OpenAI Sora's “historical footage” of the 1849 Gold Rush: promise and problems for genealogists

We have two known “Forty-Niners” in our family tree. The knowledge of these two men is based entirely on an account written in longhand (reminder: stories matter for future genealogists!) by my Great-Aunt Genevieve nearly 50 years ago. There was Edward Cowan, a '49’er who met with great success:

“There was a pin made of gold that he mined and sent it to my great grandmother … My mother had a picture of him with a beard which he had grown which touched the floor! Some of his descendants are fairly wealthy and prominent and quite proud of the fact that they are fifth generation Californians.”

Then there was Michael Scanlon, “who went to California in the Gold Rush and was never heard of again.”

I always wondered about their lives during the California Gold Rush. For Edward, was it a slow grind searching the seams and streams for tiny nuggets of gold, or did he strike it rich in the mother lode? For Michael, did he die unexpectedly? Or was he too reluctant to admit failure to his relatives? Or did something else happen? For both men, what was it like crossing the continent on horseback or foot, and living in a frontier mining camp?

We can’t answer all of those questions, but now have a glimpse into what a frontier mining town looked like:

This is not a film set, a historical re-enactment, or traditional CGI (computer generated imagery) used in Hollywood blockbusters. It’s a tool called Sora, based on a type of artificial intelligence known as generative AI, which can create convincing media content such as text, photos, and video from simple text prompts. In this case, someone typed “Historical footage of California during the gold rush” and Sora created the video.

The creator of this technology, OpenAI, says Sora can accurately “interpret prompts and generate compelling characters that express vibrant emotions.” But OpenAI also says “it may struggle with accurately simulating the physics of a complex scene” and “may struggle with precise descriptions of events that take place over time.” If you take a look at the video, you’ll note some inconsistencies (pack horses walking without humans through town) or missing elements (no commercial signs).

This points to a fundamental limitation of generative AI: the quality of the output is only as good as the millions of images and data points that went into the model. A similar prompt for activity around Potosí in the Bolivian Andes when vast silver deposits were discovered there in the 1540s might show wildly inaccurate imagery, or fail to generate anything at all. That’s because there isn’t enough modeling data for Sora to depict the people, buildings, animals, and the local landscape from nearly 500 years ago.

We will almost certainly begin to see AI-generated historical footage in the months and years to come. It’s exciting to view scenes that our ancestors may have experienced hundreds of years ago, but we also have to be vigilant for seemingly authentic historical recreations based on flawed data or assumptions.