GM's Proprietary In-Vehicle AI Is Coming
General Motors will launch a proprietary in-vehicle AI assistant later this year, marking a significant shift in how the automaker thinks about artificial intelligence inside the car. The system will be powered by an unnamed large language model provider and will integrate deeply with OnStar, GM's telemetry systems, and proprietary vehicle knowledge. It will go beyond the capabilities of Google Gemini, which currently handles conversational queries, temperature adjustments, and radio controls in eligible 2022 and newer GM vehicles. According to Anna Santos, GM's director of product management for voice and AI, there is a limit to what an AI that simply sits at the top level of the vehicle can do.
“There’s a limit to what an AI that’s just sort of sitting at the top level of the vehicle can do,” Santos said. The new assistant is meant to understand the vehicle's mechanical state, driving patterns, and owner preferences. That means it will be able to combine live telemetry with historical trends to answer questions about a warning light, suggest when service is needed, and even anticipate failures before they happen. Unlike a general-purpose assistant, it will know which vehicle it is installed in, whether it is an electric or combustion model, and how that specific car has been driven over time.
Key facts at a glance
- GM will launch a proprietary in-vehicle AI assistant later this year.
- The assistant integrates OnStar, telemetry, and predictive maintenance.
- An unnamed large language model provider is supporting the system.
- Google Gemini remains for general queries and entertainment.
- A “kids setting” command would adjust music, seats, climate, and door locks simultaneously.
Why GM needs its own car-aware assistant
Voice assistants have become a standard feature in modern vehicles, but most still operate at the surface level. They can change the temperature, select a playlist, or read a text message aloud. They generally cannot tell a driver why the check-engine light is flickering, whether the 12-volt battery is losing capacity, or how a particular set of driving conditions is affecting tire wear. According to GM, that deeper layer of vehicle knowledge is exactly where its new assistant will operate.
The distinction becomes clearer in real-world use. Santos described a “kids setting” command that would adjust music, seats, climate, and door locks simultaneously. That is not simply a single media command. It is a coordinated action across multiple vehicle subsystems, triggered by a natural language phrase and executed in a safe, deterministic sequence. For that to work, the AI needs access to the vehicle's internal architecture, which is not something a cloud-based assistant like Google Gemini has by default.
Gemini stays, but the boundaries are clear
GM is not replacing Gemini. The two assistants will coexist, and each will have a defined role. Gemini will continue to handle general conversational requests, entertainment, and other information-focused tasks. The GM assistant will handle everything the car knows about itself: battery state, diagnostic codes, service history, driving patterns, and other data generated by the vehicle's own sensors and control units.
This division of labor is commercially important. Gemini operates essentially like a phone assistant projected onto the dashboard. GM's proprietary system, by contrast, uses data that belongs to GM, not Google. “It’s data that’s going to be proprietary to GM,” Santos said. By keeping vehicle intelligence on its own side, GM avoids handing Google the most valuable and sensitive information generated by its cars.
Data ownership and the single-vendor trap
Google has been embedding AI across every product it controls, from search to Android to Android Automotive. Automakers were initially eager to integrate Google's tools, but some have begun to reassess that relationship. GM's decision to build a parallel system rather than hand all vehicle intelligence to Gemini reflects a concern similar to the one Microsoft CEO Satya Nadella has expressed about companies outsourcing their thinking to a single AI provider.
Nadella has often warned that organizations risk losing their institutional knowledge if they let one AI platform define how they operate. The same logic applies to automakers. If every car conversation is processed by an external model, the underlying data becomes harder to use for proprietary services, product development, and customer retention. A company's ability to build its own AI features depends on having access to its own data in a usable form.
GM's decision to work with an unnamed large language model provider reinforces this point. The company is not publicly committing to a single model vendor. It is keeping its options open, likely evaluating multiple providers as the technology evolves. That flexibility is especially important in a fast-moving market where model costs, capabilities, and licensing terms change rapidly.
OnStar as the foundation
GM's telemetry advantage is not new. OnStar launched in 1996 as a safety and communications system that connected drivers to human advisors. Over time, it expanded into remote diagnostics, turn-by-turn navigation, crash response, and vehicle health reports. Today, OnStar can send data about engine performance, tire pressure, oil life, and battery condition directly to GM servers. The new AI assistant is designed to turn that existing infrastructure into a conversational interface.
Rather than waiting for a driver to call a service center, the system can proactively generate a clear explanation of a vehicle issue and recommend next steps. For example, if coolant temperature has been running high, the assistant could tell the driver that the system is monitoring the issue, suggest a service appointment, and even check current dealership availability. If a sensor fails, the assistant could explain the impact on vehicle performance and clarify whether immediate attention is required.
Predictive maintenance and fleet operations
Predictive maintenance is one of the most commercially valuable applications of vehicle-aware AI. By analyzing telemetry data over time, the assistant can identify patterns that indicate future failures. A slight drop in battery voltage under load, a growing vibration in the braking system, or a gradual change in energy consumption can all signal that maintenance is needed before a breakdown occurs.
For private owners, this means fewer surprise visits to the repair shop and a clearer understanding of what the car is doing. For fleets, the value is even higher. GM is pushing its autonomous driving capabilities toward robotaxi-readiness, and an in-vehicle AI that understands predictive maintenance and telemetry feeds directly into fleet management for autonomous operations. A robotaxi that can report its own health and schedule maintenance before a problem becomes critical could save operators significant time and money.
Competitive pressure across the auto industry
GM is not alone in wanting a more car-aware AI. Many automakers are experimenting with generative AI assistants that can access vehicle manuals, answer driver questions, and provide natural language control of car functions. Some are working with major cloud providers, while others are trying to build or license smaller models optimized for automotive use. The challenge is that an AI assistant must be both helpful and safe, and vehicle control requires an extra layer of reliability.
Automakers also face a choice about how much of the AI experience should be developed in-house. Handing everything to a Big Tech platform can accelerate development, but it also weakens the automaker's connection to the customer. By building its own layer, GM can create services and subscription opportunities around vehicle data while still benefiting from a large model's natural language abilities.
Engineering and safety challenges
There are serious engineering challenges in building an in-car AI that accesses sensitive vehicle systems. Latency is a major concern: drivers are unlikely to wait several seconds for a response to a question about a warning light. Some data will need to be processed locally in the vehicle, while other queries can be sent to the cloud. The system will need to determine what can be answered quickly with onboard data and when a more powerful cloud model is needed.
Reliability and safety are equally important. An AI that can adjust seats, climate, and door locks must be designed to avoid dangerous actions. Voice commands should be rejected when they conflict with safety rules. The AI also needs to understand when it does not know something, rather than guessing about a maintenance issue. That kind of careful orchestration is difficult, and automakers have to invest heavily in testing before deploying it across a production fleet.
Privacy is another concern. The same telemetry that makes predictive maintenance possible is a highly detailed record of how a car is used. GM has a long history with OnStar data, but adding an AI layer that can reason about that data raises new expectations around consent, transparency, and data retention. If drivers believe their car is being monitored in ways they do not understand, adoption could stall.
What remains unchanged
For now, GM will continue to offer Google Gemini in its vehicles, and the user experience will include both assistants without forcing drivers to choose sides. Gemini handles the broad world of knowledge and conversation. GM's new assistant handles the narrow but powerful world of automotive insight. The two systems may even work together, with Gemini referring a vehicle-specific question to the GM assistant or the GM assistant calling on Gemini for general information.
Integration and expansion
GM has called the new assistant “the beginning of a broader AI journey.” That suggests the technology will expand over time. Early versions may focus on diagnostics and maintenance, but future capabilities could include more proactive coaching for electric vehicle drivers, trip planning that accounts for charging conditions, and personalized vehicle settings that adapt as different drivers enter the car.
The unnamed LLM provider also suggests GM is keeping its options open as the market evolves. Building a proprietary layer now does not mean choosing one AI provider forever. It means creating the interface, data infrastructure, and user experience that can sit on top of different models as they improve.
In the end, the race is not just about which AI can hold a better conversation. It is about which AI can understand the car, the driver, and the journey well enough to act with confidence. General Motors is betting that the most valuable assistant is the one that knows what the vehicle knows.