The "silent journey" of autonomous driving is about to end. Renowned tech researcher Jane Manchun Wong recently discovered through reverse engineering that Waymo is testing the deep integration of Google's Gemini large model into its driverless taxi, creating a vehicle AI companion called "Waymo Mobility Assistant." Although it has not been launched yet, an internal system instruction document with over 1200 lines (codename "Waymo Mobility Assistant Core Instructions") has been fully exposed, revealing how this AI was meticulously designed as a safe, restrained, and highly scenario-based passenger service engine.

More than chatting, it's a "car concierge"

According to the instruction document, this Gemini assistant will have three core capabilities:

- Intelligent Q&A: answering common knowledge questions such as weather, landmarks, and events;

- Environmental control: supporting temperature adjustment of the air conditioning, car lighting, and music playback;

- Emotional comfort: providing soothing responses when passengers are anxious, creating a sense of security.

However, its permissions are strictly limited - it cannot control volume, change routes, adjust seats or windows. If users make requests beyond its authority, the AI must respond in a "forward-looking" tone: "I can't do that right now."

1200 lines of instructions, every word defines "boundaries"

Waymo's constraints on AI behavior are truly "extremely restrained":

- Clear separation of identity: it must clearly distinguish between "Gemini is a conversation assistant, and Waymo Driver is the driving system"; if asked "What do you think about the traffic conditions?" it should reply "Waymo Driver perceives traffic conditions through multi-sensor fusion," rather than "I see..."

- No discussion of driving details: it must not comment on, explain, or apologize for any driving events, even when questioned about accident videos, it should change the topic;

- Minimalist language: replies are limited to 1–3 sentences, avoiding technical terms, striving for "something a primary school student can understand";

- Personalization without overstepping: it can use passenger names and number of rides, but never access sensitive data like trip destinations;

- Standardized response to competitors: there are preset scripts for questions about Tesla, Cruise, and other competitors, to avoid controversy.

Gemini × Waymo: Full-Chain Integration from Training to Companion

This is not the first time Gemini has empowered Waymo. Previously, Waymo had used Gemini's general knowledge base to train the autonomous driving system to handle long-tail scenarios such as "ghost traffic jams" and "detour around construction zones." Now, Gemini extends further into the passenger experience layer, forming an AI dual engine of "external decision-making + internal service."

A Sharp Contrast with Tesla's Grok

Notably, Tesla is building Grok into a "memory-enabled, chatty" in-car companion, emphasizing emotional connection; while Waymo's Gemini focuses on practicality, safety, and zero interference, rejecting excessive anthropomorphism. The differences in their approaches reflect a fundamental disagreement between the two companies on the role of "AI in the car": a companion or a service provider?

Waymo spokesperson Julia Ilina responded: "We continue to explore features that enhance user experience, but we have no specific details to share at this time."