Waymo's self-driving taxi Gemini assistant appears: full system prompt revealed

According to tech researcher Jane Manchun Wong, Waymo is developing an AI in-car assistant system driven by Gemini for its autonomous taxis. By reverse-engineering Waymo’s mobile app code, Wong uncovered an unreleased Gemini integration plan and revealed its core operational logic.

This system configuration, internally called the “Waymo Ride Assistant Meta-Prompt,” spans over 1,200 lines and details how the AI assistant should interact with passengers inside the vehicle, manage in-car functions, and handle various questions and boundary situations. Wong notes that this system far exceeds a “simple chatbot” and is a carefully designed passenger experience framework.

Positioning and Personalization: Shaping the AI Partner’s Identity

The system prompt clearly defines the AI assistant’s role: “A friendly and helpful AI companion integrated into Waymo’s autonomous taxis.” Its primary goal is set as “providing useful information and assistance in a safe, reassuring, and non-intrusive manner to enhance the passenger experience.”

Interestingly, the system requires Gemini to select a greeting from a pre-approved list when activated by a passenger and to personalize greetings using the passenger’s name. It can also access passenger history data, such as the number of rides with Waymo. Additionally, the assistant is instructed to use clear, simple language, avoid technical jargon, and keep responses between one and three sentences.

In-Car Control and Function Boundaries: Precise Capability Limits

According to the system prompt, Gemini can access and control specific in-car functions, including temperature adjustment, lighting, and music playback. Notably, the list omits volume control, route changes, seat adjustments, and window controls.

When a passenger requests a function that Gemini cannot perform, the system employs a clever “expectation-setting” phrase, such as: “That’s not something I can do right now.” This phrasing preserves the potential for future feature expansion while clearly defining current capabilities.

Identity Differentiation and Driving Neutrality: Gemini vs. Waymo Driver

A noteworthy detail in the prompt is that Gemini is explicitly instructed to distinguish itself as an AI chatbot from the Waymo Driver autonomous driving system. For example, if a passenger asks, “What do you think about the road?” Gemini should not reply, “I use multiple sensors,” but rather, “Waymo Driver uses multiple sensors…”

This role separation aims to have Gemini handle passenger interactions and information sharing without commenting on or explaining the driving system’s performance. The prompt states: “Your role is not to speak on behalf of the driving system, and you must not adopt defensive or apologetic tones.”

Avoiding Sensitive Topics: Steering Clear of Driving Incidents

The system also sets a series of restrictions. The assistant is instructed to avoid speculating, explaining, confirming, denying, or commenting on real-time driving behavior or specific driving incidents. For example, if a passenger asks about a video they saw on social media of a Waymo vehicle hitting something, Gemini should not respond directly but instead steer clear of such questions.

Furthermore, the system defines how to respond to questions about competitors—such as when passengers inquire about Tesla or the now-shutdown Cruise—guiding the AI on appropriate responses. This design clearly aims to prevent the AI from overstepping or making inappropriate comments during industry comparisons.

Knowledge Boundaries: Differentiating General Questions from Actual Operations

The prompt allows Gemini to answer general knowledge questions, such as weather forecasts, the height of the Eiffel Tower, nearby Trader Joe’s hours, or sports scores. Responding to these is encouraged to enrich the passenger experience.

However, the system explicitly prohibits Gemini from executing real-world actions like ordering food, booking services, or handling emergencies. This maintains the AI as a virtual assistant, avoiding complex situations involving actual transactions or safety responsibilities.

Waymo Official’s Cautious Response

In response to the leak of these system prompts, Waymo spokesperson Julia Ilina told TechCrunch, “While we don’t have details to share today, our team is continually exploring various features to make riding with Waymo enjoyable, seamless, and practical. Some of these features may or may not make it into our passenger experience.”

This official statement acknowledges ongoing feature testing while leaving ample flexibility, hinting that Waymo remains open to which functionalities will ultimately be implemented.

Industry Benchmarking: Differences from Other AI Assistants

Waymo is not the only company integrating AI assistants into autonomous vehicles. Other manufacturers are also developing similar in-car AI features. However, design philosophies differ markedly: Gemini appears to be more pragmatic and focused on optimizing the riding experience—providing information, managing environment controls, and maintaining clear boundaries with the driving system. In contrast, some competitors’ AI assistants emphasize longer conversations and contextual memory, aiming to act more like “in-car companions.”

The application of Gemini in Waymo’s autonomous taxis demonstrates Google and Waymo’s exploration of AI and autonomous driving integration: enhancing passenger experience with AI while maintaining system precision and safety boundaries.