AI could be capable of 'learning by thinking', say researchers

 ### AI and the Concept of 'Learning by Thinking':

Recent research suggests that artificial intelligence (AI) might be capable of a cognitive process called 'learning by thinking,' a concept central to great discoveries. Traditionally, cognitive scientists have studied how people learn by observing the external world, a process known as observational learning. However, there is another form of learning—often overlooked—called 'learning by thinking.' This occurs when one gains knowledge not through external input, but through internal cognitive processes such as thought experiments and self-explanations.

According to Tania Lombrozo, a professor of psychology at Princeton University and author of a review published in *Trends in Cognitive Sciences*, 'learning by thinking' is a powerful tool that humans have used to advance knowledge. For instance, thought experiments involve exploring theoretical principles by mentally working through their consequences. Similarly, learning through self-explanations occurs when one makes sense of new information by connecting it to what they already know.

Historical figures like Albert Einstein and Galileo Galilei used thought experiments to develop groundbreaking ideas like the theory of relativity and insights about gravity. Lombrozo’s review, however, indicates that this type of cognitive process might not be limited to humans; AI, particularly advanced language models, could also engage in a form of 'learning by thinking.'

### AI's Potential for Self-Correction and Thought Experiments:

Lombrozo notes that recent developments in AI, particularly in large language models, show signs of this cognitive ability. When asked to elaborate on a complex topic, AI can sometimes refine or correct its initial responses based on the explanations it generates. This is similar to how humans refine their understanding through self-explanation.

In the gaming industry, simulation engines approximate real-world outcomes, and AI models can learn by using the outputs of these simulations as inputs for further learning. Additionally, prompting AI to draw analogies or engage in step-by-step reasoning can improve its ability to answer complex questions more accurately than through simple queries.

In some cases, AI models, like ChatGPT, even correct themselves without being explicitly told, which mirrors how humans learn by thinking. Lombrozo emphasizes that this self-correction is an emerging demonstration of AI's capacity to engage in thought-based learning.

This suggests that AI could one day harness a key aspect of human cognitive ability, advancing its problem-solving skills by thinking through complex scenarios without direct input from the outside world.

### 'Learning by Thinking' in Humans and AI: A Comparative Perspective:

In humans, 'learning by thinking' occurs through processes like explanation, simulation, analogy, and reasoning, according to Tania Lombrozo’s research. Lombrozo, a professor of psychology at Princeton University, highlights how this form of learning allows people to acquire knowledge without direct external input, unlike traditional observational learning.

One clear example of 'learning by thinking' in humans involves explaining how a microwave works to a child. During the explanation, one may discover gaps in their understanding, leading to deeper learning. Another instance occurs when rearranging furniture in a room. Instead of physically moving items right away, many people mentally simulate different layouts to determine the best arrangement before taking action.

Lombrozo draws comparisons between this human cognitive ability and AI's potential for similar processes. She notes that these comparisons raise important questions: Why do both natural (human) and artificial (AI) minds exhibit these characteristics? What purpose does 'learning by thinking' serve, and why is it valuable?

Lombrozo argues that 'learning by thinking' is a form of “on-demand learning.” Humans may learn new information without knowing its immediate relevance, storing it for future use when the context demands it. This ability allows for flexible application of knowledge in different scenarios.

However, the review also poses a critical question: Are AI systems truly 'thinking,' or are they simply mimicking the outcomes of such cognitive processes? While AI can demonstrate behaviors that resemble human 'learning by thinking,' the deeper question remains whether this is genuine reasoning or a sophisticated replication of the outputs associated with human thought.

This distinction could have profound implications for understanding the nature of AI cognition and its potential for autonomous learning.