An interesting thing about contemporary artificial intelligence models, specifically large language models (LLMs): They can only output text based on what’s in their training dataset. Models, including ChatGPT and Claude, are “trained” on large databases of text. The models, when asked a quest ion, statistically create a response by calculating, one word at a time, what the most likely next word should be. A consequence of this is that LLMs can’t output text about scientific breakthroughs that have yet to happen, because there’s no existing literature about those breakthroughs. The best an AI could do is repeat predictions written by researchers, or synthesize those predictions. Adam Mastroianni, writing in his newsletter Experimental History, put this elegantly: “If you booted up a super-smart AI in ancient Greece, fed it all human knowledge, and asked it how to land on the moon, it would respond, ‘You can’t land on the moon. The moon is a god floating in the sky.'”  It’s an interesting thought experiment. What if you intentionally limited the training data? Could you create an AI system that responds as though it’s from a period in the past?  What could that reveal about the psychology or everyday experiences of the people from that era? That’s exactly what Hayk Grigorian, a student at Muhlenberg College in Allentown, Pennsylvania, had in mind when he created TimeCapsuleLLM. This experimental AI system was trained entirely on texts from 19th century London. The current release is based on 90 gigabytes of text files originally published in the city of London between 1800 and 1875.  This is, to be clear, very much a hobby project. The sample-generated text on GitHub isn’t consistently coherent, though Ars Technica did report that it has correctly surfaced names and events from the 1800s. When prompted to continue the sentence “It was the year of our Lord 1834,” the model recounted a protest: “the streets of London were filled with protest and petition,” going on to mention the policies of Lord Palmerston, who was the foreign secretary at the time.  It’s an interesting experiment, but could such a thing actually be useful? Potentially.An opinion piece published by the Proceedings of the National Academy of Sciences of the United States of America (PNAS) by collaborators including Michael E. W. Varnum, a professor of psychology from the Department of Psychology at Arizona State University, is an interesting read. It proposes that models like this could be a way to study psychology outside a modern context. The paper refers to such AI models as Historical Large Language Models, or HLLMs for short, and states that psychology researchers could use them to study the thinking of people in past civilizations.  “In principle, responses from these faux individuals can reflect the psychology of past societies, allowing for a more robust and interdisciplinary science of human nature,” the paper says. “Researchers might, for example, compare the cooperative tendencies of Vikings, ancient Romans, and early modern Japanese in economic games. Or they could explore attitudes about gender roles that were typical among ancient Persians or medieval Europeans.” It’s an interesting idea, though the paper does acknowledge this could be tricky.  “All LLMs are a product of their training corpora, and HLLMs face challenges in terms of sampling, given that surviving historical texts are likely not representative samples of people who lived in a particular period,” the paper admits, stating that historical texts tend to be written by elites, not everyday people. “As a result, it could be hard to generalize from these models.”  And there are other things to keep in mind. Research from Ghent University in Belgium shows that the ideology of the people who work on an LLM shows up in the text those models generate. There’s every reason to suspect the same problem will apply to LLMs designed to reflect past cultures.  So there are difficulties. Only time will tell if such models end up being used in psychological research, or remain the domain of hobbyists.  The post This AI thinks it’s the 1800s appeared first on Popular Science. ...read more read less