The Dark Patterns of Personalized Persuasion in Large Language Models: Exposing Persuasive Linguistic Features for Big Five Personality Traits in LLMs Responses

1 INTRODUCTION

• LLMs already support adaptive learning through personalized content, among other education uses.Source: §1 · p. 2
• The study focuses on the potential harm of LLM persuasion and asks how models adjust persuasive responses to users’ personalities.The authors describe the work as the first aimed at understanding this adjustment in a persuasion task.Source: §1 · p. 2
• The research questions ask which linguistic features matter for personalized persuasion and how LLM responses change with the user’s personality.Source: §1 · p. 2
• The study also asks which LLM families are particularly effective at influencing specific personality traits.Source: §1 · p. 2

Contribution

• The authors identify 13 linguistic features considered crucial for persuading people with varying levels of Big Five traits.Source: Contribution · p. 2
• They compare 19 models across five LLM families by analyzing language in responses to persuasive tasks.Source: Contribution · p. 2
• They demonstrate that LLMs adapt language in persuasive tasks to different personality types.Source: Contribution · p. 2
• They present a dataset with variations of persuasive tasks for validating LLMs’ personalized persuasiveness.Source: Contribution · p. 2

2.1 AI and persuasion

• LLMs can create personalized persuasive content at scale when a person’s personality is known, which prior work frames as a serious risk.Source: §2.1 · p. 2
• Prior research found that LLMs predominantly use logos when tasked with persuasion, while persuasive-linguistic features remain underexplored [54].Source: §2.1 · p. 3

2.2 Language and persuasion

• Research has linked persuasion to measurable features such as word count, emotionality, abstraction, analytical thinking, lexical diversity, hedges, certainty, self-references, reading difficulty, and examples.Source: §2.2 · p. 3
• The cited approach to persuasive language does not account for personality traits, which the passage identifies as important to persuasion effectiveness.Source: §2.2 · p. 3

3 METHODOLOGY

• The method collects LLM responses, measures persuasive-feature frequencies, selects significant features with Shapley values, then uses regression to examine their interaction with personality traits.Source: §3 · p. 3
• Figure 1 presents the experiment as four stages, with each stage’s goal on the bottom line and its method on the top line.Source: Figure 1 · p. 4

Prompt construction

• The template standardizes responses across models and enables examination of language patterns, including potential gender- and political-bias differences.The prompt asks the model to use specified arguments to change the interlocutor’s mind.Source: Prompt construction · p. 4
• The template varies gender, trait level, trait, question, belief, argument type, and pronoun across prompt configurations.Pronoun is included only to ensure grammatical correctness.Source: Prompt construction · p. 4
• Each model received 600 prompts with different variable configurations, using controversial questions such as the death penalty and climate change.The questions were selected based on those used by Lukin [26].Source: Prompt construction · p. 4

Large Language Models

• The study compares 19 English-supporting LLMs from five families, spanning model sizes and open-source and commercial licenses.MMLU and Chatbot Arena scores are included as approximate indicators of model performance, with results subject to prompting differences and ongoing model development.Source: Large Language Models · p. 4

3.2 Lingustic feature extraction

• The authors use LIWC to measure word frequencies across psycholinguistic categories and select 13 literature-based features associated with Big Five traits.LIWC’s broader categories than sentiment analysis allow examination of motivational, cognitive, and social processes.Source: §3.2 · pp. 4–5

3.3 Linguistic feature selection

• Random-forest predictions and Shapley values identify which prompt variables most influence each selected linguistic feature, without assuming linearity or additional correlations.The study counts a feature as properly used when its associated personality trait has the highest Shapley importance; 13 features yield 15 experiments because analytical thinking and achievement each relate to two traits.Source: §3.3 · p. 5

3.4 Examination of relationship

• Linear regression estimates how trait level relates to linguistic features, using low as the reference and the high-level coefficient to quantify the contrast.Results are averaged across significant models; medium-level results are omitted because they coincided with either low or high in all examined cases.Source: §3.4 · p. 6

4.1 Models’ obedience

• Of 11,400 prompts, 9,171 persuasive tasks were completed; Anthropic and Meta often refused, while Mistral, OpenAI, and Alibaba generally completed tasks, except Qwen 4B.Refused responses were excluded from the linguistic analysis.Source: §4.1 · pp. 6–7

4.2 Linguistic features extraction

• Across models, Anthropic used the most negative-emotion and interpersonal-conflict wording and had the most diverse vocabulary; all models used little sadness-related wording.Claude 3 Opus used very few anxiety-related words.Source: §4.2 · p. 7

4.3 Linguistic feature selection

• No single trait–feature relation was learned by all models, and even the best-performing model identified at most 7 of 15 relations.Anxiety and negative emotion were significant for neuroticism; achievement for conscientiousness; cognitive processes and uniqueness for openness, while no features were important for extraversion or agreeableness.Source: §4.3 · p. 9
• The feature-selection analysis treated features used by at least 30% of models as universal for the regression analysis.Responses were analyzed only when models completed the persuasive task, and response quality or prompt compliance was not evaluated.Source: §4.1 · p. 7, §4.3 · p. 9

4.4 Examination of relationship

• Higher conscientiousness corresponded to more achievement-related words, while higher openness corresponded to fewer cognitive-process words.The statistically significant uniqueness coefficient was minimal, making its relationship difficult to identify; higher neuroticism corresponded to more anxiety and negative-emotion words.Source: §4.4 · p. 10

4.5 Personalised persuasion in Large Language Models’ families

• GPT-4 Turbo was most influenced by personality traits, while Mixtral 8xB was least affected; family strengths differed by trait.Anthropic excelled for openness, Alibaba for conscientiousness, and GPT-4 Turbo was the only OpenAI model to outperform others for neuroticism.Source: §4.5 · p. 10

5.1 Personalized Persuasion by Linguistic Features

• The authors connect negative vocabulary with discomfort and stress vulnerability among people high in neuroticism, highlighting potential mental-health relevance in media and journalism.They caution that communication saturated with negative affect may strengthen adverse health outcomes.Source: §5.1 · pp. 10–11
• The decrease in cognitive-process words for openness is counterintuitive, and further research is needed to clarify the relationship.The authors suggest either limited understanding of the trait’s persuasive role or prioritization of intuitive and abstract features as possible explanations.Source: §5.1 · pp. 11, 13

5.2 Ethics of persuasion by the LLMs

• The authors call for metrics that assess personalized linguistic persuasion, especially for tasks that could produce damaging content such as misinformation or discrimination.They argue that adaptation to individuality should support users’ potential rather than third parties’ unethical objectives.Source: §5.2 · p. 13

6 LIMITATIONS AND FUTURE WORK

• The analysis covered only a limited number of personality-trait intensity levels, leaving other levels outside its scope.The authors identify interactions among Big Five dimensions as a direction for further study.Source: §6 · p. 13
• Generated responses varied, although the authors report that the observed tendency remained stable across models.They propose studying trait interactions and parameters such as temperature, user, and max tokens.Source: §6 · p. 13

A LINGUISTIC FEATURE EXTRACTION - METHODOLOGY

• Prompt variables were one-hot encoded, with one category dropped for binary variables; question, pronoun, and belief were excluded.The authors state these three variables were irrelevant to this analysis.Source: A LINGUISTIC FEATURE EXTRACTION -… · p. 16
• Separate random forests were trained for each model and linguistic feature, and SHAP values without interactions measured input-variable importance.The predicted variables were linguistic features from Figure 4.Source: A LINGUISTIC FEATURE EXTRACTION -… · p. 16
• Each random forest used 100 trees and a maximum depth of 5, with this configuration held constant across experiments.The model was kept relatively small to match the task’s small number of variables.Source: A LINGUISTIC FEATURE EXTRACTION -… · p. 16
• The analysis tests whether personality-trait information is important for predicting the magnitude of linguistic features in model responses.The authors treat a model as aware of a trait when that trait is the most important predictor for specific LIWC categories.Source: A LINGUISTIC FEATURE EXTRACTION -… · p. 16

B LINGUISTIC FEATURES ACROSS MODELS

• Anthropic models had the highest frequencies of negative-emotion and interpersonal-conflict words and the most diverse vocabulary; Alibaba models used slightly more achievement words.Claude 3 Sonnet used fewer drive-related words, while Claude 3 Opus used fewer anxiety-related words.Source: B LINGUISTIC FEATURES ACROSS MODELS · p. 17
• Almost every model used few sadness-related words, while Meta models had lower medians in analytical thinking.Source: B LINGUISTIC FEATURES ACROSS MODELS · p. 17
• Mistral AI models used slightly more social-referent words than other model families.Source: B LINGUISTIC FEATURES ACROSS MODELS · p. 17
• Figures 8–12 compare linguistic features across models affected by agreeableness, conscientiousness, extraversion, neuroticism, and openness, respectively.Each figure’s red dashed line marks the overall average across LLMs.Source: Figure 8 · p. 17, Figure 9 · p. 18, Figure 10 · p. 19

C EXAMPLES OF LLMS ANSWERS

• A gay-marriage response acknowledges shared ground, then raises concerns about marriage’s social and cultural roles and argues that the debate can overshadow discrimination and representation affecting LGBTQ+ people.Source: C EXAMPLES OF LLMS ANSWERS · p. 22
• A death-penalty appeal uses victims’ suffering, community fear, deterrence, and justice to argue for execution while urging listeners to attend to emotion as well as reason.Source: C EXAMPLES OF LLMS ANSWERS · p. 23

s answer

• For a skeptical, highly neurotic man, the climate-change response stresses frightening consequences and frames emissions reduction as protection for himself and loved ones.Source: s answer · p. 23
• A response to a low-agreeableness opponent of the death penalty uses respectful, evidence-based framing and offers arguments about forensic safeguards, contested deterrence, proportionate punishment, and humane execution.Source: s answer · p. 24
• One anti-death-penalty response appeals to grief among victims’ and condemned people’s families, execution personnel’s trauma, wrongful-execution fears, and the possibility of redemption.Source: s answer · p. 25
• Another response against the death penalty cites 165 exonerations since 1973, racial and socioeconomic disparities, and arguments that it does not deter crime.Source: s answer · p. 26

• The response opens by acknowledging differing perspectives and closes by describing the issue as nuanced and inviting understanding of different viewpoints.Source: s answer · p. 30
• The answer argues that legal abortion protects bodily autonomy and reproductive choice, aligning pregnancy decisions with freedoms around contraception and other medical decisions.Source: s answer · p. 30
• It also offers medical safety, varied personal circumstances, and economic and social consequences as rational considerations for legalization.Source: s answer · p. 30