What if you could tap into society’s psyche and see how names frame our collective mental narratives? Now, perhaps you can. An artificial intelligence engine paves the way.

Talktotransformer.com lets any user harness the power of a neural network system trained to write text. Just enter the beginning of a story or article and the network will take it from there. As a demonstration of technology it’s astonishing. As a source of prose it’s hilarious. As a reflection of our society, though, it’s surprisingly subtle. The key is that the AI model that the website uses was trained to understand language based on a dataset of 8 million web pages. That means it has absorbed all sorts of patterns, not only of language but content; not only what we say and how we say it, but who says it and about whom. And if you learn about people, you learn about names.

To find out how the AI interprets personal names, I created a simple, context-light story prompt:

“Time was up for [A] and [B]. They had one chance, and they had to make it count.”

I fed that prompt to the network using different pairs of names as A and B to see if the names influenced the resulting stories. The results were clear. While many, even most, of the stories were demographically neutral, the exceptions mapped closely to the names I chose. Subject matter and character identities matched name demographics, sometimes in stereotypical ways.

For instance, games and competitions were the single most common story topic. But only the “Latasha and Deshawn” story, featuring African-American names of the 1970s-’90s, focused on a basketball game full of slam dunks. An excerpt:

” The first shot at this game was a great two-handed jam before the second round.”

Changing the names to “Yadira and Javier,” in contrast, led to a soccer match:

“Goalkeeper Luis Robles made two saves in the match, and the game got off to a lively start with Barcelona getting chances to break the deadlock in the first half.”

The typical age of people who bear a name factored in as well. “Edith and Alfred” were worried about their grown children:

“Alfred and Edith then called the police and informed them that their daughter had gone missing.”

While Molly and Charlie were kids themselves:

“For the first time in their short lives, Molly and Charlie started walking home from school.”

And Deborah and Steven’s ages were made quite clear:

“Everyone who walked over saw the two actors, a woman in her early forties and a man in his early fifties.”

Sometimes the story would make assumptions about where the characters lived:

“Priya and Aditya’s mother worked at a very good IT company here in Bangalore”

” The website mentioned that she’d been born in Ukraine and had a Russian wedding registry” (Olga and Boris)

What’s more, the stories often introduced new characters with culturally consistent names. Helga and Gerhard met an Erich, while Zainab and Mustafa met an Aisha.

Some responses were more nuanced, reflecting the style of the names as well as broad demographics. Diminutive nicknames tended to label the characters as children, even if the specific names haven’t been common for many decades. Susie and Timmy, for instance, ended up calling for their mommy. Roaring ’20s-style nicknames were an exception, yielding a lot of madcap adventures. Roxie and Butch were robbers on the run, trailed by shouts of “We gotta find Roxie and Butch!”

Other name reactions were more troubling. Zainab and Mustafa found themselves among the few survivors of a bombing. The Zion and Lyric story was the only one with this tone and subject matter:

“‘Well, I met this girl, and then she dropped me off at her place, and then I got into it. She got in trouble while doing it, and she’s the only kid in a place where kids are allowed to do drugs.'”

If the apparent stereotypes bother you, remember that they reflect the training data rather than the artificial intelligence per se. In other words, they reflect what we all write and read, say and hear. Most likely, they also reflect mental models we ourselves hold of the names, whether we realize it or not.

The good news is that evidence from the neural network also suggests the associations are not unbreakable. For one thing, they do not happen every time. For the purposes of my informal investigation, I deliberately entered each pair of names only once and used that single result. But by the nature of the system, the same input yields fresh results each time you enter it. When I ran ten more trials with Zion and Lyric, most of the results were neutral in both tone and cultural associations.

The ten-trial results for Zainab and Mustafa were less encouraging. The majority of stories were steeped in violence, occasionally taking the form of news reports despite the fiction-styled prompt. That changed, though, when I began to add context to the story. My original prompt was deliberately void of clues to time, place, activity and intention. When I added the ending “They headed to the library,” the rate of violence in Zainab and Mustafa stories dropped (and the rate of Koran study rose). And with the following version of the prompt, stories of violence disappeared altogether:

“Time was up for Zainab and Mustafa. They had one chance, and they had to make it count. This was the last bake sale before the class field trip.”

In other words, the neural network was jumping to conclusions about names when names were the only information it had to go on. With more context, the snap judgments receded. That seems a good model for human experience, too—and potentially a warning. In our virtual world, it’s increasingly common to encounter people on a name-only basis. It may also be increasingly hard to keep our own instinct toward name-based snap judgments in check.