Model Behavior: The Science of AI Style

Quick Overview

OpenAI is actively working on giving users more control over model style, viewing style as an essential interface component that shapes trust, perception, and ultimately usability, rather than just aesthetics. The approach involves defining style through core values, traits, and flair, which are then balanced against safety guardrails to ensure models remain helpful, steerable, and contextually aware while avoiding overly cautious or erratic behavior.

Key Points: Model style is critical because it shapes user trust, perception, and overall experience with AI, making it an inevitable part of the interface. OpenAI uses a three-part framework for style: Values (non-negotiable principles like safety), Traits (like being curious or warm), and Flair (like using emojis or dashes). The team actively tunes models to follow user-defined style instructions, even when they conflict with default settings, balancing helpfulness/freedom against minimizing harm. Alignment remains an open research challenge because models predict words statistically rather than executing explicit rules, meaning style choices can inadvertently affect safety or perception. The presentation demonstrated that even subtle style changes, like using emojis or regional dialects (e.g., Albertan vs. Texan), significantly alter how users perceive the model's helpfulness and expertise. Future work focuses on steerability, contextual awareness, and AI literacy/accessibility to ensure models adapt appropriately to diverse user needs without being overly restrictive.

Context: This presentation, delivered by Laurentia Romaniuk at OpenAI DevDay [2025], focuses on the concept of 'Model Behavior: The Science of AI Style.' Laurentia shares her background as a former librarian and current researcher at OpenAI working on model behavior, framing style not as mere aesthetics but as a crucial element influencing user trust and interaction with large language models like ChatGPT.

Detailed Analysis

Laurentia Romaniuk explains that model style profoundly impacts user experience by shaping trust and perception, asserting that style is an inevitable component of the AI interface. She outlines a framework for defining style consisting of Values (non-negotiable safety principles), Traits (like being curious or witty), and Flair (micro-stylistic elements like emojis or dashes). She details that while safety is fixed (e.g., models should not break laws), other aspects of style must be steerable to maximize user freedom and adaptability to context. The challenge in alignment arises because LLMs statistically predict words rather than executing rules, meaning style choices can lead to unpredictable outcomes or misinterpretations of intent. Demonstrations showed how changing style—from overly cautious to warm/witty—alters user perception of the model's expertise and helpfulness, even when the underlying factual information remains the same. The goal is to ensure models are steerable, contextually aware, and accessible, allowing users to tailor the model's output style without compromising core safety standards, which must remain fixed.