As we increasingly delegate tasks to artificial intelligence, the need for reliable AI systems has never been more crucial. In a recent interview, Pinar Yolum, Professor of Reliable AI at Utrecht University, emphasized the importance of building AI systems that can not only perform tasks autonomously but also express uncertainty when necessary. She believes that AI systems should be capable of stating when they do not know the answer, rather than providing incorrect information – a key step in building trust.

“Trust is hard to build but easy to break,” Yolum states, highlighting that AI systems must be tested thoroughly and able to explain their decision-making processes. Moreover, she argues that AI should be able to refuse or challenge tasks that go against the norms and values it has been trained to uphold.

However, Yolum acknowledges the challenges posed by conflicting interests in the development of AI, especially when developers work for companies whose goals may not always align with the best interests of users. She suggests that regulatory frameworks could help ensure that AI developers follow best practices, just as we trust manufacturers to meet safety standards without needing to check every detail ourselves.

Looking ahead, Yolum believes that it will be possible to delegate large tasks to AI without human oversight – provided the systems are tested, their limits understood, and proper safeguards are in place.

To read the full interview (in Dutch), click here.

Photo: Bram Belloni