When an AI model is given a simple arithmetic problem framed as a short story, changing only the names of the characters should not affect the outcome. However, research shows that such minor adjustments can influence the performance of GPT models. Humans adapt easily to these variations, while language models may respond inconsistently.

According to Lewis, this is rooted in the architecture of current AI systems. These models are not inspired by human cognitive structures but are trained on vast datasets using machine learning. While this approach yields strong performance within known patterns, it can result in weaker generalisation to new or slightly altered inputs.

Lewis therefore works on neurosymbolic AI, an approach that combines symbolic AI, with explicit logic and formal representations, and machine learning. The aim is to develop models with an explainable internal structure and a stronger theoretical foundation. By explicitly incorporating logical structures, systems can move beyond pattern recognition and better organise and generalise reasoning processes.

A key part of this research involves opening up the “black box” of AI. Although language models are highly successful in practice, it is often unclear how they arrive at specific outputs. At ILLC, researchers are developing methods to better understand these internal mechanisms. The scale and complexity of modern AI models present significant challenges in this effort.

In 2024, Lewis was appointed MacGillavry Fellow, enabling her to establish her own research group. Her work builds on collaborations within ILLC, with psychologists and international partners. The long-term ambition is to develop AI systems that align more closely with human reasoning and operate in a more transparent manner.