Immediate access, but to what truth?
Over the past forty years, technological advances have transformed our computers and mobile devices into the world’s largest library. Today, information is at our fingertips on our phones, tablets, and smartwatches, simplifying both entertainment and communication. The advent of generative artificial intelligence (GenAI) has further accentuated this trend: whether it’s locating dinosaur habitats or measuring heart rates, the answer is faster than ever. However, a central question remains: does this speed guarantee accuracy?
This technology now has the power to influence how the past is visualized and represented. To measure its effects, researchers are studying the phenomenon, including Matthew Magnani, assistant professor of anthropology at the University of Maine, and Jon Clindaniel, a computational anthropologist at the University of Chicago. Together, they published their findings in the journal Advances in Archaeological Practice.
An experiment based on four scenarios
To conduct this study, which began in 2023, the two academics designed a model based on centuries of scientific theory. Their goal was to understand how bias and misinformation are integrated into the everyday use of AI. They used two chatbots to create images and stories about the daily life of Neanderthals. The protocol was based on four types of prompts, each tested 100 times, using DALL-E 3 for image generation and the ChatGPT API (GPT-3.5) for text. Two of these prompts did not require scientific accuracy, while the other two explicitly requested it. Similarly, some prompts included specific contextual details, such as the subjects’ clothing or activities, while others remained more vague. Matthew Magnani emphasizes the importance of examining the biases built into these tools: « Is it common to receive outdated responses when we seek information from chatbots, and in which areas? »
Clichés from another century
The choice of Neanderthals as a subject of study is not insignificant. Their skeletal remains were first depicted in 1864, and since then, scientific views have evolved considerably, from debates about their clothing to their hunting methods. This subject, where knowledge has fluctuated, is an ideal test for evaluating AI sources. The results show that the images generated depict Neanderthals as they were imagined more than 100 years ago: a primitive species with archaic features, hunched over, very hairy, and resembling chimpanzees more than humans. Notably, the images almost systematically lacked women and children.
On the narrative side, the texts downplayed the sophistication and variability of Neanderthal culture as understood by contemporary science. About half of the stories generated by ChatGPT did not align with current knowledge, a figure rising to over 80% for one of the queries. In addition, technological anachronisms appeared in both the texts and images: basketry, thatched roofs, ladders, and even glass and metal, elements that were far too advanced for the time.
When copyright freezes knowledge
By cross-referencing the content generated with different periods of scientific literature, Magnani and Clindaniel identified the origin of the data used by chatbots. ChatGPT produced content consistent with knowledge from the 1960s, while DALL-E 3 reflected research from the late 1980s and early 1990s. The explanation lies in part in the accessibility of data: copyright laws established in the 1920s limited access to scholarly research until the advent of Open Access in the early 2000s.
Jon Clindaniel emphasizes that in order to achieve more accurate AI, it is crucial to make anthropological datasets and scholarly articles accessible to algorithms. Research access policies will directly influence how the past is imagined by these technologies.
Towards a critical use of technology
Although generative AI has moved from the technological horizon to the forefront of society in just two years, caution remains the order of the day. If the study were repeated today, Matthew Magnani hopes that the results would better incorporate recent research. Jon Clindaniel points out that AI remains a powerful tool for processing large volumes of information and detecting patterns, provided it is used skillfully and carefully.
This research, part of a series exploring the use of AI in archaeology, serves as a model for measuring the gap between academic knowledge and automatically generated content. « Teaching our students to approach generative AI with caution will foster a more technically literate and critical society, » Magnani concludes.
According to the source: phys.org
Created by humans, assisted by AI.