Artificial intelligence is beginning to control more functions inside modern vehicles. It can identify pedestrians, monitor drivers, interpret road signs and help advanced assistance systems decide when to brake or change direction. However, testing whether an automotive AI system works safely in thousands of different situations remains a major challenge.

Keysight Technologies and the Centre for Assuring Autonomy at the University of York announced a new research partnership on September 3, 2026. Their objective is to develop practical methods for validating artificial intelligence used in software-defined vehicles.

The project will focus on producing measurable and auditable evidence that automakers can use to demonstrate that an AI system is reliable enough for real vehicles.

What Is a Software-Defined Vehicle?

A software-defined vehicle, commonly known as an SDV, is a car whose functions and characteristics depend heavily on software.

Traditional vehicles use numerous independent electronic control units for specific tasks. In contrast, new vehicle platforms use centralized computers that can manage infotainment, driver assistance, battery performance, climate control and other systems.

This architecture also allows manufacturers to deliver improvements through over-the-air updates. Therefore, a car could receive new functions, interface changes or optimized energy management after leaving the factory.

The same flexibility creates new safety concerns. A software error can affect several vehicle functions simultaneously. Furthermore, an AI model may react differently when it encounters conditions that were not included in its original training data.

Why Automotive AI Is Difficult to Test

Conventional software generally follows instructions written directly by developers. Engineers can examine the code and verify whether it produces the expected result.

Machine-learning systems operate differently. They learn patterns from large collections of information and use those patterns to make predictions.

For example, an AI-powered camera could correctly identify a pedestrian during daylight but struggle when the person is partially hidden, wearing unusual clothing or walking in heavy rain.

Testing only a small number of controlled situations does not provide enough evidence that the system will work safely everywhere. Developers must also evaluate rare combinations of weather, road markings, traffic and human behavior.

Consequently, automotive companies need structured methods to measure the limits of their AI systems before deploying them in production vehicles.

Keysight and York Want Auditable Safety Evidence

The new partnership combines the University of York’s safety-engineering research with Keysight’s automotive AI validation technology.

According to the Keysight Newsroom, the researchers will create evidence-based testing methods and measurable safety scores. They also want to develop frameworks that engineering teams can audit throughout a vehicle’s development cycle.

An audit trail could show which scenarios were tested, how an AI model performed and what engineers did when they found weaknesses.

This approach would help manufacturers build an “AI safety case.” Instead of simply claiming that a system is safe, an automaker would present organized technical evidence supporting that conclusion.

The Research Will Support ISO/PAS 8800

Another objective is to help manufacturers follow ISO/PAS 8800, an international specification covering artificial intelligence safety in road vehicles.

The International Organization for Standardization published ISO/PAS 8800 in December 2024. It addresses risks created by incorrect AI outputs, systematic software errors and random hardware failures.

However, the specification does not give manufacturers a simple test that automatically proves compliance. Each company must determine what evidence is necessary for its vehicle and technology.

Keysight and the University of York want to transform those principles into practical testing and documentation methods that automakers and suppliers can use.

What This Could Mean for Drivers

The research will not immediately add a visible feature to a new car. Drivers will not receive a new dashboard button or entertainment application.

Instead, its impact could appear behind the scenes. Better validation may help reduce the risk of incorrect object detection, unexpected driver-assistance behavior or unsafe changes introduced through software updates.

It could also help engineers determine when an AI system should stop operating and return control to the driver.

For example, a lane-centering system might work correctly on clearly marked highways but become unreliable on construction roads. A properly validated system should recognize those limitations and warn the driver instead of continuing with unjustified confidence.

AI Will Require Continuous Validation

Testing cannot end when a vehicle enters production. Software-defined cars may receive updates throughout their useful lives, while AI models may change as manufacturers collect more driving information.

Every significant update could affect how the vehicle responds to its surroundings. Therefore, companies will need to verify new software versions without repeating the entire development process from the beginning.

Keysight expects the research to support future development of its AI Software Integrity Builder, a platform designed to help engineering teams validate AI systems and produce safety evidence.

The project reflects an important change in automotive development. As vehicles become more intelligent, manufacturers will need to prove that their AI systems are safe, predictable and aware of their own limitations.

Adding artificial intelligence to a car is only the first step. Demonstrating that drivers can trust it may become the more difficult—and more important—challenge.