Sony AI has discovered a way to use ChatGPT-like technology to automatically design racing AI agents with distinct personalities, potentially transforming how future Gran Turismo games create their artificial opponents. Researchers presented a system at the NeurIPS 2025 conference that lets developers describe how they want an AI to behave in plain English, then uses OpenAI's GPT-4o to generate the underlying code that trains the racing agent. The breakthrough could eliminate weeks of manual tuning work and open the door to dramatically more varied single-player racing experiences.

How Does This Technology Actually Work?

The process is surprisingly elegant. A developer types a description of desired AI behavior, such as "win races while obeying standard motorsport racing rules and maintaining good sportsmanship." That text is handed to GPT-4o, which generates the reward function code automatically. This code trains a racing agent, and a vision-language model evaluates the results by watching the agent drive. The system repeats this cycle over several iterations until the behavior converges on what was requested.

When Sony AI researchers tested this approach with deliberately vague instructions, the results were striking. Out of ten runs using the sportsmanship prompt, every single one produced an agent that met the team's standards for clean racing. Three of those agents actually outperformed the hand-tuned GT Sophy, the legendary AI racer that made headlines in 2022 when it appeared on the cover of Nature.

What Happens When You Change the Prompt?

The real excitement emerges when developers experiment with creative instructions. The Sony AI team tested more unconventional prompts like "race as fast as possible in reverse at all times" and "drift as much as possible while otherwise obeying standard motorsport rules." Both produced working agents that did exactly what was asked: an AI that drifts through corners on command and an AI that races backward around Lake Maggiore, all generated from a single sentence.

This flexibility suggests that future Gran Turismo titles could feature dozens of distinct driving personalities without requiring the extensive manual work that went into creating the original GT Sophy. Instead of all AI opponents driving in roughly the same way, imagine racing against opponents with genuinely different characteristics and strategies.

Steps to Implement This Technology in Future Games

• Define Personality Traits: Developers could write simple descriptions for each AI opponent type, such as "aggressive late-braker," "cautious tire-saver," or "fast in wet conditions but makes mistakes under pressure."
• Generate Reward Functions Automatically: GPT-4o converts these descriptions into the mathematical rules that guide the AI's behavior, eliminating the need for expert reinforcement learning engineers to hand-craft every detail.
• Iterate and Refine: The vision-language model watches the AI drive and provides feedback, allowing the system to adjust the reward function over multiple cycles until the behavior matches the original description.
• Deploy Diverse Opponents: Once validated, these AI agents can populate single-player events with genuinely varied opponents, creating a richer and more lifelike racing experience.

Could B-Spec Mode Make a Comeback?

Long-time Gran Turismo fans immediately recognize the implications for B-Spec mode, which hasn't appeared since Gran Turismo 5. This feature let players take on the role of a race director, managing an AI driver from the pit wall using crude sliders for pace and aggression. The technology described in Sony AI's research could be the foundation for something far more sophisticated.

Imagine instructing your B-Spec driver to "qualify aggressively but conserve tires during the race," or "stay close to the leader for the first two laps, then push hard." Instead of nudging a slider and hoping for the best, players would be giving their AI driver meaningful strategic instructions in natural language. This represents a fundamental shift in how players interact with AI teammates.

Gran Turismo creator Kazunori Yamauchi himself hinted at the possibility of a Sophy-powered B-Spec mode when the technology was first revealed. Research like this brings that vision considerably closer to reality and makes it far more exciting to think about what it could actually look like.

What Are the Current Limitations?

It's important to note that this is academic research from Sony AI, not a product announcement from Polyphony Digital. The system still requires significant computational resources; each iteration involves training reinforcement learning agents on high-end graphics processing units (GPUs), a process that takes days rather than seconds. Real-time, user-facing prompting where players type race strategies into a text box during gameplay remains further out.

However, even without real-time capabilities, this technology could fundamentally change how Polyphony and Sony AI develop AI behaviors for Gran Turismo going forward. What once required weeks of expert tuning could eventually be accomplished in a fraction of the time, opening the door to a much richer variety of AI-driven racing experiences. The gap between "describe what you want" and "get an agent that does it" is closing, and the implications for future racing games are genuinely exciting.