AI has learned to recognize data like a human, — research

Chinese researchers claim that modern AI models are already capable of recognizing and sorting different categories of information just like humans.

According to scientists from the According to researchers from the Chinese Academy of Sciences and the South China University of Technology in Guangzhou, some AI models copy key functions of the human brain when performing information sorting tasks. The researchers sought to find out whether they could large language models, such as ChatGPT or Gemini, to understand and process information about various objects in a human-like manner. 

To do this, the researchers tasked ChatGPT-3.5 and Gemini Pro Vision AI models with removing unnecessary text and visual objects from a series of textual and visual objects. For this purpose, 4.7 million responses were collected regarding 1,854 different objects, including dogs, chairs, apples, cars, etc.

The researchers found that 66 of the models created by AI were based on the way people do things. These models included quite complex attributes, such as: texture, emotional significance, and accessibility for children.

The researchers found that multimodal models capable of combining text and images further mimic human thinking. In addition, the team found that brain scan data revealed a match between how AI and human brains react to objects.

According to the Chinese scientists, the results of their research show that AIs are indeed capable of close to human understanding, not just imitating responses. Future AI models could gain more intuitive, human-compatible reasoning, which is essential for robotics, education, and human-human collaboration.

However large language models, they do not understand objects in the same way as humans do on an emotional or empirical level. AI models work, by recognizing patterns in text or visual content, but unlike humans, this is not based on life experience and sensorimotor interaction. Large language models are more like a mirror that reflects millions of sources of textual and visual information to the user based on the patterns they have learned.

New training method will make AI less deceptive

The results of the study were published in the journal Nature

Source: Interesting Engineering