What Happens to Your Voice When a Mercedes Talks to ChatGPT?

The assistant in your dashboard now has a cloud behind it

A car used to hear "call Mom" and match it against a short list of commands. A car with a large language model behind it hears "I'm tired, find somewhere with decent coffee that's not off the route" and has to send that sentence somewhere with enough computing power to parse it. That trip is where the privacy questions start.

Mercedes is putting large-model assistants in its cars. The more useful question is what leaves the vehicle. What Mercedes has said publicly over the past three years is enough to trace that path, and to mark where the answers stop.

What is documented

In June 2023, Mercedes opened a US beta that put ChatGPT, delivered through Microsoft's Azure OpenAI Service, into its MBUX voice assistant. According to Mercedes' announcement, the beta covered more than 900,000 vehicles, and owners opted in by saying a phrase or using the Mercedes me app. The company said it kept control over the IT process and would use beta data to improve the service.

In 2025, Mercedes introduced the CLA, the first car on its in-house MB.OS software platform, with a new MBUX Virtual Assistant. The company has described that assistant as drawing on models from more than one vendor, including Google's Gemini via Google Cloud's automotive AI agent. The assistant is not one program. It is a front end that hands requests to whichever model suits the task.

Wake-word detection, the "Hey Mercedes" part, is generally handled on the car itself. Everything after the wake word is where the cloud usually comes in. That is the standard design across voice assistants, not a Mercedes quirk.

Why a chatbot changes the data picture

Interpretation: The older command-style system needed little context. A modern assistant is more useful the more it knows: your location, the route, the battery level, the calendar, the contacts, the time of day. Each of those becomes an input that may travel with your spoken request to a server.

That bundling is the real change. A recording of a voice saying "find a pharmacy" is one thing. The same sentence stapled to a GPS position, a destination history and a vehicle identifier is a behavioral record. Whether it is stored, for how long, and under whose name are policy questions, and the hardware cannot answer them.

There is also the matter of who listens. In 2019, Bloomberg reported that Amazon employed reviewers to listen to Alexa clips, and The Guardian reported that Apple contractors graded Siri recordings. Apple apologized and changed its program. Voice assistants improve through human review more often than their marketing suggests, a pattern documented in the earlier piece on the hidden human work behind AI. Nothing in the public Mercedes material says the company does this, but readers should check whether the privacy notice rules it out.

The car industry's baseline is low

In 2023, Mozilla's Privacy Not Included project reviewed 25 car brands and gave every one a privacy warning. According to Mozilla, 84% of the brands said they share personal data and 76% said they can sell it. Mozilla also reported that 56% would hand data to law enforcement in response to an informal request, with no court order.

In January 2025, the US Federal Trade Commission announced a proposed order against General Motors and OnStar. The FTC alleged the companies collected precise location and driving behavior data and sold it to consumer reporting agencies without clear consent. The proposed order would ban that sharing for five years.

Interpretation: I am not suggesting Mercedes does any of this. The point is that an AI assistant is being added to an industry whose default data practices have already drawn a regulator's attention. A more talkative interface sits on top of the same data-sharing machinery, and the consent screen is usually shorter than the machinery is complicated. The default setting often functions as the policy, and few drivers change it at a red light.

What is still unknown

Several answers are not in the public record, at least not in a form a typical owner can find:

• How long spoken requests and the context attached to them are kept after a session ends.
• Whether audio, text transcripts, or both are retained, and whether they can be linked to a vehicle identification number.
• Whether third-party model providers are contractually barred from training on the data.
• What the car can still do with no connection, since an offline fallback limits exposure.

Those answers live in the privacy statement and the terms of service that now ship like firmware, which can change after the sale through an app update. The version you accepted at the dealership may not be the version in force next year.

What a careful owner can check

This is general consumer guidance, not a verdict on any brand. Start in the vehicle's data and privacy menu and the companion app. Look for a switch that separates "improve the product" from "operate the feature," because they are different permissions. Check whether voice history can be viewed and deleted, and whether deleting it in the car also deletes it on the server.

Then read the section on third parties. The relevant phrases are "service providers," "affiliates" and "partners," and they mean different things. If the document says the data may be shared for "business purposes," that is a prompt to look for the definition, not a reason to relax.

What is likely to happen next

Prediction: Automakers will keep routing requests among several model vendors, because no single model is best at navigation, conversation and vehicle control. That makes the privacy chain longer, and the weakest contract in it sets the real protection level.

Prediction: Retention and training terms will become a selling point within a few model years, in the way crash-test ratings once did. Europe's GDPR and California's privacy law already give regulators grounds to ask, and the FTC's GM action shows US enforcers are willing to examine vehicle data.

Prediction: The first serious dispute will not be about a dramatic breach. It will be about a quiet policy change, such as a retention window extended or a training permission added, delivered by software update to cars already on the road.

An AI upgrade can make a car's assistant meaningfully better at understanding people. Whether it also makes the car better at remembering them is a separate question, and the answer is written in a document most buyers have never opened.