GPT-3 Displays Lower Social Intelligence Than Human Children. Chen, Y. National Collegiate Research Conference, Harvard University, January, 2021.
abstract   bibtex   
Generative Pre-trained Transformer 3 (GPT-3) [1] is a neural language model with broad language capabilities. It achieves state-of-theart or human-level performance in translation, question answering, and creative writing tasks. It is deployed in commercial products such as companion chatbots [2], however, the existing benchmarks do not specifically test for its level of social intelligence. In this project, we explore GPT-3’s social intelligence by adapting batteries and human baselines from psychology research. We found that: 1) In non-social control questions, GPT-3 performed at near-human level. 2) In social questions, it scored lower than both children and adults. 3) GPT-3’s performance did not significantly increase even when similar example’s solutions are provided.
@article{chen_gpt-3_2021,
	title = {{GPT}-3 {Displays} {Lower} {Social} {Intelligence} {Than} {Human} {Children}},
	abstract = {Generative Pre-trained Transformer 3 (GPT-3) [1] is a neural language model with broad language capabilities. It achieves state-of-theart or human-level performance in translation, question answering, and creative writing tasks. It is deployed in commercial products such as
companion chatbots [2], however, the existing benchmarks do not specifically test for its level of social intelligence. In this project, we explore
GPT-3’s social intelligence by adapting batteries and human baselines from psychology research. We found that: 1) In non-social control questions,
GPT-3 performed at near-human level. 2) In social questions, it scored lower than both children and adults. 3) GPT-3’s performance did not significantly increase even when similar example’s solutions are provided.},
	journal = {National Collegiate Research Conference, Harvard University},
	author = {Chen, Yifu},
	month = jan,
	year = {2021},
}

Downloads: 0