AI systems that understand both images and text play a key role in technologies like search engines and autonomous vehicles – but they remain vulnerable to manipulation and bias. Researchers at L3S have developed a new training method that improves the reliability and fairness of these systems.