Meta has established a school to teach AI “emotional intelligence”. AI lacks emotional intelligence and training data needs to be designed to improve its social cognitive ability. When we feel that AI is not useful, it may be because it is not “understanding” people well enough. Taking the shopping scenario as an example, an AI with emotional intelligence can understand customer needs, while an AI without emotional intelligence simply replies with prices. For AI to understand people, it needs to have the ability of “theory of mind”, being able to understand “the meaning behind words”. The Meta research team developed the ExploreToM system to test AI’s understanding ability. Stories are divided into three groups according to complexity, and traps are added to increase the difficulty. Experiments show that under complex questions, models like GPT-4o have a low accuracy rate. The more actions there are and the more complex the situation is, the more difficult it is for AI to remember the updated cognitive states of characters. Increasing the number of participants has a relatively small impact, and different combinations of action types have a large difference in model performance. In complex scenarios involving intentional misinformation, the model’s accuracy rate is even lower. Researchers transformed ExploreToM into a tool for cultivating AI’s social cognitive ability and collected specially designed “practice questions” to train the Llama-3.1 8B model, with remarkable results and without affecting other abilities. The study found that the quality of training materials is more important than the quantity. The lack of emotional intelligence in AI is due to training data issues. In the future, we need to rethink the way training data is collected and increase scenarios with cognitive differences and information asymmetry. Humans still hold the psychological high ground of true empathy and emotional intelligence, which may be the next goal for AI’s self-learning.
Posted inAI News
