbb R^{31}$. The classifiers considered are tree-based models—XGBoost, CatBoost, and Gradient Boosting—selected for interpretability and built-in feature importance. The prediction rule is

$$\hat y = \arg\max_{y\in\mathcal Y} P(y\mid \mathbf f(x)),$$

and training uses cross-entropy loss,

$\mathcal L_{\mathrm{CE}} = -\sum_i y_i \log p_i,$

with one-hot labels $y_i$ and predicted probabilities $p_i$ (Li et al., 26 Sep 2025).

The framework is notable less for proposing an opaque end-to-end detector than for making the feature space inspectable. This suggests that PLAD is intended not only to classify text but also to expose stylometric and psycholinguistic regularities associated with AI generation in social-media settings.

4. Experimental tasks and benchmark results

The experimental program evaluates three tasks: binary AIGT classification (Human vs. AI), 6-way AI provider identification, and 17-way AI-model identification (Li et al., 26 Sep 2025). Reported macro-averaged metrics are Precision, Recall, and Accuracy, with F1 additionally used in ablation analyses. The definitions are standard:

$$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}},$$

and

$$\mathrm{F1} = 2 \times \frac{\mathrm{Precision}\times\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}.$$

Against baselines including BERT-base, RoBERTa-base, and an LLM-Idiosyncrasies method, PLAD + CatBoost achieves the strongest overall results in several settings (Li et al., 26 Sep 2025). On the 17-way task it reports 35.87 precision, 36.45 recall, and 36.94 accuracy, outperforming the other listed methods on all three metrics. On the 6-way task it reports 50.06 precision, 47.34 recall, and 51.89 accuracy, while PLAD + XGBoost attains the highest reported 6-way precision and accuracy at 50.73 and 53.30 respectively. On the binary task PLAD + CatBoost reaches 88.70 precision, 87.28 recall, and 89.62 accuracy; RoBERTa-base attains 88.88 recall and 89.52 accuracy, and LLM-Idiosyncrasies attains 90.15 recall (Li et al., 26 Sep 2025).

Zero-shot generalization is also reported. When GPT-o3 and Gemini-2.5 are excluded from training, PLAD recalls unseen models at 56.3%–58.5%, compared with 25%–52% for BERT-base (Li et al., 26 Sep 2025). This indicates that the psycholinguistic feature space retains some model-agnostic signal beyond memorization of seen generators.

The results do not imply universal superiority of PLAD on every metric and every task. Rather, they show that an interpretable psycholinguistic approach can match or exceed transformer baselines in several benchmark conditions while preserving feature-level interpretability.

5. Temporal dynamics of AIGT on Xiaohongshu

Temporal analysis is conducted on the unlabeled Exploration Set covering 2023–2025 (Li et al., 26 Sep 2025). The daily fraction of posts classified by PLAD as AI versus human is smoothed using 14-day and 30-day rolling averages, and the smoothed proportion of AI content is plotted over approximately 600 days. This design emphasizes trend estimation rather than individual-post certainty.

A simple linear regression of AI proportion $|V|$0 on day index $|V|$1 is reported as

$|V|$2

indicating a long-term upward slope despite short-term plateaus that coincide with known platform governance interventions (Li et al., 26 Sep 2025). The coexistence of upward trend and local plateaus is important: it suggests that platform policy may modulate visible prevalence without fully reversing the broader diffusion of AI-assisted or AI-generated posting.

This temporal analysis should not be read as ground-truth measurement of all post-2023 authorship states. The exploration set is unlabeled by design. What is measured is the proportion of posts classified as AI by PLAD after smoothing, not directly observed authorship. The paper’s methodology therefore supports temporal inference about detected AIGT dynamics, while leaving room for future semi-supervised or active-learning annotation of real-world AI content (Li et al., 26 Sep 2025).

6. Linguistic signatures, engagement patterns, and interpretability

PLAD’s ablation study identifies the Lexical dimension as the most consequential component: removing it causes the largest drop in F1, by 9–13 points across tasks (Li et al., 26 Sep 2025). CatBoost feature-importance analysis ranks the top five features by gain as Prosodic Rhythm Consistency (17.6), Type–Token Ratio (13.2), Imperfection (5.2), Interactive & Dialogic Stance (4.8), and Emoji Density (3.0) (Li et al., 26 Sep 2025).

The reported linguistic contrast between human and AI text is correspondingly specific. AI text is characterized as “too perfect,” with high Prosodic Rhythm Consistency—uniform sentence lengths and flows—and near-zero Imperfection, interpreted as lack of disfluencies. It also exhibits higher lexical diversity, measured by Type–Token Ratio, but lower Social Grounding, reflected in fewer emojis and less dialogic stance (Li et al., 26 Sep 2025). These observations tie the detector’s empirical performance to interpretable stylistic patterns rather than only to latent classifier behavior.

Engagement comparisons by domain are conducted using $|V|$3-transform boxplots, and Mann–Whitney U tests show that human posts have significantly higher median likes, comments, and collections, with $|V|$4 in most domains (Li et al., 26 Sep 2025). Human posts outperform AI in median and tail engagement, especially in emotionally intensive domains such as Travel and Relationships. Appendix SHAP analysis further decomposes engagement associations: likes correlate most with surface features such as Punctuation Ratio and Word Frequency Entropy; collections correlate with deeper features such as Word Entropy, Phrasal Repetition, and Axiological Coherence; comments depend heavily on social-grounding features such as Empathetic Engagement and Interactive Stance (Li et al., 26 Sep 2025).

The paper also reports that “AI-Augmented” authors, defined as mixing human and AI content, achieve the highest engagement, suggesting a complementary synergy (Li et al., 26 Sep 2025). A plausible implication is that platform success may be shaped not by a strict human-versus-AI dichotomy but by hybrid authoring practices that preserve social grounding while exploiting generative assistance.

7. Applications, scope conditions, and open problems

Several applications are explicitly identified. RedNote-Vibe can support longitudinal monitoring of AIGT adoption and platform policy impact, interpretable detection tools for content moderation and authenticity labeling, and study of the human–AI co-evolution of style in social media (Li et al., 26 Sep 2025). Because the dataset couples text, metadata, and engagement, it also provides a basis for research that connects authorship signals to downstream social reception rather than treating detection as an isolated classification problem.

The work also states clear limitations. It is currently Chinese-only and focused on Xiaohongshu, so cross-platform and multilingual extensions are needed. The Exploration Set remains unlabeled, motivating semi-supervised or active-learning approaches for annotating real-world AI content. More advanced temporal models such as ARIMA and change-point detection are suggested as future directions, and multimodal signals including images and video are identified as necessary for reflecting the full “note” experience on Xiaohongshu (Li et al., 26 Sep 2025).

An important misconception is that RedNote-Vibe is merely a static benchmark with synthetic positives. The dataset includes such synthetic positives for controlled supervised evaluation, but its distinctive contribution is the combination of a five-year time span, engagement-enriched metadata, and an unlabeled post-LLM exploration partition for real-world temporal analysis (Li et al., 26 Sep 2025). In that sense, RedNote-Vibe occupies an intermediate position between benchmark corpus, observational social-media archive, and interpretable AIGT analysis framework.