Researchers at the University of Oxford and the Allen Institute for AI study how artificial intelligence (AI) creates language. They focus on large language models (LLMs).
Many think these models generate language by following strict grammar rules. However, this study shows they work differently. The models use analogies, meaning they compare new words to examples they already know, much like humans do.
To test this, researchers compare GPT-J, an open-source language model, with human choices. They look at a common English word pattern. This pattern turns adjectives, words describing things, into nouns by adding "-ness" or "-ity." For example, happy becomes happiness, and available becomes availability. Researchers invent 200 fake adjectives, like cormasive and friquish, which GPT-J has never seen. They ask GPT-J to turn these into nouns, choosing between endings like cormasivity or cormasiveness. Humans also make these choices. The study compares both sets of answers to two thinking models. One model uses strict rules, while the other uses analogical reasoning, relying on similarity to known examples.
AI relies on stored examples
The results show GPT-J acts like humans. It does not follow rules but uses analogies. For example, it turns friquish into friquishness because it resembles selfish. For cormasive, it picks cormasivity, influenced by words like sensitive and sensitivity. The model also reflects patterns in its training data. Researchers test it on 50,000 real English adjectives. GPT-J’s choices match the frequency of word forms it has seen before. It seems to store every word example from its training, using these memories to decide on new words. When faced with unfamiliar words, it asks itself, “What does this remind me of?”
Humans and AI differ in one key way. Humans build a mental dictionary, a collection of meaningful words they know, even rare ones. They recognize friquish as not a real word. AI, however, does not group words this way. It uses every instance from its training data directly. The researchers argue that AI does not think as abstractly as humans. This explains why AI needs much more data to learn language.
This study blends linguistics and AI research. It helps us understand how AI generates language, paving the way for better AI systems. The findings appear in PNAS.