Participants solved text-based analogies built from evolving letter sequences. Each puzzle required recognising a rule and applying it to new “alphabets”, including the Greek alphabet and one composed of unfamiliar symbols. This created a level playing field for children, adults and AI.

Children achieved an average score of 67 percent in the symbol alphabet, whereas AI models dropped sharply and in some cases dipped below 20 percent. According to lead researcher Claire Stevenson, this highlights a fundamental difference between human and artificial reasoning: children intuitively understand ordered structures such as alphabets, while AI relies on patterns it has previously encountered.

The researchers conclude that the ability to apply knowledge flexibly in new contexts remains a core feature of human cognition. They also note that insights from psychology – including how babies learn – increasingly inform the development of AI systems. The article Can large language models generalize analogy solving like children can? will appear in the January 2026 issue of Transactions of the Association of Computational Linguistics and is already available online. The research was co-funded by NWO within the project Learning to solve analogies: Why do children excel where AI models fail?