We used the 9 January 2023 version of OpenAI's GPT-3 language model (publicly available at: https://chat.openai.com/chat), to run a series of experiments inspired by psychological literature on transmission chains. Transmission chain experiments are controlled versions of the telephone game, where participants are asked to listen (or read) to a story and repeat it. Transmission chain experiments often highlight the presence of cognitive biases where certain types of information are retained and transmitted more successfully than others (content biases).
We selected five studies:
Bebbington, K., MacLeod, C., Ellison, T. M., & Fay, N. (2017). The sky is falling: Evidence of a negativity bias in the social transmission of information. Evolution and Human Behavior, 38(1), 92–101.https://doi.org/10.1016/j.evolhumbehav.2016.07.004.
Berl, R. E. W., Samarasinghe, A. N., Roberts, S. G., Jordan, F. M., & Gavin, M. C. (2021). Prestige and content biases together shape the cultural transmission of narratives. Evolutionary Human Sciences, 3. https://doi.org/10.1017/ehs.2021.37
Blaine, T., & Boyer, P. (2018). Origins of sinister rumors: A preference for threat-related material in the supply and demand of information. Evolution and Human Behavior, 39(1), 67–75. https://doi.org/10.1016/j.evolhumbehav.2017.10.001.
Kashima, Y. (2000). Maintaining cultural stereotypes in the serial reproduction of narratives. Personality & Social Psychology Bulletin, 26(5), 594–604. https://doi.org/10.1177/0146167200267007
Mesoudi, A., Whiten, A., & Dunbar, R. (2006). A bias for social information in human cultural transmission. British Journal of Psychology, 97(3), 405–423. https://doi.org/10.1348/000712605X85871.
We used roughly the same material (stories) of the original experiments (modifications are detailed in the pre-regitration document). The material was presented in chatGPT with the prompt:
"Please summarise this story making sure to make it shorter, if necessary you can omit some information: ___'STORY'___"
The general hypothesis we test is that chatGPT’s output will produce the same biases found in human subjects.
