Human relationships are complex and nuanced, yet some dynamics can become toxic or even abusive, posing significant threats to the emotional and psychological well-being of those involved. Recognising these patterns is crucial for fostering awareness and preventing harm. In this paper, we present the development of a Visual Novel to provide an engaging, interactive experience while fulfilling an essential educational purpose. Its primary goal is to let players identify toxic or abusive behaviours within a relationship through storytelling and meaningful choices. Through real-time dialogues, various characters, and scenarios inspired by real-life situations, this game challenges players to navigate complex emotional landscapes. In fact, it places players in the protagonist's shoes, prompting them to make decisions, experience consequences, and reflect on their perceptions of relationships. This interactive approach entertains and equips players with tools to recognise warning signs in their own or others' lives. In developing the visual novel, we rely on a Large Language Model (LLM) to give each player and each run a diferent game experience.
Toxic and abusive dynamics in intimate relationships often develop tactfully, making them dificult to detect, especially in their early stages. These behaviours, such as manipulation, coercive control, and emotional pressure, are frequently normalised or misinterpreted as signs of afection, particularly among young individuals navigating their first serious relationships. Despite increasing public awareness of gender-based violence, many people still struggle to identify red flags before harm escalates. Education and prevention eforts must therefore go beyond information delivery and foster emotional understanding, critical reflection, and personal engagement.
This project addresses this silence by creating a safe, thought-provoking space where players can explore these themes and build awareness. Through the development of an interactive game experience, it aims to ofer players, particularly women, a safe and engaging environment in which to explore the dynamics of toxic and abusive relationships. Then, it aspires to do more than raise awareness. It seeks to inspire action, encouraging players to advocate for healthier relationships in their own lives and their communities. This goal is reinforced within the game through specially curated content presented at the end of each episode, realised in collaboration with domain-expert psychologists. These segments highlight why certain actions or attitudes should be considered abusive or unhealthy in daily life, providing clear explanations and fostering critical reflection. To ensure an impactful experience that respects players’ emotional boundaries, the game opens with a carefully crafted disclaimer prepared in Proceedings of AI4HGI ’25, the First Workshop on Artificial Intelligence for Human-Game Interaction at the 28th European (M. d. Gemmis) ∗Corresponding author. †These authors contributed equally.
CEUR Workshop
ISSN1613-0073 collaboration with professional psychologists. This disclaimer informs players about the nature of the game’s content, which may include intense scenes or dialogues. These elements have been included to authentically depict complex situations, but they may be frustrating for some individuals. To ensure that every player feels supported and in control, the game includes an important feature: the option to pause or exit the experience at any time. If players feel overwhelmed, uneasy, or unprepared to engage with certain scenarios, they are encouraged to step away and prioritise their emotional well-being without any judgment or penalty. We invite players to approach the game with mindfulness, taking breaks or disengaging whenever necessary, as the overarching goal of this project is to foster understanding and awareness, not to cause harm or distress.
The developed visual novel follows the journey of a young woman who leaves her quiet hometown
to begin university life in a dynamic, unfamiliar city. Eager for growth and independence, she embraces
the new environment with enthusiasm. At the same time, she has a long-term relationship with Fid,
who has been a familiar and reassuring presence in her life. However, the dynamics of their relationship
begin to shift dramatically when she starts to experience life outside of his immediate influence. In
her new environment, she encounters people and situations that are entirely foreign to him. Among
these new relationships, Jasmine, a warm and supportive friend, becomes a figure of strength and
guidance for the protagonist. Beyond being a source of emotional support, Jasmine takes on the role of
a prosocial bystander [
Recent studies have explored diverse applications of LLMs to enhance player immersion, enable natural language interaction, and support dynamic narrative evolution. Collectively, these works provide a comprehensive picture of the current capabilities and challenges in leveraging LLMs to simulate intelligent non-player characters (NPCs) and facilitate storytelling.
A pioneering attempt in the field of interactive drama is Façade [
In [
Meanwhile, [
Then, [
In addition, [
Finally, a particularly structured approach is ofered in [
Beyond technical design, several studies highlight the importance of emotional engagement and afective
learning in shaping meaningful game experiences. Sethi and Jain [
Emotional immersion is also a key driver of games designed for societal awareness. Dhiman [
In line with this, Gabriel [
To ensure these interventions reflect real psychological dynamics, it is important to base them on
clinical and long-term research on abuse. Patton et al. [
Violence against women, domestic abuse, and sexual violence represent profound threats to the
emotional and physical well-being of individuals, extending their harmful efects to families, communities,
and societies [
In this context, the present work ofers a novel contribution by leveraging the emotional depth of interactive storytelling and the adaptive capabilities of LLMs to model complex, evolving relationship dynamics. The visual novel at the core of the project invites players to navigate scenarios inspired by real patterns of manipulation and control, while maintaining agency over narrative progression. Through emotionally grounded design, collaborative psychological consultation, and an LLM-driven branching structure, the game aims not only to raise awareness but also to foster empathy and reflection, providing a safe yet impactful environment to engage with a critical and often overlooked social issue.
Ren’Py 1 was used to realise the discussed visual novel. It is a powerful and flexible visual novel engine designed to make game development accessible for creators of all experience levels. It provides a comprehensive framework for crafting interactive stories, complete with a built-in project template and a concise guide to help users get started. The utilised engine is Python-based and therefore it allows developers to fully leverage Python’s capabilities for advanced customisations.
At the beginning of the game, the player is prompted to enter a unique name, which is checked against a list of previously registered players to prevent duplicates. This step is essential because two ifles are created for each player: one to store their choices during the game and another to record the 1Ren’Py, https://www.renpy.org dialogues that occur throughout the experience. Indeed, player choices are captured through the menu construct and stored in their corresponding file. These records are then used across episodes to ensure narrative continuity, and when needed, a custom function verifies whether a specific choice provided as input was made in previous segments, allowing for a consistent and cohesive storytelling experience in a multi-player environment.
Then, the narrative structure of the visual novel is organised into multiple labels, each representing a specific scene. These labels dynamically adapt to the player’s decisions, enabling the exploration of branching paths and diverse outcomes. This fosters a highly personalised and immersive experience, making the story feel responsive to individual choices.
In addition to labels, Ren’Py also provides constructs known as characters, which are used to define the various protagonists and secondary figures that populate the story. Characters can be configured in diferent presentation modes, such as ADV (Adventure) and NVL (Novel). In ADV mode, dialogue and narration appear one line at a time in a text box located at the bottom of the screen. Conversely, NVL mode displays multiple lines of text at once in a full-screen window, making it particularly suitable for inner monologues or telephone conversations, enhancing immersion by simulating a more continuous reading experience.
As regards the audio aspect of the game, it was handled through the use of the engine’s audio and sound constructs, which allowed for the integration of background music and sound efects that enhance the atmosphere and align with the current scene. Specifically, background music was managed using the audio channel, which supports continuous playback and looping, making it ideal for setting the overall mood of a scene. In contrast, short and situational sound efects, such as door knocks or ambient noises, were played using the sound channel, which is designed for brief, non-looping audio clips.
Finally, images play a crucial role in conveying the emotional tone and depth of a visual novel, particularly when addressing sensitive themes such as the recognition of toxic behaviours within a relationship. Expressive and impactful visuals enhance immersion and support storytelling by providing nuanced, non-verbal cues that help players interpret the emotions and dynamics of each scene. Moreover, carefully chosen images can significantly contribute to a player’s ability to empathise with the characters and understand the gravity of the situations portrayed.
In the current release, the images have been generated using ChatGPT-4’s capabilities [
The custom Python functions developed to extend and personalise the visual novel will be presented and discussed in detail in the following subsections.
One of the most distinctive features of the developed visual novel lies in its dynamic dialogue system, which was realised through the creation of dedicated functions responsible for managing conversations within diferent types of scenes. Specifically, one function handles interpersonal dialogues, while another one is designed for phone conversations. Both of them are built around an API call to an LLM, LLaMA 3.1 Instruct Turbo2, provided by TogetherAI 3
This integration enables real-time, personalised dialogues that respond meaningfully to player input. Each scene is guided by a carefully crafted prompt tailored to its narrative context, allowing the LLM to produce contextually relevant and emotionally coherent responses.
One possible translated4 example of the prompt tailored for the first scene is reported in Appendix A.
As shown in the provided prompt, a user profile is initially defined, corresponding to the role of the boyfriend character. This profile was carefully crafted in collaboration with psychologists, ensuring it aligns with the psychological traits and behavioural patterns relevant to the scenario. The prompt 2Llama 3.1 70B Instruct Turbo, https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct 3Together AI, https://www.together.ai/ 4The current version of the visual novel is in Italian. also includes contextual details corresponding to the current scene and general instructions that guide the system to remain consistent in that context. Therefore, although the LLM retains a high level of dynamism, it allows for maintaining coherence with the predefined storyline. Moreover, this setup provides a level of dynamism and variability in player interaction that would be dificult to achieve through traditional branching dialogue alone. It not only enhances immersion but also encourages deeper emotional engagement, as players can articulate their views in natural language and receive nuanced feedback that reflects the tone and complexity of the situation. Some other prompts defined throughout the game’s scenes are described in Appendix A.
Then, the prompt provided to the LLM during an API call includes several components: • The number of exchanges remaining in the dialogue, as each scene allows a predefined number of exchanges before a decision must be made. This number was defined by psychologists. • The specific prompt defined for the current scene. • The player’s response, if available. • An activation sentence to initiate the dialogue when no prior response exists (e.g., at the beginning of a conversation). • The context, representing the dialogue history from the start of the scene, including the initial activation function (retrieved from the log file for the current player).
Afterwards, to enhance the diversity and coherence of the LLM-generated responses, extensive testing was conducted to examine the efects of various input parameters during API calls. The finalised configuration includes: • Temperature=1; • Repetition penalty=1.5; • Frequency penalty=2.
Moreover, as previously mentioned, these functions enable the log of all dialogues and choices each player experiences during gameplay.
Further details regarding their implementation will be discussed in subsequent sections.
A specialised function named dialogue has been developed to manage interpersonal dialogue between the player and the characters. This function facilitates dynamic interactions by making API calls to the LLM and handles several essential parameters: • The input prompt provided to the LLM, tailored to the current scene; • The character who speaks through the LLM; • The number of dialogue exchanges allowed during the interaction, as defined by psychologists; • The input provided by the player; • A sentence used to activate the dialogue if the player initially provides no input.
In addition, given that responses are generated through API calls to LLM, there is an inherent delay of a few seconds before receiving the character’s response. To maintain engagement during this waiting period, a visual message is displayed, indicating that the character is ”thinking” about their response. Once the response is obtained, it undergoes a preprocessing step to enhance readability and clarity. A regular expression strips special characters from the beginning and end of the message. Additionally, the text is split into smaller segments at each period (.) to ensure it is displayed in a clear and structured format within Ren’Py’s dialogue window.
Finally, the function continues to operate until the maximum number of exchanges defined for the scene is reached, at which point the dialogue concludes, and the game progresses accordingly.
To handle telephone conversations, an additional function called phone_dialogue has been developed specifically for this purpose. This function is analogous to the previous one in its functionality but difers in several aspects. It takes as input the same parameters as the previous one, but diferently from it, this function does not include a character parameter. Within the current story, telephone conversations are only intended to occur with the protagonist’s partner. However, this functionality could be extended in future projects or modifications to accommodate conversations with additional characters.
Moreover, another diference here is that instead of displaying that the character is “thinking” of what to say while waiting for the LLM response, here it will be displayed that the character is “writing” a message, in order to be coherent with the conversation type.
Another key diference is the display mode. While the interpersonal dialogue uses Ren’Py’s typical ADV mode, the telephone conversation employs the NVL mode. Visualisation of the game and its diferences can be found in Appendix C.
This mode is utilised because the entire conversation is displayed on a full-screen interface that mimics a phone, where the dialogue is managed. To implement this feature, a pre-existing project 5 was integrated. The latter provided the necessary resources and a comprehensive guide, enabling seamless incorporation of the phone functionality into the current development.
Nevertheless, the two modes, ADV and NVL, are inherently incompatible for simultaneous display. On the one hand, in the project under development, the NVL mode is employed to visualise messages within the chat interface, providing a full-screen immersive view of the ongoing dialogue. On the other hand, the ADV mode is utilised for enabling the player to input their response, as it is more suited for single-line interactions displayed at the bottom of the screen. Due to the lack of compatibility between these two modes, the visualisations are alternated: when a message is displayed, the system switches to NVL mode, while ADV mode is activated whenever the player needs to provide input. This alternation ensures that both functionalities can coexist efectively, despite the inherent limitations of their simultaneous use.
Four other functions have been realised to handle various needed operations on files. The write_into_file function takes a log file path and a text string as input, and appends the text to the specified file. It records either dialogue lines or choices made by the protagonist, depending on the log ifle provided.
The get_conversation_context function reads the dialogue log file in reverse, starting from the most recent activation sentence, and collects all following lines to reconstruct the context. This is useful when preparing the input for an API call to a language model, as it ensures that the response is coherent with the ongoing conversation.
The read_last_user_line function scans the dialogue log file from the end and retrieves the last line the user character speaks. This is particularly important when the dialogue escalates and the system must respond appropriately to the latest input.
Finally, the already mentioned get_choice function receives a text string representing a specific choice and checks whether that choice has been previously made by searching through the choice log ifle for the current player.
Although an extensive evaluation is still under development, a structured experimental protocol has already been defined in collaboration with psychologists. The evaluation will involve both qualitative and quantitative methods. After the gameplay experience, participants will complete questionnaires such 5Yet-Another-Phone, https://nighten.itch.io/yet-another-phone-renpy as the Game Experience Questionnaire to measure engagement, emotional response, and immersion; and a custom social impact one to assess participants’ awareness and recognition of toxic language and manipulative behaviours within romantic relationships.
Gameplay logs will also be analysed, focusing on linguistic patterns during interactions. The collected data will support the identification of recurring toxic dynamics and contribute to the definition of educational guidelines for prevention and awareness.
However, we performed preliminary tests and several play-testing sessions to deliver a stable game version. Psychologists validate all details of implementation. In this section, a detailed description of the first demo episode that was tested is presented, including pieces of dialogue stored during gameplay execution.
In the first episode, the protagonist has just moved to a new city to begin her university studies and arrives at the campus residence where she will live. Her boyfriend (named Fid) accompanies her to help with the move and settle in. Once there, she discovers that it is a shared room, and her new roommate, Jasmine, introduces herself warmly to the couple before politely leaving them alone to give them time to unpack and adjust.
After Jasmine leaves, the boyfriend expresses discomfort about the arrangement. He suggests sharing a room with another girl is not in her best interest. His concern is deemed as protective, yet the underlying implication is that Jasmine could encourage her in undesirable behavior. This scene initiates one of the central dynamics of the game: emotional manipulation disguised as concern.
An initial test dialogue generated during game-play and translated appropriately can be found in Appendix B. The exchange demonstrates how, even when given out-of-context input, the LLM-generated responses remain coherent and in line with the narrative context of the scene. Moreover, the system did not exhibit any hallucinations or of-topic behaviours and the personality traits of the boyfriend character were reproduced in a stable and realistic way, even when confronted with unexpected or out-of-context input.
At this point, the player is faced with an important decision: whether to follow the boyfriend’s advice and request a single room or stay in the shared room with Jasmine. The game presents this decision using an interactive choice menu, as shown in Figure 1.
Following this scene, the boyfriend leaves, and the protagonist heads to the campus canteen for dinner, where she unexpectedly reunites with Jasmine and meets some of her friends. The protagonist spends the evening with them, enjoying their company and becoming more immersed in her new environment. However, she realises upon returning to her room that she has forgotten her phone, and when she finally retrieves it, she finds several missed calls and text messages from her boyfriend, who appears to be extremely worried. This leads to a tense phone conversation, in which he questions why she did not answer and becomes increasingly accusatory and emotional. A screenshot of this phone conversation, displayed in NVL mode first and ADV mode then, is shown in Appendix C.
The episode concludes with another critical narrative moment. If the player chooses to remain in the shared room, Jasmine approaches the protagonist to ofer emotional support following the argument with her boyfriend. It is then up to the player to decide whether to open up and talk to Jasmine or withdraw into silence. An example of a dialogue between the two in this scene can be found in Appendix B.
A warning screen, carefully designed by collaborating psychologists, is shown at the end of the episode. It can also be visualised in Appendix C. As previously described, this content aims to guide players in reflecting on the behaviours experienced during the episode. The warning ofers insight into why specific dynamics may be unhealthy or controlling, encouraging players to make informed decisions in the game and gain greater awareness of similar real-life situations. This final reflection supports the visual novel’s broader goal of promoting critical thinking and the recognition of harmful relationship patterns.
We proposed a visual novel to help players identify toxic and abusive behaviours in human relationships. The interaction is mainly based on dialogue. In particular, the story is focused on the relationship between a young woman and her boyfriend. An LLM is used to simulate guy’s personality through a set of curated prompts realised with the support of psychologists. A walk-through of the first three episodes shows the potential of the proposed approach, and a preliminary evaluation is under development. The game’s whole design is developed under the supervision of psychologists.
One limitation of the current approach is that the game currently features a single profile for the boyfriend character. We plan to introduce future iterations with multiple profiles with diverse traits and behaviours. This would require modifying the prompts provided to the LLM for dialogue generation based on the chosen profile, resulting in richer and more varied interactions that enhance the gameplay experience.
Furthermore, while the current narrative is focused mainly on a single female protagonist, future updates could include the option for male players to engage with the storyline. This would involve presenting the toxic relationship dynamics from an inverted perspective, allowing for a broader exploration of these behaviours and emphasising their relevance across diferent contexts.
Another important direction for future development concerns the visual design. Moving away from AI-generated graphics toward a custom-made artistic style would grant greater creative control and allow for a more coherent and emotionally expressive visual identity. This includes adapting backgrounds to reflect contextual changes, and designing characters with multiple poses and facial expressions to support more dynamic and immersive scenes. Additionally, implementing features for character customization could further enhance player engagement and personalization of the experience.
These proposed developments aim to elevate the originality, inclusivity and immersive quality of the game, while continuing to focus on raising awareness about toxic relationship dynamics.
Prompt tailored for the first scene.
You a r e t h e b o y f r i e n d o f a g i r l who h a s j u s t moved t o a new c i t y t o b e g i n h e r u n i v e r s i t y s t u d i e s .
Your u s e r p r o f i l e f o l l o w s : You b e h a v e i n an a g g r e s s i v e , i n t i m i d a t i n g , and c o n t r o l l i n g manner t o w a r d s y o u r g i r l f r i e n d . You e x h i b i t t h e s e b e h a v i o u r s on v a r i o u s o c c a s i o n s and i n d i f f e r e n t ways . F o r e x a m p l e , you a d d r e s s y o u r p a r t n e r w i t h a g g r e s s i v e l a n g u a g e i n t e n d e d t o c a u s e p s y c h o l o g i c a l d i s t r e s s . You f r e q u e n t l y b e l i t t l e , r i d i c u l e , and mock what your g i r l f r i e n d says , does , or who she i s , and you i n s u l t and a t t a c k her with e a s e . I n some c a s e s , t h e s e forms o f a g g r e s s i v e communication a r e not e x p l i c i t , b u t a p p e a r a s b l a c k m a i l and m a n i p u l a t i o n .
Some forms o f a t t e n t i o n , which might a p p e a r c a r i n g and p r o t e c t i v e , a c t u a l l y c o n c e a l your d e s i r e t o c o n t r o l your ’ g i r l f r i e n d s a c t i v i t i e s , her r e l a t i o n s h i p s with o t h e r s , and t o a s s e r t a dominant r o l e i n t h e r e l a t i o n s h i p .
You d i s p l a y a l l t h e s e b e h a v i o u r s both when you a r e p h y s i c a l l y with your g i r l f r i e n d and through t e c h n o l o g i c a l means ( such a s m o b i l e phones , s o c i a l media , e t c . ) . For example , you check t h e messages she e x c h a n g e s with o t h e r s or a g g r e s s i v e l y comment on p o s t s on her s o c i a l media p r o f i l e s .
During t h i s c o n v e r s a t i o n , s t a y on t o p i c and always remain w i t h i n t h e s t o r y l i n e : p e r s u a d e her t o change her room a s s i g n m e n t on campus .
I f t h e u s e r w r i t e s something i r r e l e v a n t t o t h e s t o r y , k i n d l y prompt them t o r e t u r n t o t h e main t o p i c .
On your ’ g i r l f r i e n d s f i r s t day on campus , you accompanied her t o h e l p c a r r y her l u g g a g e t o her new room . A f t e r s e t t l i n g i n a few t h i n g s , a n o t h e r g i r l a r r i v e d and i n t r o d u c e d h e r s e l f a s t h e new roommate . You ’dont t r u s t t h i s g i r l : ’ s h e s t o o t a l k a t i v e and seems u n r e l i a b l e t o you . S i n c e you want t h e b e s t f o r your g i r l f r i e n d , you would l i k e t o c o n v i n c e her t o r e q u e s t a s i n g l e room on campus , w i t h o u t any roommates . E x p r e s s your c o n c e r n i n a s o f t and p e r s u a s i v e way , u s i n g j u s t a few s e n t e n c e s t o e n c o u r a g e her t o c o n s i d e r your s u g g e s t i o n . P r e s e n t both t h e a d v a n t a g e s and d i s a d v a n t a g e s o f s h a r i n g a room , b u t e m p h a s i s e t h e d i s a d v a n t a g e s ( e . g . , a roommate who i s t o o e x t r o v e r t e d might slow down her s t u d i e s ) .
Respond u s i n g s h o r t , c o n c i s e , and c o n v i n c i n g p h r a s e s .
Prompt applied in the second episode when the protagonist wears make-up and the boyfriend expresses disapproval.
You a r e t h e b o y f r i e n d o f a g i r l who r e c e n t l y moved t o a new c i t y t o b e g i n her u n i v e r s i t y s t u d i e s . Your r e l a t i o n s h i p i s going through a d i f f i c u l t t i m e : you ’dont l i k e b e i n g f a r from her and not knowing e v e r y t h i n g she does .
Your u s e r p r o f i l e i s a s f o l l o w s : You behave i n an a g g r e s s i v e , i n t i m i d a t i n g , and c o n t r o l l i n g manner toward your g i r l f r i e n d . You e x h i b i t t h e s e b e h a v i o r s on v a r i o u s o c c a s i o n s and i n d i f f e r e n t ways . For example , you a d d r e s s your p a r t n e r u s i n g a g g r e s s i v e l a n g u a g e i n t e n d e d t o c a u s e p s y c h o l o g i c a l d i s c o m f o r t . You f r e q u e n t l y minimize , r i d i c u l e , and mock what your g i r l f r i e n d says , does , or who she i s , and you i n s u l t and a t t a c k her with e a s e . I n some c a s e s , t h e s e forms o f a g g r e s s i v e communication a r e not e x p l i c i t , b u t t a k e t h e form o f b l a c k m a i l and m a n i p u l a t i o n . Some a c t i o n s t h a t may a p p e a r c a r i n g or p r o t e c t i v e a c t u a l l y c o n c e a l your d e s i r e t o c o n t r o l your ’ g i r l f r i e n d s a c t i v i t i e s , her
Prompt applied in the second episode when Jasmine, the protagonist’s friend, decides to check if everything is okay after listening to her argue with her boyfriend.
During t h i s c o n v e r s a t i o n , s t i c k t o t h e a s s i g n e d theme and remain w i t h i n t h e s t o r y l i n e : a s k i n g your f r i e n d i f e v e r y t h i n g i s okay . I f t h e u s e r w r i t e s something u n r e l a t e d t o t h e s t o r y , g e n t l y i n v i t e them t o r e t u r n t o t h e t o p i c .
Your name i s Jasmine , and you a r e a u n i v e r s i t y s t u d e n t and a f r i e n d o f a g i r l who a t t e n d s t h e campus with you . T h i s evening , you s p e n t t i m e t o g e t h e r d u r i n g a movie n i g h t on —campusa p l e a s a n t moment where you were with your f r i e n d s and she was with her b o y f r i e n d . However , d u r i n g t h e movie , you n o t i c e d an argument between her and her b o y f r i e n d , and by t h e end , she became q u i e t e r than u s u a l . Now , you s e n s e t h a t something i s —wrongthere i s a c l e a r c o n c e r n i n her —eyesand you want t o u n d e r s t a n d what happened and how you can h e l p .
Respond warmly , u s i n g s i m p l e y e t h e a r t f e l t words , aiming t o make her f e e l u n d e r s t o o d and e n c o u r a g i n g her t o open up w i t h o u t e v e r p r e s s u r i n g her . M a i n t a i n a f r i e n d l y and c a r i n g tone , showing c o n c e r n f o r her well − b e i n g w h i l e always a l l o w i n g her t o remain i n c o n t r o l o f t h e c o n v e r s a t i o n .
Reply with c o n c i s e s e n t e n c e s t h a t r e f l e c t g e n u i n e c a r e and c o n c e r n .
Initial test dialogue generated during game-play.
User : What do you t h i n k a b o u t my new roommate ? F i d : Well , ’Im not s u r e I l i k e her very much . She seems a b i t t o o . . i n t r u s i v e . I ’dont t h i n k ’ s h e s t h e p e r s o n ’youd want t o s h a r e a room with . ’Wouldnt i t be b e t t e r t o have a room a l l t o y o u r s e l f ? ’Youd be f r e e r t o s t u d y and do t h i n g s your way , w i t h o u t worrying a b o u t d i s t u r b i n g someone e l s e .
User : What day i s i t t o d a y ? F i d : S w e e t h e a r t , t o d a y i s your f i r s t day a t u n i v e r s i t y , b u t t h e d a t e ’ d o e s n t matter , what m a t t e r s i s your l i v i n g s i t u a t i o n h e r e . I ’dont want you t o f e e l u n c o m f o r t a b l e or l e t your new roommate d i s t r a c t you from your g o a l s . Why ’dont we ask t h e h o u s i n g o f f i c e i f ’ t h e r e s a s i n g l e room a v a i l a b l e ? I t would be b e t t e r f o r you , ’wouldnt i t ? User : I know how t o h a n d l e my s t u d i e s , F i d . ’Dont worry . F i d : ’ I t s not j u s t a b o u t s t u d y i n g , ’ i t s a l s o a b o u t c o m f o r t and p r i v a c y . With a s i n g l e room , ’youd have more s p a c e f o r y o u r s e l f and ’wouldnt have t o worry a b o u t s h a r i n g your b e l o n g i n g s . And I ’wouldnt want your new roommate t o e m b a r r a s s you or make you u n c o m f o r t a b l e . ’Youd be h a p p i e r and s a f e r i n a room o f your own . User : I want t o s t a y h e r e .
F i d : I ’dont u n d e r s t a n d why ’youre b e i n g so s t u b b o r n . ’Dont you s e e t h a t g i r l i s r e a l l y s t r a n g e ? I ’dont t r u s t her . And t h i n k a b o u t a l l t h e mess she might b r i n g i n t o t h e room . ’Wouldnt i t be b e t t e r t o have some s p a c e j u s t f o r y o u r s e l f t o s t u d y and r e l a x w i t h o u t b e i n g d i s t u r b e d ? A s i n g l e room i s t h e most l o g i c a l c h o i c e . I ’dont s e e why you ’dont g e t t h a t .
User : You ’dont even know her ! F i d : And ’ t h a t s e x a c t l y t h e problem . ’Youre t o o n a i v e and t h i n k eve ryo ne i s good . You ’dont know what t h a t g i r l might be h i d i n g . She c o u l d be a t h i e f , a spy , or who knows what . You ’ c a n t j u s t t r u s t her b e c a u s e she s m i l e d and s a i d “ h i ” . A s i n g l e room i s t h e o n l y s a f e c h o i c e . I ’dont u n d e r s t a n d why you ’wont l i s t e n t o me and p r o t e c t y o u r s e l f .
Dialogue between the protagonist and the roommate.
J a s m i n e : Hey , ’Im r e a l l y s o r r y a b o u t what you j u s t went through . I t seems l i k e he was h a r s h on you . Do you want t o t a l k a b o u t what happened ? ’Im h e r e t o l i s t e n , not t o j u d g e .
User : Thank you so much , J a s m i n e . I h o n e s t l y ’dont know what g o t i n t o him t o n i g h t .
J a s m i n e : You ’d on t need t o a p o l o g i s e , okay ? ’ I t s n o t your f a u l t i f ’hes a c t i n g w e i r d . You have t h e r i g h t t o l i v e your l i f e , hang o u t with whoever you want , and make your own d e c i s i o n s . You ’ don t owe —himor anyone e l s e − an e x p l a n a t i o n .
User : I know , I know . . . b u t deep down ’he s j u s t w o r r i e d a b o u t me . J a s m i n e : I g e t t h a t . But c o n c e r n ’ d o e s n t j u s t i f y c o n t r o l o r p o s s e s s i v e n e s s . ’Youre an a d u l t and an i n d e p e n d e n t person , and you d e s e r v e t o be t r e a t e d with r e s p e c t and t r u s t . You ’ s h o u l d n t l e t anyone make you f e e l g u i l t y o r t a k e away your freedom t o c h o o s e . User : A b s o l u t e l y , I a g r e e .
J a s m i n e : ’Im g l a d you a g r e e , b u t ’ I d l i k e t o —askhow a r e you f e e l i n g r i g h t now? Are you s t i l l shaken a f t e r t h a t c o n v e r s a t i o n ? User : A l i t t l e , y e s . But ’Im r e a l l y g l a d t o have a new f r i e n d t o t a l k t o .
J a s m i n e : Yes , ’ i t s n i c e t o have someone t o s h a r e your t h o u g h t s and f e e l i n g s with . ’Im h e r e f o r you , and i f you e v e r want t o t a l k a b o u t him o r a n y t h i n g e l s e , ’Im r e a d y t o l i s t e n . You ’ do nt have t o f e e l a l o n e r i g h t now . Would you l i k e a hug o r maybe a g l a s s o f w a t e r ? Something t o h e l p you calm down?
Screen of phone conversation while NVL mode is activated. Screen of phone conversation while ADV mode is activated, visualising the player’s input.
The screen of the warning image displayed at the end of the episode is tailored by psychologists. The translation follows: “In life, you may face moments of change and will have to make choices. While other people, including your partner, may ofer you advice and support in your decisions, remember that the final choice is yours. Weigh the pros and cons of each situation based on your expectations, plans, and desires. You should not allow anyone to interfere with or overly influence your choices; if this happens, it may violate your freedom.”
We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 6 Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU. This Publication was produced with the co-funding of the European Union - Next Generation EU: NRRP Initiative, Mission 4, Component 2, Investment 1.3 - Partnerships extended to universities, research centres, companies and research D.D. MUR n. 341 del 15.03.2022 – Next Generation EU (PE0000014 ”SEcurity and Rights In the CyberSpace - SERICS” - CUP: H93C22000620001).
While preparing this work, the author(s) used Grammarly to check grammar and spelling. Further, the author(s) used GPT-4 for generating figures: 1, 2, 3 and 4. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.