Neural Transparency: Mechanistic Interpretability Interfaces for Anticipating Model Behaviors for Personalized AI

Millions of users now design personalized LLM-based chatbots through system prompts that shape their daily interactions, yet have limited ability to anticipate how these design choices will manifest as behaviors in deployment. This opacity is consequential: seemingly innocuous prompts can trigger excessive sycophancy, toxicity, or other undesirable traits, harming utility and raising safety concerns. To address this, we introduce an interface that enables neural transparency by exposing language model internals during the chatbot's personality design. Our approach extracts behavioral trait vectors (empathy, toxicity, sycophancy, etc.) by calculating the differences in neural activations between contrastive system prompts that elicit opposing behaviors. We quantify a chatbot's personality by projecting the system prompt's final token activations onto these trait vectors to create persona scores, which are then normalized for cross-trait comparability and visualized using an interactive sunburst diagram. To evaluate this approach, we conducted an online user study (N = 80) to compare our neural transparency interface against a baseline chatbot interface without any form of transparency. Our analyses suggest that users systematically miscalibrated AI behavior: participants misjudged trait activations for 11 of 15 analyzable traits, motivating the need for transparency tools in everyday human-AI interaction. While our interface did not alter design iteration patterns, it significantly increased user trust and was enthusiastically received. Qualitative analysis revealed nuanced user experiences with the visualization, suggesting interface and interaction improvements for future work. This work offers a path for how mechanistic interpretability can be operationalized for non-technical users, establishing a method for safer, more aligned human-AI interactions.