Mercedes is working with Microsoft and is starting an initial three-month test phase.
(Photo: Daimler AG)
New York, San Francisco Mercedes-Benz wants to improve the voice control of its vehicles and is testing the text robot ChatGPT from Microsoft partner OpenAI. As the car manufacturer announced on Thursday, a corresponding test phase should start this Friday.
Accordingly, Mercedes customers in the USA whose cars are equipped with the MBUX infotainment system can take part: more than 900,000 vehicles. Registration should be possible via app.
Drivers who call up the voice control with the command “Hey Mercedes” should be able to communicate with the car “even more intuitively” thanks to artificial intelligence (AI). The language assistant should be able to tell interesting things about destinations, suggest new recipe ideas or clarify knowledge questions. The special thing about ChatGPT is not only that complicated inputs can be understood, but also that longer dialogues with further questions are possible.
“The integration of ChatGPT is a real milestone on our way to making the car the center of digital life,” said Mercedes Chief Technology Officer Markus Schäfer. The pilot program adds the capabilities of ChatGPT to existing “functions such as navigation input, weather query and others” to enable “conversations with natural dialogue and follow-up questions”.
The system could be equipped with additional functions, such as booking a table in the restaurant or buying a ticket for the cinema, said the AI boss of Microsoft’s cloud spare, Eric Boyd. The most powerful language models from OpenAI are already running in the background for the system: GPT-4 and GPT-3.5.
The test phase of the AI voice control is initially limited to three months. Based on the results of this beta test, Mercedes is checking whether a large language model for “dialogical communication” should be offered in its vehicles in the future, a spokesman said when asked by Handelsblatt. “We’re starting here with ChatGPT as that’s the market-leading model at the moment.”
Fight against hallucinations
A topic that concerns not only ChatGPT but also all other providers of large language models are so-called hallucinations: errors and false statements by the AI. In the car, where drivers have to concentrate on the road holding, such would be particularly problematic.
Mercedes is testing a special way to suppress hallucinations: its own system cross-checks the AI information. “In order to rule out hallucinations, we rely on a plausibility check of the ChatGPT outputs,” said the spokesman. The Mercedes Intelligent Cloud should check so-called “points of interest” recommendations – i.e. references to restaurants, petrol stations or other destinations – to see whether these places really exist.
According to company circles, Mercedes uses its own validated data and, for example, search results from Google. You have “at all times the sovereignty over the IT processes in the background,” emphasized the carmaker.
Fictitious statements in court
Errors and misstatements are the biggest problem for mass deployment of large language models like GPT-4. Despite the ever-improving linguistic quality of models from companies like OpenAI, they regularly produce fictitious statements or invent complete sources.
Nevertheless, a number of companies are already using the software – with sometimes serious consequences. New York attorney Steven A. Schwartz was convicted in a court case of not only having ChatGPT create documents, but also making dramatic mistakes in the process.
>>> Read here: Is the AI Act slowing down AI development in Europe?
Schwartz represented a client in a lawsuit against the airline Avianca. During a flight, his client is said to have been injured in the knee by an airplane trolley. In the documents filed in court, Schwartz cited alleged other cases designed to support compensation for his client.
The judge examined the documents but could not find the cases cited. When asked, Schwarz admitted that ChatGPT had presented the cases to him and that he had not checked them. Judge P. Kevin Castel wrote, “Six of the cases presented appear to be falsified court decisions with falsified citations and falsified internal references.”
Tech companies like Microsoft and Google are working with hundreds of experts to reduce the susceptibility of the systems to errors. However, Microsoft CEO Satya Nadella admitted: “The results are not yet 100 percent reliable.” Both in the presentation of the AI systems at Google and at Microsoft, the carefully selected case studies contained incorrect numbers or incorrect statements.
More: Mercedes Autopilot California allows “Drive Pilot”.