EasyGenie and AI: paintings vs AI

Thanks for the responses to part I of our series about how EasyGenie uses artificial intelligence. Tami noted that while AI is a powerful tool, it is also somewhat scary. Another reader, Trent, asked:

“You said you wouldn’t use AI for showing real people or historical events, but how is that different than what painters have done since time immemorial?”

Trent has a point. Consider the iconic 1851 painting by German-born painter Emanuel Leutze shown above, depicting George Washington crossing the Delaware River.

Then consider this AI-generated painting created in Canva:

Ignoring the awkwardness of the AI-generated image (quality will surely improve in the coming years) we are presented with two exaggerated images of a real historical event. Neither were contemporaneous with Washington’s historic 1776 crossing. Leutze put paint to canvas some 75 years later, while Canva whipped up the AI-generated image 249 years later.

It’s uncertain how accurate the imagery is. Historians can't say for certain how many people were on Washington’s boat, whether anyone was carrying a flag, or if George himself took to the oars. Were the people on the boat standing or sitting? What color was his jacket? We simply don’t know.

The AI actually produced multiple versions that show Washington wearing different colors, or not holding the oar:

When it comes to the EasyGenie newsletter or blog, we wouldn’t hesitate to use Leutze’s work, or any other old painting to illustrate real historical figures or historical events. We do it all the time.

But an AI image? We would balk. Trent’s observation forced us to think about why we would treat a painting differently than an AI image of the same person or event. Here are three reasons:

1. The painting was created by a human, which provides additional context around the event as well as potential biases of the creator. 
2. It’s authentically old, and therefore closer to period styles and appearances and details. 
3. It doesn’t pretend to be historically accurate.

George Washington Crossing the Delaware was a vessel of the artists’ patriotism, and reflects the zeitgeist of the early 1850s. It was the 75th anniversary of the country’s founding, and many Americans were searching for a hero to rally around as anxiety over the state of the union began to build prior to the Civil War. 

The AI images are artifacts of the current age. They look new, with high-contrast colors like something from a video game. The details don’t look right, such as the rows of identical soldiers behind Washington. It’s jarring.

Another image from Canva is downright bizarre, showing two Washington-like figures sitting side by side, staring intently at something:

Concerningly, the AI sometimes pretends to be historically accurate. Here’s the caption that Canva displayed next to the second AI image shown above, with Washington wearing a red jacket:

“A historically accurate depiction of George Washington crossing the Delaware River in a small, wooden rowboat. … The overall aesthetic is dramatic and realistic, painted in a style reminiscent of 19th-century historical paintings, focusing on a balanced composition with emphasis on detail and atmospheric perspective.”

“Historically accurate?” No one on the boat is even holding oars!

What humans see in historical AI

What really worries us is the tendency for people to believe what they’re told or shown by AI, with no critical evaluation. Check out this image, purporting to show a real historical event. It showed up in my social media feed:

Dirty Old Boston has been sharing authentic photos of Boston’s past for many years. Yet despite the account owners’ familiarity with local history, they shared a photo that was obviously created by AI, merely citing “photographer unknown.” Later, they acknowledged it was an AI photo found on the Internet, and apparently deleted it soon after.

But the damage was done: Fake was treated as fact, and more than 1,000 people liked what they saw.

Let’s be clear: There are no photos showing the moment the molasses tank burst in Boston’s North End in January 1919. There is no “historic footage” of the 1849 California Gold Rush, which took place decades before the invention of motion pictures. There was no child jumping in front of Nicole and her friends in the 1990s:

So why is there so much fake AI history out there? Part of the problem relates to how easy it is to generate AI images and text. It literally took less than a minute for Canva’s AI generator to pump out four images and accompanying captions, based on the simple prompt, “george washington crossing the delaware in a boat.” 

Other factors come into play. There is a known problem with generative AI “hallucinating,” in essence making stuff up based on the thousands of samples the model has ingested. The result is often weird, like the two Washingtons staring at something on the water, or a row of mid-1800s commercial buildings from the Old West that don’t have a single shopkeeper’s or hotelier's sign:

In other cases, humans are deliberately prompting the AI to make a fake history image (like the “photo” of a disaster) and then trying to pass them off as real for clicks and likes. Or, an app developer decides to spice up old genealogy portraits, which is what happened when MyHeritage turned Nicole’s 1990s group photo into a “live memory.”

Bottom line: At the moment, we don’t trust AI imagery beyond a very limited set of use cases. We can live with a generic AI-generated image of an overflowing box of old documents, similar to something one might find in a stock photography library. But Washington crossing the Delaware, or the 1849 Gold Rush, or the Great Molasses Flood of 1919? We’ll stick with an old painting, illustration, or original photograph, no matter how convincing an AI depiction may appear to be.