TÜV Nord develops test methods for AI and LLMs

With more than 14,000 employees at over 500 locations, the TÜV Nord Group is active in around 100 countries. At their annual press conference, the Group presented figures as well as plans for testing large language models (LLMs) and other AI applications. The aim is to be a pioneer in testing methods to make the use of artificial intelligence safe for of all users, CEO Dirk Stenkamp explained in Hanover on Wednesday.

The European AI Act provides for such test criteria, although the concrete form is still being worked out. According to the Act, AI systems with certain risk factors must undergo a testing or certification process. Such risk factors include, for example, whether an AI system has been trained unilaterally and whether this could be problematic for certain applications. It should also not be possible to misuse AI applications - this includes ways of circumventing the barriers of an LLM. In addition, an AI system should not be vulnerable to attack.

Of course, none of these problems can be solved with a single test method. So far, there are no methods to prevent prompt injections, for example, in which the built-in barriers of AI models are bypassed with the user's own instructions. However, gradations can be made in the individual areas and areas of application can be found or excluded depending on this.

AI: hype with use cases

However, Stenkamp also points out that AI is currently a "marketing hype" that will not last. It remains to be seen what use cases there are and what the real market will be. AI will certainly remain an important topic. TÜV Nord itself has pooled all of its AI expertise in a TÜV AI Lab in Berlin. "With the TÜV AI Lab in Berlin, which was launched in 2023, we are driving forward the development of methods and standards for testing safety-critical AI applications in particular," says the head of TÜV Nord.

TÜV Nord also already has AI in use, including a chatbot and a sensor data platform. In addition, wind turbines, for example, are already being inspected using AI-based image analysis. TÜV Nord has developed a GPT for all employees. All of its own apps are designed to support employees and it is hoped that AI will reduce bureaucracy - for example by eliminating paperwork.

TÜV Nord's portfolio also includes a stake in 3spin Learning - a company that creates software for immersive learning experiences. Using augmented reality, repair work on a wind turbine can be learned and you can also test your ability to work at height. However, such applications have been costly to develop up to now, and the companies are working together on new ways to shorten the development process and make it cheaper.

LLMs can help train soft skills in future learning scenarios. For example, a chatbot can simulate an unruly flight passenger or an unhappy caller.

TÜV Nord is also working closely with business and politics in the area of the energy transition. In future, it will be possible to have the battery of an e-car checked at TÜV, and the focus is on the used car market for e-cars. There is a certification offer for sustainable raw materials, as well as a seal that confirms that products are suitable for use in hydrogen applications.

(emw)