ChatGPT outperforms UCLA students in SAT test

Taylor Webb, a metacognition researcher with the UCLA (University of California, Los Angeles) led a research to assess ChatGPT’s reasoning ability. The task provided to ChatGPT was to predict the next image in a convoluted shape array by providing its text equivalent. ChatGPT had an accuracy of 80%. An average of 40 ULCA students was lesser in comparison. Certain analogy questions were also feeded to the algorithm. Here too the AI-powered language model outperformed the students of UCLA. 

However the scenario is not so straightforward as well. It did not perform well in a task wherein it had to match a poem with a prose conveying similar meaning. Social interaction, mathematical reasoning and understanding of physical space were regions where-in ChatGPT was still lacking. 

Taylor Webb clarified that there is no basis to assert that ChatGPT has now achieved the capacity to match human intelligence. Pattern recognition and inference ability are the strengths of the software. There is however ambiguity with respect to how its reasoning process functions. It is not clear whether they mirror human reasoning or have developed an alternate means of performing the task. The difference could also stem from variance in training methods.