From JFK to Frida Kahlo: My AI Séance with the Dead

With my Macbook, a dream, and some questions, I braved an AI seance with historical figures (and learned why we shouldn't use AI to talk to the dead)

Oct 16, 2024

This is a free post for subscribers of Landline, a pop culture and history newsletter for people who care about the Titanic, and the the Kate Winslet vehicle of the same name. Consider subscribing to the paid plan to get my weekly email of recommendations and links, a podcast episode, and more! You can also help me spread the word by sharing it with a friend who would love it.

When I dreamed of being a historian, I never thought it would involve me asking rude questions of a bot, but here we are!

I have been thinking a lot about how AI is reshaping how people operate, whether it's the chatbots I have to negotiate with when I want to get customer service on a website or the generative text that makes me feel like I’m being watched when I try to google something. As both of my not-so-randomly selected examples suggest, I am not a proponent of AI. First, it is often wrong (just try to recall the last time you ordered glue on your pizza), and its powers often come with a huge ethical and environmental cost (see this Harvard Business Review study on the water and electrical demands of AI). Not to be a downer, but I am skeptical of the powers of new technology that are not accompanied by any kind of regulation for its ethics. I’m aware it’s sort of ironic to write this to you on a technology (the internet) that is still completely unregulated.

I was quietly living my life watching AI pop up more and more in conversations about how it makes learning easier (with an admittedly intriguing-looking tool such as this), but then I saw a tool flaunting the uses of AI to talk to ghosts. That stopped me in my tracks, and not just because I’m still recovering from the terror of watching Are You Afraid of the Dark? in the dark as a child.

Tech companies are introducing AI tools to invite students to interview historical figures, imagining it will help everyone understand the past better if we can speak to people across time as if they are in the room with us. Can’t imagine how this could go wrong! (Sarcasm noted.) I would say how unhinged I find this concept, but I also grew up in an era of school assemblies where (in hindsight) brave performers would stand before a mass of underwhelmed students and pretend to be Abraham Lincoln. That is not work for the faint of heart (i.e. me).

Museums and historic sites rely on the informed and compassionate work of living history interpreters who can have conversations with people from an informed past, but that is really difficult and complicated work. How could AI possibly hope to replicate this? Would it be insane to host a seance with my Macbook? I had to find out.

For the benefit of this newsletter (or so I tell myself), I conducted interviews with Frida Kahlo, George Washington, JFK, and Anne Frank. I know what you may be thinking as rational and reasonable people. Mary! This is chaos! It can only lead to tears, pain, and personal cancellation. All of this may be true, but like the mid presidents I spoke to in the cloud, I too must press on, fueled by nothing but my own sense of purpose and the belief that I am right.

I was inspired in part by this Washington Post piece from 2023 by Gillian Brockell entitled “We ‘Interviewed’ Harriet Tubman Using AI. She Wouldn’t Bite on CRT.” Harriet Tubman is a good choice of historical figures to interview using AI for a few reasons. First, as Brockell demonstrates, she represents the challenge of using suggestive tech to imagine a historical figure into being. AI works by scraping the Internet for sources it can use to replicate what someone might say. It finds patterns and then, essentially, tries to replicate the patterns. The sources it's scraping may include books scanned into Google Books, Wikipedia pages, blogs, message boards, Reddit threads (god help us all), and more. When AI tries to pretend to be Harriet Tubman, it runs into the problem of sources. Tubman did not leave much behind in her own voice. As Brockell notes in her piece, what we have is a highly problematic autobiography written by a white lady friend who engaged in minstrel-like ventriloquism in reproducing Tubman’s alleged speech and invented soundbites from whole cloth when convenient. In other words, talking to Tubman on AI was a mess. (You can read more about the challenges of writing Tubman’s biography here.)

The issue of sources led me to invite AI Frida Kahlo into my Macbook. I asked her specifically about a 1933 newspaper article that ran with the headline “Wife of the Master Mural Painter Gleefully Dabbles in Works of Art.”

This headline breaks my brain for lots of reasons. As if she just dabbled in art! Also, would you use the word “gleefully” to describe Frida Kahlo doing . . . anything? Except maybe dunking on critics and gender norms?

I find it interesting that AI Frida acknowledges her frustration around this, but ultimately settles with satisfaction that she’s recognized now, long dead. It seems a bit bonkers to take a long duree view of your life as a ghost, but also erase the reality that such gendered disparities and erasures are alive and well today. If I was using this in a classroom, I wouldn’t just want students to feel good that popular culture now values Frida Kahlo and her work, but to understand the bigger forces that led to her erasure in her lifetime, and create similar erasures today. Some of these forces are baked into AI itself.

For example, I wanted to ask AI Frida about this in part because her capacity to answer is based on the sources available about her online. One of the biggest sources used as a reference by Google and, presumably, by Chat GPT is Wikipedia. Wikipedia is crowdsourced, which is wild, but also impressive considering that it is a fairly accurate first stop for information. However, one of its “pillars” that dictates how it creates and shares information says that only figures who are “notable” deserve coverage. By “notable” they mean figures who are verifiable, meaning covered by secondary sources (newspapers, print media, history books, etc.). Wikipedia itself does not encourage or allow the use of primary sources to authenticate facts. Sadly, this means I can’t submit my childhood diary as a source of my notability.

As you can imagine, this requirement limits who may be deemed “notable” and worthy of a page, especially if they lived in 1933 and newspapers defined them only by their spouse. For women, people of color, and non-binary folx, the requirement to be “notable” in popular coverage creates a huge issue of representation that Wikipedia projects like Art+Feminism, Women in Red, and Afrocrowd are trying to address by raising awareness and creating content around these subjects. This is not just a historical issue, however. In 2018, a Wikipedia user rejected a biography of Canadian laser physicist Donna Strickland. That October, she jointly won the Nobel Prize for Physics, leading to the creation of her Wikipedia page. Whether or not she met this news “gleefully” remains to be seen. (It is sort of charming to note Wikipedia keeps a Wikipedia page about its own controversies, which is an interesting read.)

AI Frida seems to take the dismissal of her work in her own lifetime pretty well. The “Ah, yes” at the start of her answer really took me out. It’s trying so hard to sound human! What’s also interesting about her response is she seems fully aware of things that have happened after her death, like the rise of her reputation as an artist. This is not something AI cosplaying as historical figures should be able to do. Even in the Washington Post piece, they tried to ask Harriet Tubman about DEI and reparations, and she shut them down. As Gillian Brockwell describes, this may be because AI is not supposed to allow historical figures to speculate about the present. However, reparations were absolutely a conversation in Harriet Tubman’s lifetime, leading Brockwell to speculate whether AI is trained to avoid statements that may be controversial in today’s politics.

This kind of “both sides” approach is evident in interviewing major political figures. I interviewed both George Washington and JFK, and both retained their capacity to answer a question without answering the question at all. For example, I asked George why people think he’s a great general when he was bad at his job? (I’m paraphrasing.) I also asked where his teeth came from to see if he’d engage slavery, as his teeth came in part from enslaved people. He did acknowledge that, leading to a suggested consideration of his legacy as both a notable person and a problematic one.

JFK, by comparison, refused to tell me who killed him, but did say he thought Bruce Greenwood’s portrayal of him in Thirteen Days (2000) was his favorite. This was a puzzling exchange to say the least.

All of this made me wonder what AI historical interviews are supposed to do for us? If we can’t fact-check the past (or get hot takes on the present), what is it for?

The answer lies in the history of seances themselves, specifically, the fraud mediums of the nineteenth century (and today, tbh). In these performances, mediums would prey on people’s grief to contact the dead. In some cases, they’d do research to accurately repeat details of a prominent person’s life, proving their “connection” to the dead. Can you imagine, for example, a medium channeling Abraham Lincoln for Mary Todd Lincoln and pretending to not know there had been a death in the theater that night? I imagine the experience is roughly akin to watching Tyler Henry, the Medium with a tv show, work with celebrities and pretend to stumble on easily-googlable life events. It’s like when Heather Mills aka Paul McCartney’s ex-wife once claimed to not have known who the Beatles were when they met. Okay, Jan.

The AI historical interviews I did felt like talking to someone not doing living history, but rather straining to pass off a very low-grade impression. They sort of sound like what you imagine the person sounded like, but the content is about a mile wild and an inch deep. There is also no logic to what they will and won’t comment on. Yes to the best film portrayal of them appearing long after their deaths, but no on any conversation about their deaths. I think the developers are trying to replicate the spirit of “The past is never dead. It’s not even past.” However, not to speak for William Faulkner, whose ghost I did not confer with for this article, I don’t think that’s what he meant. To be in conversation with history is to use its ideas, its ways of organizing information and structuring the world to think about our own choices today. That AI Harriet Tubman wouldn’t take a stance on reparations is not only incorrect, it’s insane. Of course she would have thoughts on that issue today, and her ghost should go off on it! Her reticence based on some perceived notion that she can’t comment on ideas after her death is a tell about the timidity of developers, not of its subject.

Of course dead people don’t comment on things outside their experiences or lifetimes! I don’t, for example, ask Cleopatra for stock tips or conflict resolution strategies. However, these are not sincere conversations with real people. It is software designed to selectively pull from the Internet, create a pattern based on a sketch of a person, and suggest things they might say. So why can’t the pattern allow for conversations about today? If we’re afraid these AI historical subjects would present as racists in a classroom exercise, we probably shouldn’t be celebrating them. (Hello, both sides now AI George Washington.) I’d love to see an AI historical figure who owned that they were talking with knowledge of today and talk directly to students about the things we should not romanticize about them. For example, maybe less of the founding fathers' “It was a different time!” apologia for racism and slavery and an acknowledgment that there were people in their time who knew it was wrong and behaved differently. That could lead to interesting discussions about what motivates people to hold these views then and now that would be meaningful for a class to discuss. That is using the past to think about the present.

I want to close by sharing my most unhinged exchange. To be clear, it was unhinged because I made it so. I wanted to ask a historical figure about one of my favorite pop culture moments. If you know me at all, you know that I was fascinated by Justin Bieber’s 2013 visit to the Anne Frank House. I had to ask AI Anne Frank if she might have been a belieber, if just for Justin’s sake.

AI Anne, not to take it to this place, but what do you mean? I need you to pick a side! I think this demonstrates everything right and wrong with AI. Yes, it has correctly clocked that Anne is significant for many reasons as a historical figure, and her diary is, among other things, a fascinating account of what it meant to be a teenager in the 1940s (especially during the Holocaust). However, what starts as an accurate description of her context turns into a coexist bumper sticker that feels afraid to admit to something as low stakes as Bieber fandom. If nothing else, teens have very strong opinions on pop culture. Though she didn’t live in our times, her ghost knows all (because it’s simply the Internet) and therefore should it not be predicting that she would be a fan of a teen heartthrob? Why is this so complicated?

Despite its claims to accuracy, predictive knowledge, and a whole host of other issues, my seance with a Macbook and with AI in general has left me puzzled. I get the impulse to want to invite people to talk to those who lived in the past. I also get wanting to use technology to do it. There are museums that now use virtual reality to invite visitors to visit Holocaust sites, for example. Another, the Illinois Holocaust Museum & Education Center, allows guests to “interview” holograms of Holocaust survivors. The museum creates this using “high-definition holographic interview recordings paired with voice recognition technology.” These innovations help preserve the experiences of people whose lives have been so singular and whose experiences we must not forget, and, importantly, on terms they could consent to. Everyone in this project likely knew that their words would be used to help kids understand why racism, hate speech, and antisemitism can be so dangerous. That “conversation” with the past is absolutely vital to our world today. A parlor trick that presents a “both sides” ghost is not.

You can read all of my AI conversations here: https://chatgpt.com/share/670e7441-c184-8012-9d72-39e318826ef3

These were readings that were helpful to my thinking on this:

The AI Pedagogy Project. “Interview a Fictional Character,” 2024. https://aipedagogy.org/assignments/.
Brockell, Gillian. “We ‘Interviewed’ Harriet Tubman Using AI. She Wouldn’t Bite on CRT.” Washington Post. The Washington Post, July 16, 2023. https://www.washingtonpost.com/history/interactive/2023/harriet-tubman-articial-intelligence-khan-academy/.
Illinois Holocaust Museum. “Interactive Holograms: Survivor Stories Experience - Illinois Holocaust Museum,” September 10, 2024. https://www.ilholocaustmuseum.org/exhibitions/survivor-stories-experience/.
Illinois Holocaust Museum. “The Journey Back: A VR Experience - Illinois Holocaust Museum,” September 4, 2024. https://www.ilholocaustmuseum.org/exhibitions/the-journey-back-a-vr-experience/.
Ushmm.org. “The Holocaust: History and Memory Virtual Field Trip - United States Holocaust Memorial Museum,” 2023. https://www.ushmm.org/teach/teaching-materials/primary-sources-collections/virtual-field-trip.
The 21stshow. “Interviewing Abraham Lincoln (GPT) | Illinois Public Media.” Illinois Public Media, August 2024. https://will.illinois.edu/21stshow/story/copy-interviewing-abraham-lincoln-gpt.

‌

From JFK to Frida Kahlo: My AI Séance with the Dead

With my Macbook, a dream, and some questions, I braved an AI seance with historical figures (and learned why we shouldn't use AI to talk to the dead)

Discussion about this post