AI thinking is closer to human than it seems — study

AI models such as ChatGPT learn language not so much through abstract rules as through "memories" of the examples they have seen, according to researchers from the University of Oxford and the Allen Institute for AI. They described an experiment comparing the choices made by humans and the GPT-J open large language model (LLM) when forming nouns from fictional adjectives with -ness or -ity.

Tech Xplore writes about it.

How linguists tested AI's ability to generalise

The team created 200 fictional English words like cormasive and friquish and asked the model to choose the form cormasiveness or cormasivity. GPT-J's answers coincided with human answers where AI found analogies with real words: friquish — friquishness was similar to selfish — selfishness, while cormasive was similar to sensitive — sensitivity. Two classical cognitive models — rules and analogies — showed that LLM behavior is closer to the second approach.

In addition, it turned out that GPT-J "remembers" every example from the training texts: its predictions for almost 50,000 real adjectives accurately reflected the corpus statistics. The difference with humans is that we store generalised vocabulary forms in our memory and immediately see that friquish does not yet belong to the language, while the model works with all individual occurrences without combining them.

"Although LLMs can generate language in a very impressive manner, it turns out that they do not think as abstractly as humans do. This probably contributes to the fact that their training requires so much more language data than humans need to learn a language," Janet Pierrehumbert, the professor at Oxford University, explains.

Co-author Valentin Hofman from Ai2 adds that the results will help to create more efficient and understandable AI systems.

As a reminder, Microsoft employees are prohibited from installing and using the DeepSeek app on company devices. It was announced during the hearing in the U.S. Senate by the President and Vice Chairman of the Corporation Brad Smith. According to him, the ban is related to the fact that the app stores data in China, and the chatbot's answers may be influenced by "Chinese propaganda".

We also wrote that Apple's Senior Vice President of Internet Services, Eddy Cue, stated in court that by 2035, people would probably no longer need an iPhone. In his opinion, the development of AI can change the technological landscape and bring new market players to the fore.