---
title: 'PLAD: PsychoLinguistic AI Text Detection'
url: https://www.emergentmind.com/topics/psycholinguistic-aigt-detection-framework-plad
type: topic
---

# PLAD: PsychoLinguistic AI Text Detection

PsychoLinguistic AIGT Detection Framework (PLAD) is an interpretable framework for detecting AI-generated text (AIGT) in social media, introduced together with the RedNote-Vibe dataset for Xiaohongshu/RedNote. PLAD represents each post as a 31-dimensional psycholinguistic feature vector and classifies it with tree-based models such as Gradient Boosting Classifier, XGBoost, and CatBoost. Its stated purpose is not only binary human-versus-AI attribution, but also provider and fine-grained model identification, as well as downstream analysis of how linguistic signatures relate to social-media engagement and temporal diffusion of AIGT [2509.22055].

## 1. Conceptual framing and research problem

PLAD is motivated by a problem formulation in which social-media AIGT detection differs from standard detection on news, essays, or other relatively static text domains. In the RedNote-Vibe study, social media is described as dynamic, engagement-driven, temporally evolving, and stylistically noisy, with content shaped by audience interaction, trends, and platform governance. On that account, AIGT detection is treated not as a one-off binary classification task on clean text, but as a longitudinal, socially situated detection problem in which interpretability is necessary because the same linguistic evidence is also used to study engagement effects and temporal dynamics [2509.22055].

The framework is also positioned against two limitations in prior detector design. Black-box neural classifiers, especially transformer-based systems, may yield strong predictive performance but typically do not explain why a post is classified as AI-generated. Classical stylometric detectors, by contrast, are characterized as relying on relatively shallow statistics that may miss psychological and communicative properties relevant to authentic social-media writing. PLAD therefore adopts psycholinguistic features grounded in emotion, cognition, social interaction, style, and discourse. The underlying premise is that human social-media writing carries signatures of autobiographical grounding, social awareness, irregular cognitive production, and evolving discourse structure, whereas AI-generated text tends to be smoother, more uniform, and less socially grounded [2509.22055].

## 2. RedNote-Vibe and the supervised detection setting

PLAD is developed and evaluated on RedNote-Vibe, described as the first longitudinal five-year dataset for social-media AIGT analysis. The source platform is Xiaohongshu/RedNote, and data collection begins from ten dominant content categories in RedNote’s official behavior report: Career, Wellness, Travel, Health, Food, Pets, Education, Sports, Fashion, and Relationships. The authors manually expand tags to around 50 representative tags per domain, crawl 120,000 notes from January 2020 to July 2025, and retain 98,714 notes after topic filtering with Qwen-2.5-7B. Each retained sample includes title, text, tags, publication timestamp, likes, comments, collections, and topic domain [2509.22055].

To obtain labeled AIGT for supervised learning, the study generates parallel AI-written versions of pre-LLM human posts. Seed notes are restricted to content published before November 2022 so that they can be treated as human-authored. Each seed note is provided to an LLM together with title, content, and domain; the prompt asks the model to imitate style and tone, preserve personal characteristics such as colloquialisms, punctuation, and occasional grammatical imperfections, match length, and avoid direct copying. The model pool contains 17 LLMs from 6 providers, with each model generating at least 1,000 samples. The paper also defines an exploration set of real post-LLM-period posts from 2023 to 2025 without ground-truth labels; that set is used for temporal and engagement analysis rather than supervised evaluation [2509.22055].

RedNote-Vibe supports three supervised tasks. The first is binary human-versus-AI classification. The second is 6-way provider identification. The third is 17-way model identification. The paper reports macro-averaged precision, recall, and accuracy, and compares PLAD against statistics-based methods such as StyloAI, Binoculars, and Ullah et al., and against model-based methods including BERT-base, ALBERT-base, RoBERTa-base, Sniffer, POGER, and LLM-Idiosyncrasies. At the same time, the benchmark is not fully documented as a time-aware generalization setup: exact train/validation/test split ratios, random seeds, optimization settings, and a clearly temporal supervised split are not specified in the provided text [2509.22055].

## 3. Computational architecture and inference pipeline

PLAD is explicitly presented as a feature-based supervised classification pipeline rather than a deep end-to-end neural architecture. The input text is converted into a 31-dimensional feature vector,
$$
\mathbf{f}(x)\in\mathbb{R}^{31},
$$
where each component is a psycholinguistic or structural measurement. The paper states that feature extraction proceeds through two routes. One route consists of directly computable statistical features such as emoji density, type-token ratio, and other structural measures. The second route consists of semantically based features derived through a proxy LLM used for psychological text analysis. For these features, the authors construct evaluation rubrics that translate theoretical constructs into measurable criteria and then prompt a proxy LLM to score the text on those dimensions. To reduce proxy-model bias, a verification mechanism using Chain-of-Thought reasoning is applied [2509.22055].

PLAD is not a direct LIWC implementation. The paper explicitly contrasts its method with classic LIWC-style dictionary counting and instead characterizes the representation as a mixture of custom structural/statistical features, lexical diversity and entropy measures, emotion and social markers, discourse and cohesion features, and higher-order cognitive or value-oriented dimensions inferred via LLM-based rubric scoring. The exact identity of the proxy model, the rubric prompts, the scoring scales, and the full inventory of all 31 features are not specified in the provided text, although the paper states that detailed feature lists and example criteria appear in an appendix [2509.22055].

Once the feature vector is extracted, PLAD trains a tree-based classifier. The evaluated back-ends are Gradient Boosting Classifier, XGBoost, and CatBoost. The intended prediction rule is
$$
\hat{y}=\arg\max_{y\in\mathcal{Y}} P(y\mid \mathbf{f}(x)),
$$
and training uses cross-entropy loss,
$$
\mathcal{L}_{\text{CE}}=-\sum_i y_i \log(p_i).
$$
At inference time, the workflow is straightforward: compute the 31 psycholinguistic features for a post, feed them into the trained tree-based classifier, and output either a binary human-versus-AI label, a provider label, or a fine-grained model label, depending on the task [2509.22055].

## 4. Psycholinguistic organization of the feature space

PLAD organizes its representation into four feature families, described as “dimensions of human language expression.” The main paper does not enumerate all 31 features, but it does identify a set of named features and associated conceptual targets.

| Dimension | Theoretical emphasis | Named features in the paper |
|---|---|---|
| Emotional and Social Grounding | autobiographical memory, Theory of Mind, social connection | Interactive and Dialogic Stance, Emoji Density, Empathetic Engagement |
| Cognitive Architecture | perspectival complexity, dialectical reasoning, temporal coherence | Perspectival Complexity, Axiological Coherence |
| Lexical Identity and Stylistic Signature | idiolect, irregularity, self-correction, lexical uniqueness | Prosodic Rhythm Consistency, Type-Token Ratio, Imperfection, Lexical-Stylistic Personalization, Word Frequency Entropy, Sentence Burstiness, Punctuation Ratio |
| Cohesion and Textual Flow | semantic progression, repetition, discourse-level cohesion | Phrasal Repetition Frequency |

The first dimension, Emotional and Social Grounding, is linked to emotional intensity, personal grounding, social connection, and markers such as emoji usage patterns. It is motivated by the claim that authentic social posts are often rooted in interpersonal orientation and lived experience. The second dimension, Cognitive Architecture, targets conceptual complexity, narrative structure, nuanced argumentation, and tolerance for ambiguity; the paper describes it in terms of perspectival complexity, dialectical reasoning, and temporal coherence. The third dimension, Lexical Identity and Stylistic Signature, is described as the most empirically important family and is motivated by idiolect, lexical uniqueness, stylistic consistency, and imperfections generated by real-time cognition. The fourth dimension, Cohesion and Textual Flow, concerns how ideas are organized across the whole text and includes semantic progression, discourse-level cohesion, and repetition-related effects [2509.22055].

Because the full 31-feature inventory is not listed in the provided text, the representation remains only partially recoverable. Still, the paper and appendix fragments identify at least the following named features: Emoji Density, Interactive and Dialogic Stance, Empathetic Engagement, Perspectival Complexity, Axiological Coherence, Prosodic Rhythm Consistency, Type-Token Ratio, Imperfection, Lexical-Stylistic Personalization, Word Frequency Entropy, Phrasal Repetition Frequency, Punctuation Ratio, and Sentence Burstiness. The text also mentions emotional intensity, personal grounding, social connection, dialectical reasoning, temporal coherence, semantic progression, and repetition patterns as feature concepts, while leaving most operational definitions unspecified [2509.22055].

## 5. Detection performance and interpretability

PLAD is evaluated on binary detection, 6-way provider identification, and 17-way model identification. The strongest reported results are summarized below.

| Task | Best PLAD variant | Reported result |
|---|---|---|
| Binary human vs AI | CatBoost | Precision 88.70, Accuracy 89.62 |
| 6-way provider identification | XGBoost | Precision 50.73, Recall 48.77, Accuracy 53.30 |
| 17-way model identification | CatBoost | Precision 35.87, Recall 36.45, Accuracy 36.94 |

In binary classification, PLAD with CatBoost attains the highest reported precision and accuracy, while LLM-Idiosyncrasies attains the highest recall at 90.15, and RoBERTa-base remains competitive at 89.52 accuracy. In 6-way provider identification, XGBoost is best overall, with CatBoost reported as similar. In 17-way model attribution, CatBoost exceeds the strongest baselines listed in the paper, including RoBERTa-base and LLM-Idiosyncrasies. The study therefore presents PLAD as superior or consistently competitive, especially on harder attribution tasks where psycholinguistic features are argued to capture provider- and model-specific stylistic traces [2509.22055].

A more stringent robustness check is the provider-level zero-shot setting in which GPT-o3 and Gemini-2.5 are excluded during training but tested at inference while other models from the same providers remain in the training set. In that setup, PLAD substantially outperforms BERT-base on recall for unseen target models: for GPT-o3, PLAD achieves 56.34 zero-shot recall versus 25.35 for BERT; for Gemini-2.5, PLAD achieves 58.46 versus 52.31. On seen-model accuracy, the results are mixed but still competitive: 46.46 versus 42.42 in the GPT-o3-held-out setting, and 43.72 versus 45.23 in the Gemini-2.5-held-out setting. The paper interprets this as evidence that PLAD captures family-level style imprints more robustly than fine-tuned neural baselines [2509.22055].

The ablation study supports the four-way decomposition of the feature space. Using CatBoost, the full system reaches macro-F1 scores of 36.16 on the 17-way task, 48.66 on the 6-way task, and 87.98 on binary classification. Removing Lexical Identity and Stylistic Signature causes the largest degradation, reducing the scores to 26.81, 38.66, and 85.00 respectively. Removing Emotional and Social Grounding also causes a large drop, especially on 17-way classification. This places lexical/stylistic regularity at the center of the learned discrimination, with social-emotional grounding as the next strongest cluster [2509.22055].

Interpretability is one of the main reasons the paper favors tree-based models. The feature-importance analysis ranks Prosodic Rhythm Consistency first with importance 17.6, Type-Token Ratio second with 13.2, and Imperfection third with 5.2. Emotional and social features also enter the top tier: Interactive and Dialogic Stance has importance 4.8 and Emoji Density 3.0. The paper’s summary characterization is that PLAD identifies AI content as being “too perfect, too uniform, and lack of social characteristics.” The appendix adds a complication: some prose describing higher values of “imperfection” suggests the feature may be coded in the direction of stable fluency rather than literal imperfection, but the broader interpretation remains that AI text appears more regular, more uniform, and less idiosyncratic than human text, even when prompted to imitate social-media style [2509.22055].

## 6. Engagement modeling and temporal dynamics

PLAD is also used as an analytic feature space for studying social-media response. For likes, collections, and comments, the paper trains separate CatBoost regressors with default hyperparameters and interprets them with SHAP. This is explicitly described as associational feature-importance analysis rather than causal inference. The main substantive result is that different engagement metrics align with different portions of the PLAD feature space. Likes are driven most by Punctuation Ratio and Word Frequency Entropy, suggesting sensitivity to surface readability and rhythm. Collections are most influenced by Word Frequency Entropy, Phrasal Repetition Frequency, and Axiological Coherence, implying stronger association with content richness and internal value consistency. Comments depend not only on Punctuation Ratio but also on Lexical-Stylistic Personalization, Empathetic Engagement, and Interactive and Dialogic Stance, indicating that deeper engagement is tied to interpersonal and dialogic features [2509.22055].

At the content level, the paper finds that human-authored posts have higher median engagement and more high-engagement outliers than AI-generated posts across the six largest topical domains. Engagement distributions are log-transformed because of heavy tails. The strongest human advantage appears in domains such as Travel, Career, and Relationships, which the authors interpret as domains where personal experience and emotional resonance matter. Fashion is identified as an exception with more similar patterns. The paper also notes that AI posts receive lower absolute engagement but show a flatter decline from likes to comments, suggesting relatively more discussion per like in some cases [2509.22055].

At the author level, 829 authors with at least four posts are partitioned into Traditional Authors, AI-reliant Authors, and AI-augmented Authors according to the proportion of AI-generated posts. The reported shares are 68.7%, 27.7%, and 3.5% respectively. The AI-augmented group attains the highest engagement across all three metrics, which the authors interpret as evidence that strategic human-AI collaboration may outperform purely human or purely AI production [2509.22055].

The temporal analysis applies PLAD to the unlabeled exploration set to estimate the proportion of AI-generated content over time. Across the last 600 days, the trend shows a steady increase, with reported linear slope \(= 0.012\) per day and \(0.355\) per month, smoothed by 14-day and 30-day rolling averages. Growth becomes clearer after mid-2024, with plateaus and dips that the paper says coincide with platform-governance interventions reported in news sources. The appendix also includes a pairwise feature-correlation analysis over all \(N=465\) feature pairs: the average absolute Pearson correlation is 0.1884 with standard deviation 0.2290; only 6 pairs exceed \(|r|>0.7\), 89 fall in \(0.3<|r|<0.7\), and 370 are weakly correlated with \(|r|<0.3\). The authors use this to argue that the 31-feature representation is relatively orthogonal and non-redundant [2509.22055].

## 7. Limitations, misconceptions, and research context

Several limitations are explicit or strongly implied in the published description. First, PLAD is platform-specific: it is developed on Xiaohongshu/RedNote, a platform centered on lifestyle sharing and personal experience, so the same features may behave differently on Twitter/X, Reddit, or formal writing. Second, the setting is language- and culture-dependent, since some cues—emoji use, punctuation habits, colloquialisms, discourse markers—are likely specific to this Chinese social-media environment. Third, the exploration set is unlabeled, so temporal prevalence estimates depend on PLAD itself. Fourth, the supervised AI data are synthetic paraphrastic variants generated from human seeds, which may not fully match real-world AI-assisted posts. Fifth, feature drift is a serious concern because newer models may better mimic human irregularity and social grounding. Sixth, some features rely on a proxy LLM evaluator, so the system may inherit that model’s bias or instability. Seventh, the paper does not present a formal adversarial robustness study against deliberate injection of “human-like” imperfections, emojis, or discourse markers [2509.22055].

There are also several recurrent misconceptions that the paper itself helps to correct. PLAD is not a deep end-to-end neural architecture; it is a feature-based supervised pipeline with tree-based classifiers. It is not a classic LIWC implementation; instead it combines directly computed stylometric or structural signals with LLM-scored psycholinguistic variables. And although it is introduced together with a longitudinal dataset, its main supervised benchmark is not a fully documented temporal generalization test. The longitudinal contribution is stronger in descriptive analysis of estimated AI prevalence and engagement than in formal year-sliced robustness evaluation [2509.22055].

Within the broader AIGT-detection literature, PLAD occupies an interpretable, psycholinguistically oriented position rather than a probability-only, robustness-only, or latent-disentanglement position. POGER recovers pseudo white-box generation-probability features through proxy-guided efficient re-sampling and is explicitly characterized as complementary rather than psycholinguistic [2402.09199]. DP-Net emphasizes dynamic perturbation training for cross-domain generalization and adversarial robustness [2504.21019], while DRGD focuses on disentangling AI-detection semantics from generator-aware artifacts for unseen-generator generalization [2604.13692]. PADBen shows that detectors often fail in an intermediate paraphrase-laundering region, especially for authorship obfuscation [2511.00416]. FourierGPT proposes a spectrum of relative likelihood with psycholinguistic roots in surprisal and information-density theories [2406.19874]. REVEAL moves in a different direction by generating explicit reasoning chains before classification [2604.19172]. A separate conceptual line of work maps stylometric features to cognitive constructs such as lexical retrieval, discourse planning, cognitive load management, and metacognitive self-monitoring, offering a broader theoretical backdrop for interpretable AI-text detection [2505.01800].

In that landscape, PLAD’s distinctive contribution is the claim that social-media AIGT can be detected and analyzed through an interpretable 31-feature psycholinguistic representation that is simultaneously useful for attribution, engagement analysis, and temporal monitoring. Its strongest empirical evidence lies in fine-grained attribution, zero-shot provider-level generalization to unseen models, and the interpretability of feature importance. Its main unresolved issues are reproducibility, platform transfer, and adversarial robustness [2509.22055].

Source: https://www.emergentmind.com/topics/psycholinguistic-aigt-detection-framework-plad