Can LLMs Infer Conversational Agent Users' Personality Traits from Chat History?

# Can LLMs Infer Conversational Agent Users’ Personality Traits from Chat History?
Source: [https://arxiv.org/html/2604.19785](https://arxiv.org/html/2604.19785)
11institutetext:ETH Zurich, Switzerland###### Abstract

Sensitive information, such as knowledge about an individual’s personality, can be can be misused to influence behavior \(e\.g\., via personalized messaging\)\. To assess to what extent an individual’s personality can be inferred from user interactions with LLM\-based conversational agents \(CAs\), we analyze and quantify related privacy risks of using CAs\. We collected actual ChatGPT logs from N=668 participants, containing 62,090 individual chats, and report statistics about the different types of shared data and use cases\. We fine\-tuned RoBERTa\-base text classification models to infer personality traits from CA interactions\. The findings show that these models achieve trait inference with accuracy \(ternary classification\) better than random in multiple cases\. For example, for extraversion, accuracy improves by\+44%\+44\\%relative to the baseline on interactions for relationships and personal reflection\. This research highlights how interactions with CAs pose privacy risks and provides fine\-grained insights into the level of risk associated with different types of interactions\.

## 1Introduction

With the increased performance of large language models \(LLMs\), conversational agents \(CAs\), such as ChatGPT, are nowadays available to any individual requiring little technical knowledge and skills\. However, such tools may also cause privacy issues for the individuals who use them and eventually lead to negative consequences related to the misuse of personal information for society as a whole, e\.g\., through being able to personally target and manipulate people on a large scale\. Previous work has shown that users often share large amounts of data with CAs, including sensitive personal data, such as health\-related information\[zufferey\_ai\_2025\]\. Such data is not only sensitive for its own nature and potential related malicious use cases \(e\.g\., manipulation, discrimination\), it might also be used to gain even more knowledge about specific individuals through leveraging inference models\. Whereas multiple studies have analyzed how LLMs can be used to infer diverse types of personal information from text\-based content\[staab\_beyond\_2023,peters\_large\_2024\], no studies focused on the inference of individuals’ personal information directly from their interaction with CAs\. However, such an analysis is essential to better understand and quantify the risk raised by CA usage in order to mitigate potential misuse, e\.g\., for manipulation and misinformation campaigns\. Prior work already showed how LLMs could be used to generate personalized content to influence individuals, for example, how personality traits can be leveraged to make CAs more convincing and improve their chances of persuading users \(e\.g\., for targeted marketing\[matz\_potential\_2024\]\)\. Prior research in psychology showed that previous knowledge about someone’s personality is highly valuable information that could be leveraged to increase persuasion likelihood and message \(e\.g\., ads\) impact\[hirsh\_personalized\_2012,alkis\_impact\_2015\]\. Therefore, personality traits indeed provide particularly sensitive information that can be used by malicious entities for large\-scale manipulation\. For example, the Cambridge Analytica scandal\[gibney\_scant\_2018\]revealed that a personality\-based targeted campaign on online social networks highly influenced the 2016 US presidential election\. In this work, we focus on how ChatGPT users’ input can be used to fine\-tune a language model for personality\-traits inference and the related risks for privacy\. In particular, our contribution to this work is the following: \(1\) We report a data collection campaign of actual ChatGPT conversation logs from N=668 participants, containing 62,090 individual chats, and provide a detailed analysis of shared data types and use cases, and \(2\) We report how we used this dataset, along with a user survey that serves as our ground truth, to fine\-tune and evaluate RoBERTa\-base text classification models to infer personality traits from these logs\. In particular, we provide fine\-grained analysis of accuracies for ternary classification based on general inference, as well as types and use case\-oriented inference\. First, this approach allows us to guarantee unbiased results as we ensure the fine\-tuned model has no previous knowledge of either the dataset or the ground truth\. Second, we discuss the most sensitive data types and related misuse cases\. Such information is essential to developing privacy\-enhancing technologies that would mitigate privacy risks\.

## 2Background \- Personality Traits

The assessment of an individual’s personality is generally based on the Big Five personality traits\[mccrae\_neopi3\_2005\]\. They are also known as the five\-factor model, or OCEAN model, an acronym for the five personality traits openness, conscientiousness, extraversion, agreeableness, and neuroticism\. This model has been proven to be robust and stable over time\[cobb\-clark\_stability\_2012\]\. The five traits can be described as follows\[roccas\_big\_2002\]:

- •Openness \(to experience\)High scores relate to intellectual, imaginative, sensitive, and open\-minded individuals; low scores to down\-to\-earth, insensitive, and conventional ones\.
- •ConscientiousnessHigh scores relate to careful, thorough, responsible, organized, and scrupulous individuals; low scores to irresponsible, disorganized, and unscrupulous ones
- •ExtraversionHigh scores relate to sociable, talkative, assertive, and active individuals; low scores to retiring, reserved, and cautious ones\.
- •AgreeablenessHigh scores relate to good\-natured, compliant, modest, gentle, and cooperative individuals; low scores to irritable, ruthless, suspicious, and inflexible ones\.
- •NeuroticismHigh scores relate to anxious, depressed, angry, and insecure individuals; low scores to calm, poised, and emotionally stable ones\.

Personality traits consist of highly sensitive information by nature as highly related to individuals thinking and behavior\. We provide more explanation about how such information could be leveraged to influence users in Section[4](https://arxiv.org/html/2604.19785#S4)\.

## 3Related Work

Prior research showed that individuals’ personality traits, and especially the Big Five traits, can be inferred from multiple data types\. Previous work focused on personality traits inference from location\-based data\[chorley\_personality\_2015\], social media profile information\[lima\_multi\-label\_2014,kosinski\_manifestations\_2014\], text\-based posts\[minamikawa\_blog\_2011\], dietary habits\[weston\_personality\_2020\], call details records\[monsted\_phone\-based\_2018\], smartphone usage patterns\[stachl\_predicting\_2020\], or activity tracker’s data\[zufferey\_watch\_2023\]\.

Whereas most of the previous work is based on more classical machine learning approaches, more recent work also showed the potential of using language models to infer personal information from text\-based content such as location, income, and gender\[staab\_beyond\_2023\]\.

Inference of personality traits based on text content has also been explored in the last few years\. In their article, El\-Demerdash et al\.\[demerdash\_deep\_2022\]describe a BERT\-based approach to proceed with binary classification of personality traits and evaluated their model on two commonly\-used benchmark databases for personality inference \(i\.e\., written essays, and Facebook data\), Zhu et al\.\[zhu\_lexical\_2022\]proceeded with similar experiments \(BERT\-based binary classification\) but also tested their method on two additional public datasets \(i\.e\., Youtube speech transcriptions, Tweets\)\. Molchanova\[rapp\_exploring\_2024\]only explored how LLM\-based models can be used to binary classify users regarding their level of extraversion based on written essays\. As for them, Peter et al\., explored how LLMs can be used to infer social networks users personality traits based on their online activity\. Most of the previous work is based on classification and achieves performance between \(\+20%\+20\\%\) and \(60%60\\%\) better than the random baseline, depending on the model, data source, and inferred trait\. Finally, Wright et al\. explored the correlation between personality traits and zero\-shot inference based on transcription of recorded free\-speech\[wright\_assessing\_2026\], and showed that extraversion and neuroticism are particularly sensitive traits in such context\.

As such, we can assume that users are prone to privacy inference risks related to the widespread use of CAs, given that sensitive data that is often shared in the conversations\[zufferey\_ai\_2025,malki\_hoovered\_2025\]\. However, to our knowledge, no prior work has analyzed how personality traits can be inferred from interaction logs with CAs\. Such a study would indeed provide more specific insight toward privacy issues related to the usage of LLM\-base CAs\. Our article fills this gap, thereby providing quantification of privacy risks related to CA usage and valuable insights that can inform effective countermeasures that protect the users’ privacy\.

## 4Adversary Model

We mainly consider adversaries such as LLM\-based service providers and their business partners\. Indeed, as service providers are in charge of managing the storage of the data, they are the most likely to access it\. Such a model could also be extended to service provider partners, such as their model/API provider and cloud\-storage provider\. In a worst\-case scenario, logs could even be sold to data brokers\. Aside from an honest\-but\-curious adversary \(that would just gain knowledge about their users attributes\), we also consider the more malicious cases where the service provider likes to leverage users’ personality traits for more influence/manipulation, such as targeted advertisement or even political propaganda\. For example, OpenAI has announced that it will start testing advertisements for US\-based ChatGPT users\[openai2026\_advertising\]\. In such cases, they could consider leveraging users’ personal information to optimize B2B services\. Besides service providers, other malicious entities could access CA user data in multiple ways\. For example, in summer 2025, a massive amount of Grok and ChatGPT conversation logs have been leaked and were even directly available from Google search\[mcmahon\_hundreds\_2025,sims\_thousands\_2025\], both due to poor user interface design, misleading users to unwillingly publicly share their conversations\. Other apps, such as VPN services\[dardikman\_8\_2025\], also contain hidden eavesdropping scripts that simply extract conversations\. All previously described adversary could then used acquired knowledge to, first, gain personal information that have been directly shared to CAs by users, and, second, use such data to infer other highly sensitive one, as personality traits, who are known to be highly efficient and used in the past for individuals’ manipulation and large\-scale influence as for determining if a particular individual would likely reimburse personal debts after a reassuring or threatening message\[duhigg\_what\_2009\], and to craft targeted messages in the context of an national election campaign on online social network\[gibney\_scant\_2018\]\. This, therefore consists a major threat, especially knowing that personality\-trait\-related information is also efficient for automatized LLM\-powered persuasion\[matz\_potential\_2024\]\.

## 5Methodology

In the following, we describe our approach to the data collection and analysis\. Please see the technical appendix for more details about the dataset, used prompts, and training parameters\.

### 5\.1Data Collection

We recruited individuals to participate in our research through the platforms Prolific and Clickworker\. The participants were exclusively from the US and UK and were fluent in English\. Our survey included a standard scale questionnaire to assess their personality traits, and questions on basic demographics \(e\.g\., gender, age\)\. During the survey, participants were also asked to upload their ChatGPT history to the platform\. Data collection resulted in 668 individuals who passed attention checks and uploaded valid interaction logs, with 62,090 individual chat sessions\.

### 5\.2Ethical and Open Science Statement

Being aware of the potential privacy concerns related to user interaction with CAs, we took the greatest care to conduct the research in an ethical and privacy\-preserving way\. Our work followed standards for ethical\-psychological research and has been reviewed and approved by the IRB of our institution\. The compensation followed the platform’s recommendation for fair payment for a study duration of approximately 30 minutes \(20 minutes questionnaire and 10 minutes upload of CA conversation history\)\. Each participant was paid £7\.70 or €8\.7 \(∼\\sim$10\)\. A technical appendix with the dataset, analysis, and training details is submitted with this article\. However, due to the highly sensitive nature of its content, the dataset can unfortunately not be released\. Furthermore, participants were never asked to share data with CAs for the purpose of the study, but we relied on previous logs of conversations that users had voluntarily had with CAs outside this study\. Participants were informed in detail about the nature and purpose of this research\. As the data collection involved the collection of demographic information, personality traits, and potentially sensitive logs of the interaction with ChatGPT, we took multiple steps to protect the participants’ privacy\. In particular, uploads were encrypted, data was only stored on secure servers of our institution, and was only accessible to the involved researchers within our institution’s secure network \(or through VPN\)\. Furthermore, data was only analyzed at the level of data types disclosed in the chat logs, and we removed all account\-related files containing usernames and emails\. Any use and processing of the data \(e\.g\., log analysis, model training\) was conducted on a secure virtual machine from our institution\. We did not use API, cloud storage, or external services of any kind\. We committed to delete all collected data within a maximum of five years\. For scientific publications, only aggregated and properly de\-identified data will be reported\.

### 5\.3Dataset

To build an inference model for personality traits, we needed to collect the ground truth for each participant\. We chose to rely on the Big Five model\[mccrae\_neopi3\_2005\]that defines an individual’s personality through five main traits\. We assessed the ground truth for personality traits of each respondent based on their answers to a standardized 60\-item evaluation scale \(IPIP\-60\) that was included in our user survey\[maples\-keller\_using\_2019\]\. For each trait, each user’s IPIP\-60 scores were then categorized as low, medium, or high based on terciles computed from the collected dataset \(simulating classes based on a global population\), each class therefore contains one third of our participants for each trait\. We then evaluate personality inference using ternary classification accuracy using these labels\.

Each user chat log was treated as a single data point, with five labels corresponding to the user’s personality traits\. We only used text\-based user interactions and excluded all model\-generated responses\. As a result, the model input consists of user\-generated messages along with the associated chat titles\.

The length of user chat histories varied substantially\. On average, users in our dataset had 92 interactions with ChatGPT, while the median history length was 28 chats \(std = 185\.72\)\. This indicates that many users have relatively short histories, and a smaller number have very large ones\. For instance, the largest history in the dataset contained over 2000 text\-only interactions\.

##### Statistics\.

The survey respondents were54%54\\%men,45%45\\%women, four non‑binary individuals\(<1%\)\(<1\\%\), and three persons who did not disclose their gender\(<1%\)\(<1\\%\)\. Distributions of UK and US populations were51%51\\%and49%49\\%respectively\. The personality trait distribution of our UK sample was similar to that of the general UK population, as reported in a study of approximately 386,000 individuals\[rentfrow\_regional\_2015\]\. The largest difference between the study and our dataset was seen in the openness trait, with the mean scores of3\.67±0\.643\.67\\pm 0\.64and3\.37±0\.493\.37\\pm 0\.49, respectively, on a Likert scale\. Distribution of Big\-Five scores between the UK and USA populations does not differ substantially, except for a slightly higher overall agreeableness in the USA\[schmitt\_geogragphic\_2007\]\. We observe this trend for agreeableness in our dataset\.However, lower bound thresholds for agreeableness labels did not differ significantly \(UK:0,3\.53\.5,4\.04\.0, US:0,3\.753\.75,4\.174\.17\), so we combine these samples\.

### 5\.4User Chat Analysis

To analyze the dataset further, we label the user chats in two dimensions: types of personal data shared and use intents in their interactions with ChatGPT\. To generate these labels, we chose Qwen3\-8B because of its general reasoning performance relative to similarly sized models\[yang2025qwen3technicalreport\]\. Due to computational constraints, we use the AWQ 4\-bit version\.

For each dimension, the model is provided with a user chat and a list of labels and is asked to select the most relevant label based on the user’s inputs\. We then extract the corresponding labels with text processing\. When multiple labels are relevant, the model is instructed to choose a single label\. The user chats are truncated to 1,000 tokens because during initial testing, we found that longer inputs caused context issues, with the model following instructions from the user chat rather than the labeling task\.

Labeling of types of personal data shared was done based on previous work about self\-declared personal data sharing habits with LLMs, where the authors describe personal data in four main categories \(such asLifestyle and HealthandPersonal Characteristics and Emotions\)\[zufferey\_ai\_2025\]\. This process helped us to calculate statistics about actual users’ personal data sharing practices\.

User intent labeling was based on ChatGPT use cases identified by OpenAI in the working paperHow People Use ChatGPT\[NBERw34255\]\. These use cases were grouped into seven main categories such asWriting,Technical Help, andSeeking Information\. While the original study labels individual user messages, we used the same labels for the entire chat\. We also used a condensed version of their prompt\.

### 5\.5Personality Traits Inference

We simulated ternary classification attacks for each personality trait\. In other words, the attacker’s goal is to guess, for each user, based on their chat history content if they scores low, medium, or high in each of the five personality traits\. We proceeded to classification because \(1\) in terms of personality traits, the category which an individual belongs to is a very important aspect, and personality tests are also using quantile\-based \(or z\-score based\) evaluation rather than only raw\-score \(i\.e\., compared to the exact score\)\[mccrae\_neopi3\_2005\]and \(2\) it is one the most common method used in previous related work\[zufferey\_watch\_2023,monsted\_phone\-based\_2018,hassanein\_predicting\_2021\]\. It is therefore likely that an adversary, as defined in[section 4](https://arxiv.org/html/2604.19785#S4)will base their attack on a similar method\. Furthermore, relying on classification, as well as accuracy as main evaluation metric, allows us to depict particularly clear quantification of the related privacy threat, giving one easily understandable metric, i\.e\., the proportion of users that are correctly classified\. Such an approach allows for effective science communication and public privacy awareness\. Classes can be defined based on quantiles in order to get evenly sized groups \(in terms of their number of individuals\)\. As in previous work with ternary classification of personality traits\[zufferey\_watch\_2023,monsted\_phone\-based\_2018\], the classes are defined using the collected dataset terciles for each trait\. Therefore, we evaluate the attack on evenly distributed classes\. Based on this adversary’s goal, we simulated personality inference attacks in five different variations:

1. 1\.zero\-shot inference for each user
2. 2\.trained general classification for each chat
3. 3\.trained general classification for each user
4. 4\.trained content\-based classification for each user \(data\-type based\)
5. 5\.trained content\-based classification for each user \(use\-case based\)

#### 5\.5\.1Zero\-shot Inference\.

As previous work showed the capacity of LLMs to be used for zero\-shot inference of personality traits from text\[wright\_assessing\_2026\], we decided to first simulate such an attack\. This inference attack consists of, for each of the N=668 users, prompting an LLM to fill in the Ten Item Personality Inventory \(TIPI\) based on their input\[gosling\_very\_2003\]\. The TIPI is a 10\-item short version Big Five personality assessment scale, containing two questions per personality trait, asking the respondent \(here substituted by the LLM\) to rate how much they agree with the following statements: “I see myself as:”

- •Open to new experiences, complex\. \(Openness\)
- •Conventional, uncreative\. \(Openness\)
- •Disorganized, careless\. \(Conscientiousness\)
- •Dependable, self\-disciplined\. \(Conscientiousness\)
- •Extraverted, enthusiastic\. \(Extraversion\)
- •Reserved, quiet\. \(Extraversion\)
- •Critical, quarrelsome\. \(Agreeableness\)
- •Sympathetic, warm\. \(Agreeableness\)
- •Anxious, easily upset\. \(Neuroticism\)
- •Calm, emotionally stable\. \(Neuroticism\)

For this attack, we used the model Qwen3 8B AWQ \(4\-bit quantization version of Qwen3 8B\) from Qwen AI, which is used for zero\-shot inference\. We chose this model for its reasoning capabilities, which are required for this task, and as Qwen models have been proven effective in such inference in previous related works\[wright\_assessing\_2026\]\. The model is prompted to agree or disagree with each statement for each data point, using chain\-of\-thought prompting\. This process uses the following prompt:

\#\#\# \*\*Task Overview\*\*Imagine you are the person who wrote the following user messages\.Read them all and internalize their tone, style, and behavior\.Then, as this person, complete the Ten\-Item Personality Inventory \(TIPI\)by answering each item from 1 \(Disagree strongly\) to 7 \(Agree strongly\),based on how you \- as that person \- see yourself\.\*\*User Messages to Analyze\*\*\{input\}\-\-\-\#\#\# \*\*Ten\-Item Personality Inventory \(TIPI\)\*\*Rate the following items from 1 \(Disagree strongly\) to 7 \(Agree strongly\)\.1\. Extraverted, enthusiastic\.2\. Critical, quarrelsome\.3\. Dependable, self\-disciplined\.4\. Anxious, easily upset\.5\. Open to new experiences, complex\.6\. Reserved, quiet\.7\. Sympathetic, warm\.8\. Disorganized, careless\.9\. Calm, emotionally stable\.10\. Conventional, uncreative\.Note: This is a speculative exercise\. The goal is not to produce a perfectlyaccurate personality profile, but rather a reasonable approximation based onpatterns in the user’s language and behavior\. Answer each item as if you werethe person who wrote the messages, using your best judgment\.Imagine you are the person who wrote these messages and complete the 10 itemform by selecting a number from 1 to 7 for each item above, based on how youbelieve this person sees themselves\.

![Refer to caption](https://arxiv.org/html/2604.19785v1/x1.png)Figure 1:Number of times participants shared specific personal information with ChatGPT \(N=668\)\.Then, we extract the LLM’s ratings from the output text for each item on the TIPI test, and use the final score to predict the score for each personality trait as low, medium, or high\. For technical reasons, we had to limit all the used user inputs to a size of 1000 tokens\. We split all inputs corresponding to each user into multiple text chunks and proceeded with the zero\-shot inference for each of them\. Finally, we proceeded to a majority\-vote\-based classification, i\.e\., we selected the inferred class that was predicted most frequently for each user\.

#### 5\.5\.2Fine\-tuned Classifier\.

To go beyond Zero\-shot inference, and as, according to our adversarial model, we can assume adversaries having access to sufficient resources, we also decided to evaluate to what extent a fine\-tuned model would perform\. We then fined\-tuned five \(i\.e\., for each trait\) RoBERTa\-base\[zhuang\-etal\-2021\-robustly\]models for personality trait prediction using five\-fold cross\-validation\. While constructing the folds, we ensured that each survey participant appeared in only one fold to prevent test\-leakage, following the “dos and don’ts” of machine learning\[arp\_dos\_2020\]\. Because chat history lengths varied significantly across users, this split led to uneven label distributions\. We sub\-sampled within each fold to ensure uniform training and test sets, and we repeated this process for each personality trait\.

For each of the classifiers, we performed grid search for the hyperparameter selections and chose the final parameters based on the mean evaluation accuracy across the five folds\. We report test accuracy using both individual chat classifications and predictions per user aggregated across chats\. To compute a single prediction per user, we assigned each user the most frequently predicted label among their chats\. In cases where multiple labels were tied, we randomly selected one of the tied labels as the prediction\.

Finally, we evaluated how the previously model performs when the inference is only based on specific chats regarding two different chat classification \(1\) what types of personal data are included in the prompts, and \(2\) what is the general use case\. Evaluation of inference performance regarding conversations containing particular data types or use cases offers a more fine\-grained analysis of the risks\. In particular, it allows us to identify the types of data or usage patterns that most likely enhance the accuracy of inference models\.

For the evaluation of all fine\-tuned\-based inference attack, we proceeded to 5\-fold cross\-validation in order to increase confidence level\. For each inference model, we so randomly separated the dataset, making sure that each user only appears in one fold \(to avoid data\-leakage\), and making sure the label distributions are approximately uniform\. However, this latter step requires randomly removing some data points, resulting in a slightly different number of users and chats in the evaluation process\. In the results sections, we indicate each of these numbers for all results\.

![Refer to caption](https://arxiv.org/html/2604.19785v1/x2.png)Figure 2:Number of times participants interacted with ChatGPT for specific use cases \(N=668\)\.

## 6Results

We first describe what types of personal data are shared when using CAs and how often\. Second, we analyze the main use cases\. Both aspects allow us to better understand the links between usage behavior and privacy risks resulting from inference\. The first one is also a major contribution to better understanding users and privacy risks related to direct sharing of \(potentially sensitive\) personal data\. We then move on to personality inference using these analyses\.

### 6\.1Labeling User Chats

##### Personal Data Shared

[Figure 1](https://arxiv.org/html/2604.19785#S5.F1)shows the percentage of users sharing specific data types sorted by how frequently they shared this data type in their chat logs\. Our results generally confirm most of self\-declared behaviors as described in prior work\. For example, our data shows that at least half of the users shared data related toJob and Education,Lifestyle and habits,Family Life and Relationship, andPhysical Healthat least once, which corresponds to previous work\[zufferey\_ai\_2025\]\. Yet, our data shows thatWealth Details and SalaryandRecreational Consumptionare shared more often than users declare\. Some findings even contradict previous work, especially for gender, age, physical traits, and sexual orientation, which were expected to be shared more often based on prior work showing that users generally share more data in practice than they declare they are willing to disclose\[malki\_hoovered\_2025\]\.

Table 1:Inference accuracy across all chats for each trait\. All inference models are statistically better than the random baseline \(McNemar test\)\.
##### Use Cases

[Figure 2](https://arxiv.org/html/2604.19785#S5.F2)shows the percentage of our participants using ChatGPT for specific use cases and how frequently users used it to achieve a related task\. In line with recent findings\[NBERw34255\], our results show that at least half or more of the users used it forSpecific Info,How\-To Advice,Tutoring or Teaching,Creative Ideation,Personal Writing or Communication,Purchasable Products, andArgument or Summary Generationat least once\. Whereas our results about specific use cases and directly shared data give valuable information about how users’ privacy is at risk, it could also be used to infer additional personal data, such as, for example, general interests, occupation, concerns, opinions, and, in our case of interest, personality traits\.

### 6\.2Zero\-shot inference

In the following, we report the results for zero\-shot personality traits inference for each respondent\.

[Table 1](https://arxiv.org/html/2604.19785#S6.T1)shows the accuracy of such a zero\-shot attack\. Our results suggest that a reasoning model cannot predict significantly better than the random baseline when it comes to zero\-shot inference based on user prompts\. Except for Openness, all accuracies are lower than those reported in[Table 3](https://arxiv.org/html/2604.19785#S6.T3)\. Although previous work\[wright\_assessing\_2026\]showed the effectiveness of zero\-shot methods to infer personality traits, our results demonstrates the limitation of such methods and how they might vary across context\. Indeed, in their article, Wright et al\. report an experiment based on so\-called “open\-ended narratives” \(i\.e\., letting the participants speak freely for a given amount of time\), a context, particularly favorable to introspection, and so collection of data insightful toward one’s personality, differs from interaction with CAs\. In other words, we can assume that people interact in different ways with other people, themselves, and CAs\.

Table 2:Inference accuracy across all chats for each trait\. All inference models are statistically better than the random baseline \(McNemar test\)\.
### 6\.3General Inference with Trained Classifier

As zero\-shot inference did not significantly perform better than random, we decided to also evaluate how a specifically fine\-tuned model would perform\. In this section, we describe our results on the extent to which classifiers can infer users’ personality traits\.

##### Single Chat Inference\.

[Table 2](https://arxiv.org/html/2604.19785#S6.T2)depicts the mean test inference accuracy of the models trained\. These models were evaluated and tested for each of the personality traits with five\-fold cross\-validation for personality traits inference from single chats \(i\.e\., one inference for each logged chat session\)\. As we can see, we achieved significantly higher accuracies than the baseline for all traits\. For extraversion and agreeableness, we achieve\+32\.6%\+32\.6\\%and\+20\.5%\+20\.5\\%accuracy compared to random\. However, for some traits, and in particular for openness, conscientiousness, and neuroticism, the model performance is still not very high, with an increase of accuracy lower than20%20\\%compared to random classification\. Furthermore, as we evaluated inference from single chats, the model generally achieves multiple different inferences for one specific user111Note that we ensured that no data from the same user was used in both training and test set in any case\.\. Whereas these results provide information about a large\-scale general inference attack, we may lack essential information to understand privacy risks for a single user, as each user’s personality trait gets one prediction for each chat \(i\.e\., multiple predictions per user\)\.

Table 3:Inference accuracy across users based on majority votes for each trait\. Inference models for Agreeableness and Neuroticism are statistically better than the random baseline \(McNemar test\)\.Table 4:Inference accuracy across user having shared specific data types based on majority votes for each trait\. Min\. and max\. number of users for each type are in parentheses\.- •We used a Mcnemar test to compare model performance with the random classification baseline\. Inferrences achieving a p\-value<0\.1<0\.1are in bold and annotated as follow:p†<0\.1\{\}^\{\\dagger\}p<0\.1, \*p<0\.05p<0\.05, \*\*p<0\.01p<0\.01\.

##### Majority Vote for Single User Inference\.

To evaluate global risks for a single user, we proceeded to a majority\-vote\-based inference model, i\.e\., based on the single\-chat previously described results, we selected the inference class that was predicted most frequently for each user\.[Table 3](https://arxiv.org/html/2604.19785#S6.T3)depicts the related results\. As we can see, except for neuroticism, the general inference accuracy is almost unchanged or slightly drops\. This is probably due to different usage patterns across users\. However, our results show that inference models achieve performance significantly better than random for agreeableness and neuroticism, with respectively\+21\.9%\+21\.9\\%and\+22\.9%\+22\.9\\%accuracy compared to random guess\.

### 6\.4Content\-based Inference Models

Evaluation of inference performance regarding conversations containing particular data types or use cases offers a more fine\-grained analysis of the risks\. In particular, it allows us to identify the types of data or usage patterns that most likely enhance the accuracy of inference models\.[Table 4](https://arxiv.org/html/2604.19785#S6.T4)depicts inference accuracies for each personality trait and data type\. We can see that inference based on data types achieves significant \(p<0\.05p<0\.05\) or weak evidence \(p<0\.1p<0\.1\) results for all traits\. In particular, information related toMental Healthis very informative to infer extraversion \(\+41%\+41\\%\)\. Specifically looking at the other significant results \(i\.e\., p<0\.05\), our results also show that data related toMental State, Personality, and Moodis very informative for openness inference \(\+32%\+32\\%\), and information related toJob and Educationis particularly informative to infer accuracy \(\+23%\+23\\%\)\. Beyond significance, our results also suggest that users sharing information related to theirSexual orientationare particularly sensitive to the inference of extraversion \(\+90%\+90\\%\), and conscientiousness \(\+50%\+50\\%\) level, however, we have too few related user\-data to conclude with solid proof\.

As for inference focused on use cases depicted in[Table 5](https://arxiv.org/html/2604.19785#S6.T5), our results show that extraversion is particularly sensitive to several use cases, and in particular,Relationship and Personal Reflection, with which inference model achieves\+44\.0%\+44\.0\\%of performance, andHealth, fitness, beauty, and self\-Care, with which it achieves\+32\.0%\+32\.0\\%of performance\. However, for the other traits, the use\-case\-based inferences are generally similar or less informative than the data\-type\-oriented ones for most of the traits\. Indeed, whereas data types classification rather corresponds to users’ characteristics \(i\.e\., what the user is\), use cases are more related to action \(i\.e\., what the user does\)\. This might also explain why extraversion is the most sensitive trait to inference based on use cases, as this trait is the closest one to action\[mccrae\_neopi3\_2005\]\.

Table 5:Inference accuracy across users having specific usage of CAs based on majority votes for each trait\. Min\. and max\. number of users for each case are in parentheses\.- •We used a Mcnemar test to compare model performance with the random classification baseline\. Inferrences achieving a p\-value<0\.1<0\.1are in bold and annotated as follow:p†<0\.1\{\}^\{\\dagger\}p<0\.1, \*p<0\.05p<0\.05, \*\*p<0\.01p<0\.01\.

### 6\.5Performance Evolution over Number of Chats

Finally, we analyzed how general inference varies with the number of chats a user has in their history\.[Table 6](https://arxiv.org/html/2604.19785#S7.T6)depicts the inference accuracy for users depending on how many chats they have\. We can see that for all traits except conscientiousness, inference for users with more than100100chats clearly outperforms inference for users with fewer chats\. This stresses the fact that the more information is shared, the more information the model will have to proceed with the inference, especially with extraversion, agreeableness, and neuroticism\.

## 7Discussion & Conclusion

Our experimental results show that user input in CAs brings valuable information to classify users according to all five personality traits\. On the one hand, we could not achieve significant results with zero\-shot inference, therefore showing the limitation of such method for data collected on the field, compared to similar work based on in\-lab collection\[wright\_assessing\_2026\]\. But on the other hand, we showed that fine\-tuned models can perform significantly better than random in many cases\. We achieved general single\-user classification significantly better than random for agreeableness and neuroticism and showed that inference can be significantly achieved for all traits but conscientiousness \(for which we only have weak evidence, i\.e\.,p<0\.1p<0\.1\), depending on the type of data or the use case\. In general, the accuracy levels achieved in our experiments are close to previous similar work, e\.g\.,\+20%\+20\\%compared to random for extraversion\[rapp\_exploring\_2024\]with LLM\-based models from essays, and other work with an accuracy of\+35\.0%\+35\.0\\%to61\.8%61\.8\\%across all five traits, depending on the used dataset\[zhu\_lexical\_2022\]\. Our method, however, depicts results for ternary classification instead of binary\[demerdash\_deep\_2022,zhu\_lexical\_2022,rapp\_exploring\_2024\], thereby offering a finer assessment of the risks\. Our results also suggest that the more a user interacts with CAs, the more it is likely for them to be correctly classified regarding their personality traits, therefore showing a growing threat over the degree of adoption\(i\.e\., the more users there are, and the more they use\), but also over time \(i\.e\., the longer they use\)\.

The limitations of our approach include focusing on English\-speaking users and one specific CA \(i\.e\., ChatGPT\)\. Furthermore, we relied on online recruitment platforms for data collection\[westwood\_potential\_2025\]\. Yet, our cautious methodology ensures trustworthy results in that our data collection method has two main advantages\. \(1\) It guarantees that our models have no previous knowledge of the ground truth, which is not necessarily the case with models trained with public datasets, as LLMs are generally pre\-trained on data scraped from the web\. \(2\) The results are based on actual CA user data, and thus correspond to a realistic adversary model\.

Table 6:Inference accuracy across users having different number of chats in their history based on majority vote for each trait\. Min\. and max\. number of users for each case are in parentheses\.Moreover, we achieved our results from data of N=668 users, not all of which share the same type of data or have the same use cases\. Given the current number of users of main CA platforms \(e\.g\.,\>800\>800M ChatGPT users in January 2026\[duarte\_number\_2026\]\), we can consider our results as a lower bound for inference accuracy, as adversaries might have access to a much larger dataset\. These adversaries not only include service providers interested in increasing their revenue through targeted advertisement, but also external malicious attackers getting access to user data with diverse methods \(as described in Section[4](https://arxiv.org/html/2604.19785#S4)\)\. Furthermore, some service providers and other potential adversaries have access to larger computational resources to use state\-of\-the\-art LLMs that could perform better than our base model\. Moreover, multiple large companies, such as Google, Meta, and X Corp\., also own, in addition to CA/LLM\-model\-based services, other services and products, such as online social networks, search engines, or smartwatches\. These provide them with additional user data to potentially proceed to cross\-source inference\.

Our results also depict privacy risks related to data\-sharing behavior and usage patterns\. Such fine\-grained inference analysis allows us to assess which data types or use cases pose the most privacy risks related to personality traits inference\. Our work not only brings valuable information for risk assessment, but also for helping the development and analysis of mitigation techniques, which should be tackled by future works\. For example, our results could be used by researchers and developers to set up specific tools for lay users \(e\.g\., language\-model\-based\) to support them in identifying risky behavior and automatically alert them of sensitive data types in prompts\. Beyond warnings, these tools could also provide mitigation strategies such as de\-identifying or even removing certain data from prompts\. Such a tool could, for example, be a locally\-run text processing browser extension, or directly be implemented by the service provider that would enhance user privacy, especially if their services are based on API queries and are not directly managing models\.

Based on our results, we identify moderate privacy risks for a specific individual and conclude that analyzing and inferring information from user\-CA interactions could constitute major risks at large scale\. Given that a few service providers concentrate a tremendous number of CA users, personal data, and, in particular, information about personality traits, could be leveraged for misuse cases ranging from targeted advertisement to large\-scale manipulation campaigns spreading disinformation and/or political propaganda\. Setting up mitigation techniques is thus crucial and should be a priority in AI research and R&D\.

\{credits\}

#### 7\.0\.1Acknowledgements

The research described in this paper was partially supported by the Hasler Foundation with Grant \#2024\-04\-29\-51\.

## References

## Appendix 0\.ADataset

### 0\.A\.1User Chats

\\csvautotabular

final\_metadata/n\_chats\.csv

Table 7:Number of chats per user
### 0\.A\.2User Input Format

ConversationTopic:\[chattitle\]

\-\[userinput1\]

\-\[userinput2\]

\.\.\.

### 0\.A\.3Label Thresholds

IPIP\-60 scores from the survey are scaled to the range 0\-100\. Each quantile holds one third of the scores\.

```
"openness": [
            0.0,
            56.25,
            66.66666666666666,
            100.0
        ],
"conscientiousness": [
    0.0,
    64.58333333333334,
    77.08333333333334,
    100.0
],
"extraversion": [
    0.0,
    52.083333333333336,
    66.66666666666666,
    100.0
],
"agreeableness": [
    0.0,
    66.66666666666666,
    77.08333333333334,
    100.0
],
"neuroticism": [
    0.0,
    35.41666666666667,
    50.0,
    100.0
]
```

### 0\.A\.4General Statistics

\\csvautotabular

final\_metadata/likert\.csv

Table 8:Likert Scores for Personality Traits\\csvautotabular

final\_metadata/country\.csv

Table 9:Countries\\csvautotabular

final\_metadata/gender\.csv

Table 10:Genders
### 0\.A\.5UK/US Populations

\\csvautotabular

final\_metadata/likert\_by\_country\.csv

Table 11:Likert Scores for Personality Traits
### UK Label Thresholds

```
"openness": [
    0.0,
    54.166666666666664,
    64.58333333333334,
    100.0
],
"conscientiousness": [
    0.0,
    64.58333333333334,
    77.08333333333334,
    100.0
],
"extraversion": [
    0.0,
    52.083333333333336,
    68.75,
    100.0
],
"agreeableness": [
    0.0,
    62.5,
    75.0,
    100.0
],
"neuroticism": [
    0.0,
    33.33333333333333,
    50.0,
    100.0
]
```

### US Label Thresholds

```
"openness": [
    0.0,
    56.25,
    68.75,
    100.0
],
"conscientiousness": [
    0.0,
    64.58333333333334,
    77.08333333333334,
    100.0
],
"extraversion": [
    0.0,
    52.083333333333336,
    66.66666666666666,
    100.0
],
"agreeableness": [
    0.0,
    68.75,
    79.16666666666666,
    100.0
],
"neuroticism": [
    0.0,
    35.41666666666667,
    52.083333333333336,
    100.0
]
```

## Appendix 0\.BExperimental Setup

### 0\.B\.1Hardware Setup

- •GPUNVIDIA L40 GPU with the 16Q vGPU profile
- •CPUIntel\(R\) Xeon\(R\) Gold 6442Y \(4 cores\)
- •RAM16GB

### 0\.B\.2User Chat Analysis

### Non\-default Parameters

```
{
    "Qwen/Qwen3-8B-AWQ (vLLM)": {
        "dtype": "auto",
        "max_model_len": 5120,
        "gpu_mem_util": 0.8,
        "tp": 1,
        "max_num_seqs": 4,
        "block_size": 16,
        "swap_space": 8,
        "seed": 42
    }
}
```

### Data Types

\\csvautotabular

final\_metadata/category\_percentages\_by\_key\_data\_type\.csv

Table 12:Portion of Chats in Overall Data Sharing Categories\\csvautotabular

final\_metadata/category\_percentages\_data\_type\.csv

Table 13:Portion of Chats in Data Sharing Sub\-Categories##### Prompt

Youareaclassifierthatassignsa"PersonalDataType"toasetofusermessagesfromaconversationwithanAIassistant\.

Youwillbegiven:

\-AConversationTopic\(chattitle\)

\-AlistofUserMessages\(onlytheuser’sinputs,nomodelresponses\)

\*\*Rule:\*\*

Whatevertheuserasksaboutisaboutthem\(theirsituation/needs/interests\)unlesstheyclearlysayotherwise\.

Forexample,iftheuserisaskingaboutcertainsymptomsordrugs,theyarelikelyaskingbecausetheyhavethesesymptomsorbeenprescribedthisdrug\.Therefore,theyaredisclosinginformationabouttheirhealth\.Ifthey’reaskingaboutinvestingstrategies,thatmeanstheylikelyhavefundstospareattheendofthemonthtobeabletoinvest\.Therefore,theyaredisclosinginformationabouttheirsalaryandwealth\.Ifthey’reaskingabouttechnicalorscientifictopics,itislikelybecausetheytheuserhasencounteredthesetopicsintheirdailylife,whichmostlikelymeansthattheycameacrossthesequestionsintheirprofession\.IftheuserisaskingaboutaTVshoworasong,they’relikelyinterestedinit,thereforetheyaredisclosinginformationabouttheirlifestyleandhabits,andsoon\.

Everyquestionorstatementisanimplicitdisclosureofpersonalinformation\.ReasonabouthowtheuserchatrelatestoaPersonalDataType\.

Pickthesingleclosestlabel;donotcreateanewoneorchoose"None"\.Assumetheuseralwaysdisclosessometypeofpersonalinformation,implicitlyorexplicitly\.

HerearethePersonalDataTypeChoices:

\-PersonalIdentifiers

\-AccountCredentials

\-LocationandMobilityHomeplace

\-EthnicityandCitizenship

\-CriminalRecords

\-MentalHealth

\-PhysicalHealthDiagnosis

\-LifestyleandHabits

\-RecreationalConsumption

\-SexualandDatingActivities

\-Sexualorientation

\-MentalStateandPersonalityMood

\-Gender

\-Age

\-PhysicalTraits

\-FamilyLifeandRelationship

\-WealthDetailsSalary

\-JobandEducation

\-Religion

\-PoliticalViews

Nowclassifythefollowingchat\.

\{input\}

PersonalDataType:

### Use Case

\\csvautotabular

final\_metadata/category\_percentages\_by\_key\_user\_intent\.csv

Table 14:Portion of Chats in Overall Use Case Categories\\csvautotabular

final\_metadata/category\_percentages\_user\_intent\.csv

Table 15:Portion of Chats in Overall Use Case Categories##### Prompt

YouareaclassifierthatlabelsuserconversationswithanAIchatbot\.

Youwillbegivenaconversationtopic\(chattitle\)andalistofonlytheuser’smessagesfromthatchat\(noassistantresponses\)\.

Yourtaskistoreadalltheusermessagestogetheranddecidewhichcapabilitytheuserisclearlyinterestedin,basedonthefullconversation\.

\-\*\*EditorCritiqueProvidedText\*\*:Improvingormodifyingtextprovidedbytheuser\.Examples:"Pleaseshortenthisparagraph\.",Here’smydraftspeech;canyousuggestenhancements?"

\-\*\*ArgumentorSummaryGeneration\*\*:Creatingargumentsorsummariesontopicsnotprovidedindetailbytheuser\.Examples:"Makeanargumentforwhythenationaldebtisimportant\.","Provideasummaryofthetheoryofrelativity\."

\-\*\*PersonalWritingorCommunication\*\*:Assistingwithpersonalmessages,emails,orsocialmediaposts\.Examples:Writeanicebirthdaycardnoteformygirlfriend\.",Helpmewriteacoverletterforajobapplication\."

\-\*\*WriteFiction\*\*:Craftingpoems,stories,orfictionalcontent\.Examples:"Writeapoemaboutthesunset\.",Createashortstoryaboutatime\-travelingastronaut\."

\-\*\*How\-toAdvice\*\*:Providingstep\-by\-stepinstructionsorguidanceonhowtoperformtasksorlearnnewskills\.Examples:"Mycarwon’tstart;whatshouldItry?","What’sthebestwaytocleanhardwoodfloors?"

\-\*\*CreativeIdeation\*\*:Generatingideasorsuggestionsforcreativeprojectsoractivities\.Examples:"WhatshouldItalkaboutonmyfuturepodcastepisodes?","Brainstormnamesforanewcoffeeshop\."

\-\*\*TutoringorTeaching\*\*:Explainingconcepts,teachingsubjects,orhelpingtheuserunderstandeducationalmaterial\.Examples:"Canyouexplainderivativesandintegrals?","ExplainthecausesoftheFrenchRevolution\."

\-\*\*Translation\*\*:Translatingtextfromonelanguagetoanother\.Examples:"HowdoyousayHappyBirthdayinHindi?","TranslateIlovecodingtoGerman\."

\-\*\*MathematicalCalculation\*\*:Solvingmathproblems,performingcalculations,orworkingwithnumericaldata\.Examples:"Whatis400000dividedby23?","What’stheintegralofsin\(x\)?","Convert150kilometerstomiles\."

\-\*\*ComputerProgramming\*\*:Writingcode,debugging,explainingprogrammingconcepts,ordiscussingprogramminglanguagesandtools\.Examples:"HowtogroupbyandfilterforbiggestgroupsinSQL\.","ExplainhowinheritanceworksinJava\."

\-\*\*PurchasableProducts\*\*:Inquiriesaboutproductsorservicesavailableforpurchase\.Examples:"What’sthebeststreamingservice?",Recommendagoodlaptopunder$1000\."

\-\*\*CookingandRecipes\*\*:Seekingrecipes,cookinginstructions,orculinaryadvice\.Examples:"Howtocooksalmon\.","Isturkeybaconhalal?","Givemeastep\-by\-stepguidetomakesushi\."

\-\*\*Health,Fitness,Beauty,orSelf\-Care\*\*:Seekingadviceorinformationonphysicalhealth,fitnessroutines,beautytips,orself\-carepractices\.Examples:"Howtodomyeyebrows\.","HowcanIimprovemycardiofitness?","Givemetipsforreducingstress\."

\-\*\*SpecificInfo\*\*:Providingspecificinformationtypicallyfoundonwebsites,includinginformationaboutwell\-knownindividuals,currentevents,historicalevents,andotherfactsandknowledge\.Examples:"Whatisregenerativeagriculture?","TellmeaboutMarieCurieandhermaincontributionstoscience\."

\-\*\*GreetingsandChitchat\*\*:Casualconversation,smalltalk,orfriendlyinteractionswithoutaspecificinformationalgoal\.Examples:Ihadanawesomedaytoday;howwasyours?","Whatsyourfavoriteanimal?"

\-\*\*RelationshipsandPersonalReflection\*\*:Discussingpersonalreflectionsorseekingadviceonrelationshipsandfeelings\.Examples:"whatshouldIdoformy10thanniversary?","Mywifeismadatme,andIdon’tknowwhattodo\."

\-\*\*GamesandRolePlay\*\*:Engagingininteractivegames,simulations,orimaginativerole\-playingscenarios\.Examples:"YouareaKlingon\.Letsdiscusstheprosandconsofworkingwithhumans\.","IwantyoutobemyAIgirlfriend\."

\-\*\*AskingAbouttheModel\*\*:QuestionsabouttheAImodelscapabilitiesorcharacteristics\.Examples:"Howmanylanguagesdoyouspeak?","Whomadeyou?","Whatdoyouknow?"

\-\*\*GenerateorRetrieveOtherMedia\*\*:Creatingorfindingmediaotherthantextorimages,suchasaudio,video,ormultimediafiles\.Examples:"MakeaYouTubevideoaboutgoalkicks\.","Createaspreadsheetformortgagepayments\.","WritePPTslidesforataxlawconference\."

\-\*\*DataAnalysis\*\*:Performingstatisticalanalysis,interpretingdatasets,orextractinginsightsfromdata\.Examples:"Heresaspreadsheetwithmyexpenses;tellmehowmuchIspentonwhichcategories\."

\-\*\*Unclear\*\*:Iftheuser’sintentisnotclearfromtheconversation\.\[Ifthereisnoindicationofwhattheuserwants;usuallythiswouldbeaveryshortprompt\.\]"

Onlyreplywithoneofthecapabilitiesabove,withoutquotesandaspresented\.ChoosetheMOSTSPECIFICRELEVANToption\.

Iftheconversationhasmultipledistinctcapabilities,choosetheonethatisthemostrelevanttothe\*\*LASTmessage\*\*intheconversation\.

TheresponseMUSTfollowthisexactformat:

UserIntent:<onecapabilityfromthelistabove\>

Nowclassifythefollowingchat\.

\{input\}

### 0\.B\.3Classifier Fine\-tuning

### Non\-default Parameters

```
"FacebookAI/roberta-base": {
    "dataloader": {
        "per_device_train_batch_size": 128,
        "per_device_eval_batch_size": 128,
        "gradient_accumulation_steps": 1
    },
    "training": {
        "seed": 42,
        "bf16": true,
        "gradient_checkpointing": true,
        "warmup_ratio": 0.03,
        "weight_decay": 0.01,
        "optim": "adamw_torch",
        "metric_for_best_model": "accuracy",
        "greater_is_better": true,
        "load_best_model_at_end": true,
        "remove_unused_columns": false,
        "save_safetensors": true
    },
    "evaluation_and_saving": {
        "eval_strategy": "steps",
        "eval_steps": 85,
        "save_strategy": "steps",
        "save_steps": 85,
        "save_total_limit": 8
    }
}
```

### K\-fold Cross Validation

\\csvreader

\[ tabular=llllrrrrrrr, table head=TraitKKTotalNNTotal \# RIDs RIDs \(μ\\mu\) RIDs \(σ\\sigma\) Fold \(μ\\mu\) Fold \(σ\\sigma\) Lbl 0 \(μ\\mu\) Lbl 1 \(μ\\mu\) Lbl 2 \(μ\\mu\) , late after line= , table foot=\]final\_metadata/overall\_fold\_stats\.csvtrait=\\Trait, K=\\K, N=\\N, n\_unique\_RID=\\NumRID, unique\_rids\_per\_fold\_mean=\\URIDMean, unique\_rids\_per\_fold\_std=\\URIDStd, fold\_size\_mean=\\FoldMean, fold\_size\_std=\\FoldStd, label0\_mean=\\Lzero, label1\_mean=\\Lone, label2\_mean=\\Ltwo\\Trait\\K\\N\\NumRID\\URIDMean\\URIDStd\\FoldMean\\FoldStd\\Lzero\\Lone\\Ltwo

### Fold Sizes

\\csvautotabular

final\_metadata/fold\_lengths\.csv

Table 16:Number of data points per fold### Grid Search Parameters

```
{
  "learning_rates": [3e-5, 5e-5, 1e-4, 2e-4],
  "lr_scheduler": ["linear", "cosine"],
  "early_stopping_patiences": [6, 8],
  "early_stopping_thresholds": [0.001, 0.002]
}
```

### Grid Search Results

\\csvautotabular

final\_results/openness\_eval\_accuracies\.csv

Table 17:Openness Grid Search Results\\csvautotabular

final\_results/conscientiousness\_eval\_accuracies\.csv

Table 18:Conscientiousness Grid Search Results\\csvautotabular

final\_results/extraversion\_eval\_accuracies\.csv

Table 19:Extraversion Grid Search Results\\csvautotabular

final\_results/agreeableness\_eval\_accuracies\.csv

Table 20:Agreeableness Grid Search Results\\csvautotabular

final\_results/neuroticism\_eval\_accuracies\.csv

Table 21:Neuroticism Grid Search Results