The voice machine

Last fall, The Atlantic magazine revealed that Meta’s large language model AI, LLaMA, had stripped, in secret and without permission, 170,000 books in an attempt to teach itself English. It was an extraordinary number to consider. All those words! Words that represented the thoughts, experiences, imagination and souls of their writers, ingested, without any concern for the well-being of the author(s), financial or otherwise. All so that AI could learn to speak to us in a way that might make us, the spoken to, feel as though we were engaging with a real person. Without actually having to engage with a real person, of course. So much art plundered and redirected so that AI can better provide the user with driving directions, or press releases, or assistance in writing a term paper, or a facsimile of art. Art to create fake art.

The Atlantic story included a search engine, allowing you to plug in an author’s name and see how many of their titles existed in the database of stripped work. Almost immediately, many of the writers I know, or follow on social media, began searching for themselves. Some had had their entire body of work pulled. Some, one or two titles. Many of them took to social media to express their outrage over the results (the outrage mixed with a subtle, but palpable, dose of pride, it must be said. They were AI worthy! Even if AI was coming for their livelihood). Curious, I went and searched myself. My 2018 memoir about being 40, single and childless had not hit The New York Times bestseller list but it had done well and five years on, remained in print with healthy sales. I continue to hear regularly from readers about it and it still pops up on various lists.

I typed in my name.

Nothing.

I didn’t bother to post this nonresult. But a few days later, again out of curiosity, I searched for a few other authors I know. Black women who had written New York Times bestsellers. Books on female friendship. Memoirs. Black history. They weren’t there. Neither was Elliot Page’s memoir, which debuted at number one on the New York Times hardcover nonfiction best-seller list. Playwright Jeremy O. Harris was missing, too, despite picking up a record-setting 12 Tony Award nominations in 2021.

Hmm.

One of the reasons I wrote my memoir is that I had arrived at a certain age and encountered an absence of stories that seemed to reflect, in any way, the life I was leading. I felt like I’d been dropped into a cultural black hole. The book was less an account of my doings than it was evidence of existence. Prior to selling the book, I had written essays on similar topics that had gone viral, which is to say, my sense that I was isolated was not based on the experience of feeling alone in my life, so much as the frustration of seeing many people leading similar lives to mine (numbers backed up by studies that show there are more unmarried women right now than ever in history) and never seeing these lives accurately reflected in stories. I wanted to provide some language to give life to that experience where there had been little before.

Not finding my book in the LLaMA list felt like a re-erasure of sorts. I had put myself forth. Spoken. Been heard. And now had disappeared again. It was a strange Catch-22: On the one hand, I had not been fed, non-consensually and unpaid, into the digital maw. On the other, my absence from that list also represented one more small example of our human fallibilities, one that AI is training to reinforce.

It will reinforce on an extraordinary scale: One often-referenced estimate projects 90 percent of the internet could be artificially generated in the next few years. We are, as I write this, rebuilding the foundation of how the world will likely operate for years, if not centuries, to come. The materials we’re using, algorithms and databases, will determine what will rise up to become the future sources of truth — the guidebooks for the future to understand this past.

Among the more important questions that need to be asked when you consider this is: Who will get to tell our truths? Who will be deemed worthy of a story? Whose words, thoughts and experiences will be guiding us forward, and who will be left out? Who is this world being built for?

That certain voices had been left out of LLaMA’s great vacuuming up of culture was not necessarily a surprise. The limits of AI’s training models are well-documented. AI has had difficulty recognizing Black women especially, for instance, frequently categorizing them as men. A recent Bloomberg investigation found AI-generated images for higher paying jobs automatically were “dominated by subjects with lighter skin tones.” These bias issues, now seemingly being repeated in the sources for language models, speak directly to the persistent problem of Silicon Valley itself, one that has long plagued the startup world: It is dominated by white men.

Indeed, after perusing the LLaMA book database, it looked hauntingly similar to the bookcase of a hypothetical college-educated man who, having invited you back to his apartment after a date, is eager to impress you with his interests. Here is the Shakespeare, here is the Edith Wharton, here is the James Baldwin. It is the bookcase of a Manhattan brownstone with the curtains left open at night, and lights ablaze, so passersby can be impressed. Look! it says. We read! Or at least, we perform reading.

As The Atlantic pointed out, it is a very Western bookcase. “A system trained primarily on the Western canon, for example, will produce poor answers to questions about Eastern literature.” Or many other things.

In his novel “The Book of Laughter and Forgetting,” Milan Kundera (also not in the database) wrote, “One morning (and it will be soon), when everyone wakes up as a writer, the age of universal deafness and incomprehension will have arrived.”

I have thought of that line often in the last decade as it felt like Twitter, now X, could be described as everyone waking up as a writer. While certain elements of Twitter/X did feel deaf and incomprehensible, it was also a cacophony of voices and experiences that shifted how we thought and spoke about the world and our place in it. The age of AI that is barreling down upon us seems to suggest the opposite.

Deafness and incomprehension may not seem likely on a generative AI model whose number one author source is Shakespeare (“There’s nothing either good or bad, but thinking makes it so!”). But if the people programming it look exactly the same as the people who have largely always programmed how we communicate (be it with a quill and ink or code) then all this change has the potential to fast-track us into a new, albeit even more punishing, version of a past era. Whether that’s progress or not is up for debate.

It’s not surprising then that communities who have been ringing the alarm bell the loudest over AI’s bias are the ones who suffer most from either their exclusion or the distortion of their identity. For instance, computer scientist Timnit Gebru is one of several high-profile women of color who saw this coming, writing in a widely read 2021 paper that: “White supremacist and misogynistic, ageist, etc., views are overrepresented in the training data, not only exceeding their prevalence in the general population but also setting up models trained on these datasets to further amplify biases and harms.”

If past is prologue, then we have a good sense of where we’re headed if these biases are allowed to prevail. And it’s directly back into a history that only allowed for one voice.

That solitary voice has been variously referred to as the voice of the victor, and the voice of the person who buys ink by the barrel. These days it’s the voice of the algorithm. What they all have in common is that they are almost always male and almost always white. The result is centuries of human existence during which most of the stories, and subsequently laws, are being written and made by very few people.

We know what that past looks like. It looks like a lack of information. Like a slice of Swiss cheese, where the holes are missing context. Often, it looks like how our laws are applied. Joan Didion (15 entries) famously wrote, “We tell ourselves stories in order to live.” But in truth, the stories we tell determine how we live.

In my case, it looks like every story about being a woman ending happily at the altar. The past few years have seen an increasing panic over the declining rates of marriage in America. They are at their lowest ever recorded. In a February interview with The New York Times about his book “Get Married: Why Americans Must Defy the Elites, Forge Strong Families, and Save Civilization,” University of Virginia professor Brad Wilcox said he believed one of the challenges facing the institution of marriage (2021 clocked the highest number of unmarried Americans ever recorded) is the “soulmate” myth.

“Our culture, our pop culture in some ways especially, will often give us what I call the soulmate myth. And it’s this idea that there’s some perfect person out there waiting for you and that once you find them and love them and then marry them, you’ll have this perfect connection that engenders intense emotional connection, sense of romance, passion that in turn leads you to be happy and fulfilled most of the time.”

Wilcox goes on to detail that we need more realistic storytelling when it comes to marriage in the hopes it will let people know what they are getting into, and subsequently lead to less disappointment (on the part of women, presumably, since research tells us men are happier and do better when married).

Those stories do exist, as it happens. And not just on numerous women-focused websites or magazines. In 1987, the sex educator and author Shere Hite (five entries) published “Women and Love,” the third part of her bestselling Hite Report, analyzing the emotional relationships between women and men. Hite found that the majority of married American women (to the tune of four-fifths) believed their marriages to be unequal and that there was widespread dissatisfaction. The Hite Report sold an estimated 50 million copies, and yet this result was so at odds with the established marriage plot, or what Wilcox calls the soulmate myth, Hite was chased out of America and driven into exile in Europe. One wonders if these stories had been taken more seriously, if they had been included in the algorithm as it existed at the time, would we live in a time and in a country where people, marriages and families are healthier?

What is happening now feels similar. This is not a moment when only a few people have access to platforms, or where there are few ways in which to make their voices heard. Quite the opposite. We have a cacophony. And yet, what those who are programming AI are seemingly doing, intentionally or otherwise, is choosing to not utilize these voices and this knowledge, both from our past and present, at the risk of our future.

AI is developing at a furious pace. In less than two years it has upended how we communicate. Yet it’s useful to remember that what’s currently emerging hovers somewhere between absurd and comical. Google’s latest foray into large language models, Gemini, appeared to rewrite history, replacing white historical figures with actors of color à la the “Hamilton” musical. Even when this was corrected, the resulting voice was less like talking to an actual human than being stuck in conversation with what journalist John Herrman described as a “a customer service rep denying a claim at an insurance company.” Which admittedly, as those of us who’ve been stuck in a chatbox trying to get a phone connected, or any other kind of service supplied, knows can be terrifying in another sort of way, but is not quite apocalyptic.

It’s easy to laugh. And it makes for fun coverage, allowing us to ward off the panic, voiced by many experts, that AI will render most human work unnecessary and/or accidentally set off a nuclear apocalypse. But another thing history has taught us is that things that begin as a joke can quickly turn deadly serious.

Right now, AI in all its foibles remains, ironically, very human. It’s a mess. It is an unwieldy technology that has collided with an unwieldy world. There is the potential for this all to be very exciting. For everyone to be able to see themselves in a technology that will likely infiltrate our lives in ways that even amid all this change, can be hard to imagine. But we must imagine it. This moment is demanding from us, not new stories, necessarily — we are not a world lacking for stories — but an insistence that we imagine ways in which they will remain heard.

Glyniss MacNicol is a writer and podcaster whose latest book, “I’m Mostly Here to Enjoy Myself,” will be published this summer.

This story appears in the May 2024 issue of Deseret Magazine. Learn more about how to subscribe.