Can AI Have a Personality? Prompt Engineering for AI Personality Simulation: A Chatbot Case Study in Gender-Affirming Voice Therapy Training

Abstract—This thesis investigates whether large language models (LLMs) can be guided to simulate a consistent personality through prompt engineering. The study explores this concept within the context of a chatbot designed for Speech-Language Pathology (SLP) student training, specifically focused on gender-affirming voice therapy. The chatbot, named Monae Jackson, was created to represent a 32-year-old transgender woman and engage in conversations simulating client-therapist interactions. Findings suggest that with prompt engineering, the chatbot maintained a recognizable and consistent persona and had a distinct personality based on the Big Five Personality test. These results support the idea that prompt engineering can be used to simulate stable personality characteristics in AI chatbots.

Can AI Have a Personality? Prompt Engineering for AI Personality Simulation: A Chatbot Case Study in Gender- Affirming Voice Therapy Training

This thesis investigates whether large language models (LLMs) can be guided to simulate a consistent personality through prompt engineering. The study explores this concept within the context of a chatbot designed for Speech-Language Pathology (SLP) student training, specifically focused on gender-affirming voice therapy. The chatbot, named Monae Jackson, was created to represent a 32-year-old transgender woman and engage in conversations simulating client-therapist interactions.

The research is grounded in a single hypothesis: that carefully constructed prompts can influence an AI chatbot to behave as if it has a stable, predefined personality. Monae’s behavior was shaped through detailed persona development and iterative prompt refinement. The prompts were designed to encourage consistency in tone, emotional response, and conversational boundaries and to guide the chatbot’s conversational behavior throughout each session.

While the scope of the study is limited, the findings provide a starting point for further research into the use of AI personas in educational and therapeutic simulations.

This thesis investigated whether structured prompt engineering could be used to simulate a coherent and consistent AI personality across multiple interactions and assessments.

This detailed persona is stored externally in a text file, which the chatbot loads and incorporates into its initial system prompt at the beginning of each session. This approach ensures that the persona’s characteristics remain consistent throughout interactions while allowing easy updates and modifications to the background information as needed.

The integration of comprehensive and accurate persona details ensures that the chatbot can respond in a manner that is both contextually appropriate and emotionally resonant. This level of detail is particularly important for therapeutic applications, where the authenticity of interactions can significantly impact the training outcomes for speech-language pathology (SLP) students and clinicians

Central to the application’s behavior was the use of OpenAI’s GPT API. Specifically, the ‘chat.responses‘ endpoint was used to send structured messages to the GPT-4o model and receive text responses. Each interaction began by embedding a background prompt into the API call—this prompt included a detailed description of the chatbot’s persona (Monae Jackson), her emotional profile, conversational boundaries, and role in the training context. The prompt served as a guiding framework, allowing the model to maintain a consistent character across multiple interactions.

Certain instructions within the prompt appeared to have a stronger influence on guiding the chatbot’s behavior than others. For example, the directive “Be expressively stubborn, and stay dug into your personality as Monae” was included to reduce the likelihood of the model reverting to the overly cooperative or assistant-like behavior typical of default ChatGPT responses. Similarly, the instruction “Do not willingly give information” was designed to shape Monae into a more reserved and realistic client, rather than one who readily discloses personal details without prompting.

These two statements were particularly effective in reinforcing Monae’s role as a guarded, emotionally complex individual. By limiting unsolicited disclosure and encouraging a more assertive interaction style, these instructions helped produce a more believable and educationally valuable simulation. Development proceeded using an iterative testing approach. Initial tests focused on simple roleplay responses and evaluating the model’s ability to retain character tone. As interactions were tested, adjustments were made to the system prompt and message formatting to improve consistency, emotional realism, and boundary-setting. Specific language was added or refined to correct drift in tone or assistant-like behavior, particularly when the model defaulted to generic helper responses rather than remaining in the role of a patient.

Sentiment analysis was incorporated into the system as a mechanism for managing conversational boundaries. Specifically, it was used to help the chatbot detect and respond to sustained negative interactions. When Monae encountered clinician input that reflected disrespect or insensitivity, the sentiment analysis tool assessed the polarity of the user’s language. If negative sentiment was detected more than once during a session, the chatbot was programmed to interpret this pattern as emotional harm and to end the conversation accordingly. This design choice aligned with the prompt’s directive for Monae to assert boundaries and take minor offenses seriously, while also introducing a quantitative method to trigger a realistic behavioral outcome—termination of the session. The inclusion of this mechanism supported the chatbot’s role as a simulated patient with emotional awareness and contributed to maintaining the integrity of the persona in extended or potentially inappropriate interactions.