“AI needs a better understanding of reality to be safe and effective”

Albert Gatt, Professor of Natural Language Generation, delivered his inaugural lecture on 9 December. He delved into the complex connections between language and the world and pointed out that current AI models are not fully there yet. Gatt’s research focuses on enhancing AI’s understanding of reality, which is essential for effective and safe AI. Gatt was appointed professor at Utrecht University in 2023.

At the heart of Gatt’s lecture is the concept of grounding: the idea that the meaning of words stems not just from their relation to other words, but from real-world experiences and perceptions. He argues that for AI systems to truly comprehend language, they must effectively bridge the gap between symbolic representations (such as words) and the non-symbolic realm of perception, experience and fact.

Current state of AI

Gatt points out that while modern AI models, such as ChatGPT and multimodal systems like CLIP, have made remarkable strides in processing language and visuals using massive datasets, they often fall short of achieving full grounding capabilities. For example, these models excel in straightforward tasks, such as object recognition, but struggle with abstract concepts like temporal relationships and complex interactions.

More humanlike

Gatt explains that his group is exploring new approaches to improve AI’s understanding of the real world. One way to do this is by studying how humans describe objects and events, and apply this to AI, making AI more humanlike. How do we answer questions like "What is happening in this picture?" or "Why is it happening?". Ultimately, this could make AI systems more intuitive and context aware. The techniques explored by his group also aim to make AI systems more interpretable for humans, enabling us to understand why they behave the way they do.

Misinformation and misinterpretation

The practical applications of AI are growing. However, mistakes in understanding can lead to significant consequences, such as spreading misinformation. Ensuring that AI is reliable and interprets language accurately is essential for safe and effective use.

As AI becomes more integrated into everyday life, Gatt’s research highlights a critical goal: designing systems that don’t just sound intelligent but genuinely understand the meaning behind their words.