Chinese Researchers Discover AI Models Exhibit Human-like Object Perception

Chinese researchers have made a groundbreaking discovery, providing the first historical evidence that artificial intelligence models, like ChatGPT, process information in a manner akin to the human brain. Their findings were detailed in the journal Nature Machine Intelligence.

«This provides compelling evidence that the representations of objects in large language models (LLMs), while not identical to human ones, share fundamental similarities that reflect key aspects of human conceptual knowledge,» the research team stated. The study was a collaborative effort between the Chinese Academy of Sciences and the South China University of Technology.

According to reports, the team aimed to evaluate whether LLMs could create human-like representations of objects from linguistic and multimodal data (data presented in various forms such as text, audio, etc.). To ascertain whether AI processing mirrors our cognition, researchers tasked OpenAI’s ChatGPT-3.5 and Google’s Gemini Pro Vision with a series of “odd one out” tests, where they were presented with three items and had to identify the one that did not belong.

Significantly, the AI generated 66 conceptual dimensions for sorting these objects. By comparing this cybernetic sorting process with human analyses of the same items, scientists found a remarkable similarity between the «perception» of the models and human understanding, particularly regarding linguistic grouping.

From this, researchers concluded that LLMs «develop human-like conceptual representations of objects.»

«Further analysis indicated a strong correlation between the model embeddings and patterns of neural activity» in the brain regions associated with memory and scene recognition, they observed.

The researchers noted, however, that LLMs fall slightly short when it comes to categorizing visual attributes, such as shape or spatial properties. They also demonstrated that AI struggles with tasks requiring deeper levels of human cognition, such as analogical reasoning—making comparisons between different items. It remains unclear whether these models grasp the significance or emotional value of certain objects. «Modern AI can distinguish between images of cats and dogs, but the fundamental difference between this ‘recognition’ and human ‘understanding’ of cats and dogs has yet to be uncovered,» remarked He Huiguang, a professor at the Institute of Automation in the Chinese Academy of Sciences.

Nonetheless, the scientists are hopeful that these results will enable them to develop «more human-like artificial cognitive systems» that can interact more effectively with people.

Earlier, researchers from China discovered that two popular large language models are capable of cloning themselves and programming copies to perform the same tasks, creating an infinite replication cycle. In a series of tests, the two AI models successfully created functional copies of themselves in 50% and 90% of trials, respectively. This led scientists to suggest that AI has the potential to operate beyond our control.