The Challenging Task Of Equipping AI With The Theory Of Mind – OpEd

Suppose you are sitting with your friend in a restaurant and after telling you they are thirsty they call out the waiter. In that case, you can readily infer, albeit they share no extra information, they are about to request a cup of water. The ability to understand and ascribe a mental state to others is called the theory of mind, which humans need it to navigate the world and reason. In recent years, researchers have started to test the possibility of enhancing AI systems with this capacity so as to improve their abilities in environments that require performing tasks with humans. 

Many years of research reveal that we rely on the theory of mind in many crucial aspects of our life such as language learning,  moral Judgment, perspective thinking, and  discerning intention during dynamic actions. To measure how humans utilize the theory of mind, many tests have been developed such as  false-belief tasks, unexpected content, and false photographs. It has been found that normal humans are usually expected to pass these tests but others with deficits and impairments fail to do so, for instance, those who fall into the autistic spectrum. Researchers still debating about the origin of the theory of mind and the role of evolution in developing it. 

Current existing AI frameworks such as deep learning and neural networks, despite having been proven successful in many areas, did not reach the technical advancement to understand the complexity of the theory of mind, these shortcomings have an obvious impact on its performance in various applications, enhancing these frameworks with the theory of mind will a great impact on how they operate. An ideal example is self-driving cars that need to discern the goal of other cars and pedestrians to decide in their direction, another example is the robots that while interacting with humans to provide any service should understand their goals and intentions.  

Researchers appear to be divided in acknowledging the achievements made thus far in adding the theory of mind to AI. for instance, most recently, researchers tested many large language models’ (LLMs) abilities to achieve the Theory of Mind, these models tested using the classic false-belief tasks, and their results showed that models published before 2022 showed no ability to solve Theory of Minds tasks, but a January 2022 version of ChatGPT system solved 70% of these tasks, this performance is compared to a seven-year-old child, another version of the system even reach 90% which is compared to a nine-year-old child.

According to the study, this ability emerged Spontaneously as a byproduct and was not intentionally designed and programmed. But this claim has been questioned by other researchers who showed that testing the same system with small variations will turn the result on its head and lead to machine failure, another 2022 that tested GPT-3 against a couple of known tests found that the system falls by more than fifty percent to reach human accuracy. 

In February 2018, researchers at DeepMind published a paper that claims to design a neural network with the theory of mind capacity in a simple grid world environment and that their system passed the classic theory of mind test, but other researchers showed the system while producing great results, did not exhibit innate theory of mind and that they system was trained to reach this level. 

But the theory of mind should be mutual between humans and machines and not only one-sided as some AI scholars have recently argued, that’s as so much we need the machines to understand and infer what humans agents think about their goals, humans are required to arrive at the correct interpretation of the machine end goals and intentions. 

This ongoing debate between researchers about the ideal way to incorporate the theory of mind into AI and how to measure the success reveals that we are still in very early stages and It will take time for machines to genuinely exhibit the theory of mind, but these attempts here and there foretell we are heading in the right direction. 

We are also obviously in dire need of more transparency in accessing these models so that the public can assess any purported capacities and manage our expectations from them.