Can AI-Generated Persuasion Be Detected? Persuaficial Benchmark and AI vs. Human Linguistic Differences

AI summary

Overview

• Research area: Natural Language Processing — automatic persuasion detection, LLM-generated content detection, and computational stylometry.
• Technical level: Intermediate. The detection experiments use standard zero-shot binary classification with F1 scores, and the linguistic analysis uses effect sizes (Cohen's d) and Wilcoxon signed-rank tests; familiarity with these is helpful but not required.
• Scope: The paper asks whether LLM-generated persuasive text is harder to detect automatically than human-written persuasion, and introduces Persuaficial, a six-language benchmark of roughly 65,000 generated persuasive texts, alongside the first large-scale linguistic comparison between human and machine persuasion.

What This Paper Is About

Large language models can produce persuasive text at scale, which raises concerns about propaganda, manipulation, and other harmful uses. Prior work had studied how well LLMs detect persuasive language and how persuasive LLMs can be, but no prior work had asked whether automatically detecting AI-generated persuasion is more difficult than detecting human-written persuasion. The authors build a controlled multilingual benchmark, Persuaficial, and use it to test detection difficulty across generation styles, then analyse what linguistically separates machine from human persuasive writing.

Key Contributions

1. Persuaficial benchmark: A new multilingual persuasion benchmark of approximately 65,000 texts generated with four controllable generation approaches across four LLMs, covering English, German, Polish, Italian, French, and Russian.
2. First detectability comparison of human vs. AI persuasion: An analysis across four LLMs and 16 controllable generation settings (4 models × 4 approaches) testing whether LLM-generated persuasion is harder to detect than human-written persuasion in a zero-shot setting.
3. First linguistic characterization of AI vs. human persuasion: An analysis covering 196 distinct linguistic features, contrasting human-written and LLM-generated English persuasive texts.
4. Released resources: The codebase, dataset, and all prompts are released (github.com/ArkadiusDS/Persuaficial).

Main Findings

• Detection difficulty depends on the generation approach. Paraphrased persuasion is only marginally harder to detect than human-written texts (on average 0.67% lower F1). Open-ended generation is on average 9.75% easier to detect, and rewriting with intensified persuasion is on average 5.33% easier.
• Subtle persuasion is the hard case. Rewriting for subtle persuasion reduces detection performance substantially, by 20.42% on average across English datasets and detectors. For example, on the SemEval 2023 sample, GPT 4.1 Mini scores 0.7398 F1 on human-written texts but 0.4031 on subtly rewritten texts; Llama 3.3 70B drops from 0.7459 to 0.4577.
• Open-ended generation is the easiest case. When models are prompted to generate persuasive text freely, detectors perform best; on SemEval data GPT 4.1 Mini reaches 0.8964 F1 (up 21% over human), and on non-English data open-ended generation frequently yields F1 scores above 0.9.
• Patterns hold across languages and detectors. The same trends appear in German, French, Italian, Polish, and Russian, suggesting the effects generalize across domains and are independent of which LLM is used as the detector.
• AI persuasive texts show higher lexical diversity and content density. Features such as L_CONT_T (proportion of unique content-word forms relative to total tokens), LTOKEN_RATIO_LEM (ratio of unique lemmas to total tokens), and L_CONT_A (proportion of content-word tokens) tend to be higher in AI text — for instance Cohen's d of 1.41 for L_CONT_T under open-ended generation with GPT 4.1 Mini.
• AI persuasive texts repeat words less and use fewer function words. Lower ST_REPET_WORDS and lower L_FUNC_A (proportion of function words) are signals of AI authorship; higher function-word use points toward human writing.
• Punctuation differs. AI texts, particularly Llama-generated ones, have a lower density of periods, while commas (L_PUNCT_COM) and dashes (L_PUNCT_DASH) appear more frequently.
• Complex syntax is more human. Marked syntactic constructions such as inversions (SY_INV_PATTERNS) are rarer in AI text and more typical of human persuasive writing.
• Generation style shapes style markers. Intensified persuasion shows more comparative and superlative adverbs (L_ADV_COMPARATIVE, L_ADV_SUPERLATIVE); subtle persuasion shows more modal hedging with "might" (VT_MIGHT) and more narrative/declarative framing (SY_NARRATIVE); open-ended generation relies more on imperative and infinitival constructions.
• Statistical significance. Wilcoxon signed-rank tests confirmed that all twenty top features per scenario show significant distributional differences between human-written and AI-generated persuasive texts.
• Quality of the generated data. Pre-generation human evaluation on 200 English summary sentences found about 91.2% accuracy. Post-generation evaluation of 400 generated English texts found 88.2% overall accuracy when all three criteria (factual faithfulness, persuasiveness, instruction faithfulness) had to be satisfied jointly, rising to 97.69% for persuasion-related criteria alone.

Methodology in Plain English

The authors started from three human-annotated persuasion datasets: SemEval 2023 Task 3, the DIPROMATS 2024 Task 1 dataset, and ChangeMyView (3,051 conversations). For each of six languages they sampled 1,000 passages, half persuasive and half non-persuasive; German was the exception, with only 420 non-persuasive texts available, so all were used alongside 580 persuasive ones.

Each persuasive passage was fed to four LLMs — Gemma 3 27b it, Llama 3.3 70B, Gemini 2.0 Flash, and GPT 4.1 Mini — under four controlled prompts: paraphrase the passage preserving persuasion, rewrite it so the persuasion becomes more subtle, rewrite it to intensify persuasion, or (after summarizing the passage into a short factual sentence) generate a persuasive text freely. This produced 16 generation configurations and roughly 24,000 English texts plus about 41,000 non-English texts.

Detectability was then tested by having four LLMs classify texts as persuasive or not in a zero-shot setting with temperature set to 0, balanced across classes, and scored with F1. For the linguistic analysis, each human text and its AI counterpart were encoded with the open-source StyloMetrix tool into a 196-dimensional feature vector. Using 1,500 matched pairs (500 per dataset across three datasets), the authors computed paired Cohen's d per feature per model and generation approach, and ran Wilcoxon signed-rank tests to check whether the distributional shifts were statistically meaningful.

Why This Matters

Research impact. The paper reframes AI persuasion detection as a question about generation method rather than a single binary property of "AI text," and provides a controlled multilingual benchmark plus the first stylometric comparison of human and AI persuasion. It also shows that detection performance, not just text style, is shaped by prompt design.

Real-world applications:

• Content moderation: platforms can anticipate that subtly rewritten AI persuasion will slip past current detectors and prioritise it for review.
• Journalism and fact-checking: newsrooms can use the linguistic markers reported here (lexical diversity, content-word density, low function-word use, fewer inversions) as interpretable signals rather than opaque model scores.
• Media literacy and public awareness: the finding that subtle persuasion is the hardest to detect is directly relevant to how AI-assisted political messaging is communicated about.
• Detection tool development: StyloMetrix's interpretable, reproducible feature vectors offer a transparent alternative to black-box classifiers for persuasion analysis.

Industry relevance. Trust-and-safety teams, LLM providers, and policy researchers all depend on knowing where detection breaks down. The result that a single prompt — "make the persuasion more subtle" — cuts F1 by 20.42% on average across detectors has direct implications for the reliability claims of any deployed persuasion-detection system.

Future Directions

• Extending the linguistic analysis beyond English, since the authors restricted it to English because of limited high-quality persuasion-annotated data in other languages.
• Building detectors that are robust specifically to subtly persuasive generation, since this setting caused the largest and most consistent degradation.
• Investigating why subtle and intensified generation produce opposite effects, and whether the prompting strategy itself can be inferred from text features.
• Using Persuaficial as a training or evaluation resource for more interpretable detection models that operate on linguistically grounded features rather than opaque representations.

Target Audience

NLP researchers working on persuasion, propaganda, or AI-generated text detection; computational social scientists studying AI influence and communication; trust-and-safety and content-moderation practitioners who need to know the limits of current detection systems; and linguists or stylometricians interested in measurable differences between human and machine writing. The paper is also useful for readers who want an interpretable, feature-level account rather than a purely black-box detection result.

Authors’ abstract

Large Language Models (LLMs) can generate highly persuasive text, raising concerns about their misuse for propaganda, manipulation, and other harmful purposes. This leads us to our central question: Is LLM-generated persuasion more difficult to automatically detect than human-written persuasion? To address this, we categorize controllable generation approaches for producing persuasive content with LLMs and introduce Persuaficial, a high-quality multilingual benchmark covering six languages: English, German, Polish, Italian, French and Russian. Using this benchmark, we conduct extensive empirical evaluations comparing human-authored and LLM-generated persuasive texts. We find that although overtly persuasive LLM-generated texts can be easier to detect than human-written ones, subtle LLM-generated persuasion consistently degrades automatic detection performance. Beyond detection performance, we provide the first comprehensive linguistic analysis contrasting human and LLM-generated persuasive texts, offering insights that may guide the development of more interpretable and robust detection tools.