We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
cover of episode Using AI to Understand the Thoughts of the Dead

Using AI to Understand the Thoughts of the Dead

2024/11/22
logo of podcast Science Quickly

Science Quickly

AI Deep Dive AI Insights AI Chapters Transcript
People
M
Michael Varnum
R
Rachel Feltman
Topics
Rachel Feltman: 探讨了利用AI探索过去,甚至与已灭绝文明的成员‘对话’的可能性,并采访了该领域专家的观点。 Michael Varnum: 指出了研究人员在了解古代文明成员心理和行为方面面临的挑战,即无法直接获取数据。传统方法只能通过间接途径(如档案数据和文化产品)来推断过去人们的价值观和情感,这不够直接。文章提出了一种新方法:利用像ChatGPT这样的程序模拟现代参与者,并令人惊讶地复制了行为科学中许多经典效应,这启发了研究人员尝试用古代文本训练模型来模拟古代人的行为。目前,研究人员只是尝试用历史文本训练大型语言模型,并测试其对特定知识的理解,但尚未将这些模型用于现代实验或调查。 Michael Varnum进一步讨论了历史文本数据中存在的偏差问题,因为在大部分人类历史上,识字率很低,且只有少数人的文字记录流传至今,这些记录往往偏向精英阶层。他提出了几种解决方法,例如根据现代人口的社会阶层影响进行模型微调,或对模型的回应进行加权处理,以及将历史大型语言模型的结果与其他历史记录和分析进行比较,以验证其可靠性。他还提到,当前大型语言模型更贴近西方和英语国家人口的心理特征,这与训练数据中这些社会群体被过度代表有关,但也表明使用不同语料库可以捕捉到不同文化的特征。 Michael Varnum还展望了这项技术的未来应用,例如帮助研究人员更深入地了解人类心理的普遍性和进化基础,通过研究古代社会来验证某些心理现象的普遍性,例如男女在性策略上的差异。他认为,AI不仅可以模拟参与者或编码数据,还可以生成新的研究假设,未来AI可能在社会心理学研究中扮演更重要的角色,甚至可能生成研究想法,这可能会改变研究人员的工作方式。 Rachel Feltman: 提出了关于历史文本数据偏差的问题,以及如何利用AI技术来解决这些偏差,并探讨了这项技术在研究弱势群体历史方面的应用潜力。

Deep Dive

Key Insights

Why is there a need to use AI to simulate ancient participants in social psychology studies?

There is a need because we lack direct ways to study the mentality and behavior of people from ancient civilizations. Traditional methods, such as archival data and cultural products, are indirect and limited. AI can potentially provide more direct insights by simulating ancient participants based on historical texts.

What kind of experiments have already been conducted using AI to simulate modern participants?

Experiments have replicated 70 different large-scale survey experiments using simulated participants from ChatGPT, with results correlating at about 0.9 with real human data. This suggests that AI models can capture significant aspects of human psychology.

Why is using historical texts to train AI models challenging?

Using historical texts is challenging because they are biased towards the perspectives of literate, elite, and educated individuals from the past. This can lead to a skewed understanding of ancient populations. Researchers need to account for these biases by using additional historical records and weighting responses.

What does the acronym WEIRD stand for in the context of behavioral science?

WEIRD stands for Western, Educated, Industrialized, Rich, and Democratic. It refers to the fact that the majority of participants in behavioral science come from these types of societies, which are not representative of the global population.

What are some potential dream use cases for AI in studying ancient psychology?

Potential use cases include testing the universality of certain psychological traits by extending the temporal window back to ancient societies. For example, researchers could examine differences in preferred sexual strategies between men and women in societies that lived hundreds or thousands of years ago.

How might AI be used beyond simulating participants in social psychology research?

AI could be used to generate new hypotheses for social psychology research. A recent study found that GPT-4 generated hypotheses that were considered more compelling and probably true by social psychologists, suggesting that AI could become a valuable tool for generating research ideas.

Shownotes Transcript

Writings and records are how we understand long-gone civilizations without being able to interact with ancient peoples. A recent opinion paper suggested we could feed chatbots writings from the past to simulate ancient participants for social psychology studies. Similar survey experiments with modern participant data closely matched the outcomes of the real people they were based on. We speak with the opinion paper’s co-author Michael Varnum, an associate professor at Arizona State University, about what the limits of this spooky proposal are and what the ghosts of cultures past could teach us today. 

Recommended reading:

“Large Language Models Based on Historical Text Could Offer Informative Tools for Behavioral Science,” by Michael E. W. Varnum et al., in Proceedings of the National Academy of Sciences USA, Vol. 121, No. 42, Article No.  e2407639121; October 9, 2024

https://www.pnas.org/doi/10.1073/pnas.2407639121

Inside the AI Competition That Decoded an Ancient Herculaneum Scroll

https://www.scientificamerican.com/article/inside-the-ai-competition-that-decoded-an-ancient-scroll-and-changed/

E-mail us at [email protected]) if you have any questions, comments or ideas for stories we should cover!

Discover something new every day: subscribe) to Scientific American and sign up) for Today in Science, our daily newsletter. 

Science Quickly is produced by Rachel Feltman, Fonda Mwangi, Kelso Harper, Madison Goldberg and Jeff DelViscio. This episode was hosted by Rachel Feltman. Our show is edited by Jeff DelViscio with fact-checking by Shayna Posses and Aaron Shattuck. 

The theme music was composed by Dominic Smith.

Learn more about your ad choices. Visit megaphone.fm/adchoices)