A few weeks ago, a chatbot read a tranche of my text messages spanning more than a dozen years and declared that I’d been a bad friend.
First, some context: I’m writing a memoir about my love life, or lack thereof, so I’ve been going through old correspondence and quizzing friends about their memories of the years when I was single (2012–22) and not so single (2022–now). One of those friends is Michael—who, like me, is unmarried; who, like me, has an extensive history of encounters facilitated by dating apps; and who, like me, was long looking for love but often found himself disappointed.
My text history with Michael goes back to 2014, when we first met. Sometimes, we’re in touch daily. At other times, we go weeks, months, even a year without communicating (for reasons that have never been entirely clear to us). Six years ago, I moved from New York, where Michael lives, to Los Angeles, so since then, our friendship has been conducted largely via phone. This means that an already rich history of friendship has become richer, not just because of the accumulation of our shared experiences, but also because of the way our communications have been automatically documented and archived.
A text thread, when you think about it, is an incredible repository of primary source material—in effect, a data set—containing clues to how people conduct countless aspects of their lives, whether mundane or profound. For a memoirist, such material can be a gold mine. (I suspect that future historians will agree.) More freewheeling than, say, letters or emails, texts capture the immediacy of a writer’s feelings and observations and, because of that, can seem particularly “authentic.”
[Read: Archivists aren’t ready for the ‘very online’ era]
My initial quest for love-life insights within my archive with Michael yielded somewhat frustrating results. When I searched for words such as guy and date and sex, I quickly became overwhelmed and called him to complain: “There’s too much material for me to handle.” A few minutes later, my phone dinged. Michael had sent me an Excel spreadsheet containing our entire text history and suggested that I upload it into ChatGPT. I adored the idea. I’d been playing with chatbots for about a year and a half. For my memoir, I’d used them to pull out important strands from the transcripts of audio and Zoom interviews. I confess that I also once used a chatbot to give me an astrological analysis of a potential romantic relationship. So, with little hesitation, I took Michael’s advice, fed ChatGPT our data, and awaited the results.
They weren’t all that surprising, at least when it came to the bot’s analysis of my romantic life. (Among other things, it told me that I had bad taste in men.) But I noticed that it shared some wholly unexpected stuff about my friendship with Michael (for instance: “Neither has to construct a respectable version of the self for the other”). Interesting, I thought. ChatGPT had expanded its feedback beyond an analysis of my love life to an analysis of the relationship framing the discussions of my love life. I wondered: If I asked it to go deeper on Michael and me, what more might it say?
Never before have we had the means to synthesize, so quickly, words poured unadulterated into a single device. A few years ago, one technologist imported into an app he created every WhatsApp message he’d ever sent or received, and had the algorithm analyze his interactions with people including his wife and father. As he told a New Yorker reporter, “A lot of people don’t realize that, essentially, their WhatsApp history—across all their chats—is the diary they didn’t know they’d been writing.”
But conducting wide-scale analyses of entire relationships like this also has the potential to launch us into sticky ethical territory. Many people think of text messages as ephemeral and private (despite evidence of the various ways we can be easily surveilled). Our text threads with family and friends—the people we’re closest with—can therefore be quite honest, quite raw. “We had no idea when we started emailing and texting in the ’90s, 2000s, that these were records we’d be creating between us, which until that time were evanescent,” Jeff Hancock, a communication professor at Stanford University who studies AI-assisted conversation, told me. “Who knew that this would be a massive new thing,” he said, the fact that “private conversations between individuals would now have a record?”
The idea that a person can so casually take the contents of their written communications—with or without other people’s consent—and ask an AI tool to ingest and analyze that data seems disquieting, if not downright dystopian. Hancock said that people’s use of chatbots to audit what we previously thought of as closed conversations will no doubt become more prevalent. This might feel like an intrusion, he acknowledged, but he could see net-positive benefits. Imagine a scenario, he suggested, in which “you’re like, Hey, look, I want to do right by my sister. I’ve been annoyed and frustrated with her.” A bot analyzing your text history might come up with the perfect thing to say the next time you call her.
[Read: The people who outsource emotional labor to AI]
Even so, Hancock noted that there are probably “lots of ways this will go wrong.” Consider chatbots that might not only analyze relationships as they evolve but also predict outcomes—which sounds a little Minority Report to me. And then there’s the question of accuracy: Chatbots are designed to serve, which means that they tend to communicate what they think people want to hear. Sometimes, an AI might “really unlock stuff” for a person, Hancock said. “But they’re sycophantic.” To be useful, the AI needs “to also take the perspective of the other person.”
When I fed my text thread with Michael into ChatGPT, I found out—as I’d first hoped—a few things about my love life. I learned that I’m overly credulous when it comes to men’s stated intentions (“You can recognize that a man is unavailable, evasive, selfish, or inconsistent,” the bot said, “and nevertheless continue to treat his expressions of desire as evidence that something more meaningful might develop”) and that I’m looking for love in all the wrong places (“You sometimes recognize a man’s limitations but behave as though they are negotiable”).
Yet when the bot pivoted to spicy comments about Michael and me, it really drew me in. It said, for instance, that I have selective hearing and can be dismissive after he has offered well-meaning advice: “You ask him what he would do, but instead of seriously entertaining his answer, you almost instantly construct the case against it.” So, with Michael’s permission, I came up with a new query and pressed harder: “Please analyze this data set of text messages between Michael Wertheim and Anna Holmes, two friends, and craft a portrait of their friendship over the years.” I asked ChatGPT to look not just at what he and I had said to each other but at how we’d said it. After three minutes and 53 seconds, it told me that it had analyzed the entire archive—32,245 messages, from September 13, 2014, to June 25, 2026. It had a lot to say.
“Their language with each other is extraordinarily unguarded,” it told me. “They discuss work humiliation, money, parents, family dysfunction, depression and anxiety, illness, aging, sex, dating apps, hookups, rejection, medications, bodily problems, loneliness and death.” And: “They can move in seconds from something serious to something ridiculous, from mortality to television, from depression to a bad date, from career anxiety to cat litter.” And: “Their friendship eventually acquires some of the infrastructure of family.”
When I told a friend about this experiment and read her a few lines, though, she made a smart observation: It sounded like ChatGPT was presenting a pretty rosy picture of my part in the friendship. What if, she asked, I pushed it for a more candid assessment?
So I asked it to be more brutally honest. In response, it conceded that it had been “somewhat restrained” in its analysis—and proceeded to give me some real talk, telling me, in bold type, “There are moments where I think you are plainly the harder person to be friends with.” It also shared a rather damning 10-point list of my transgressions, including a paragraph or two of elaboration with each item. Among its observations: “You can disappear into your own life and assume the friendship will simply survive it.” “You can center your interpretation of his experience over his own.” “You can be relentless when you think Michael is handling something badly.” “You can be absorbed enough in your own distress that Michael has to announce his needs explicitly.” Ouch.
It stung because I’ve always worried that this is true; the bot’s assessment felt like a confirmation of that fear. It seemed to be saying: You’re easily distracted, myopic, over-opinionated, selfish.
If I wanted to give myself some grace, I could remind myself that even though ChatGPT had a corpus of communication between us, this was still merely a snapshot of the relationship, not a complete picture. Surely the AI was blind to bits of nuance and didn’t always understand context or intent. Yes, I can disappear into my own life and assume that the friendship will survive—but that’s the case with many close friendships. And yes, I can be “relentless,” but Michael has known from the beginning that I’m trenchant when giving advice because I assume that he wants honesty rather than enablement. (Though I probably need to ask him if that’s true.)
ChatGPT also represents only one AI-generated opinion. As I was writing this article, Apple updated its Siri AI to go through iPhone users’ text messages and emails to propose next steps (such as adding a calendar event based on a conversation with a friend). I realized that a tool for real-time friendship analysis might now be built right into my phone. To test it, I asked Siri to examine my text history with Michael—and it was much kinder to me than ChatGPT, even when I asked it to be less sycophantic: I’m a reliable and supportive confidant, it said; if our relationship is asymmetrical, this is because it “appears to revolve more around his needs and schedule than yours.”
What to make of this? Perhaps only that an AI’s analysis is subjective and possibly unreliable, kind of like the opinions of one’s real-life friends.
Text messages aside, Michael and I communicate meaningfully in other ways—via voice calls, in person. Those interactions aren’t memorialized in a text thread, so they’re perhaps not as easy to analyze. But they offer something that a back-and-forth over a smartphone doesn’t, and that I’d argue can be just as revealing: tone, timbre, facial expressions, eye contact. Put another way, they offer feeling.
After all of this AI analysis, I felt a mixture of fascination and unease. The insights that ChatGPT offered were more compelling than I’d suspected they would be, yet I’m hesitant to feed more of my texts to a bot. I’m conscious of the need for consent. And I’ve decided never to upload another person’s chats without permission, because the violation would be twofold: the downloading of the texts in the first place, and the analysis of them in the second. I also worry that gaps in a chatbot’s reading might lead me to wrongly or unfairly assess relationships that are important to me, or new and tender.
Take the text thread between my boyfriend and me. Our communications over iMessage are fairly thin. I have no doubt that if a bot were to analyze our texts, it would characterize me as overly chatty and him as reticent and at times unresponsive. In real life, I know that if he’s unresponsive, this is typically because he’s at work. But on a day when I might be feeling less generous toward him, I could imagine recalling the AI’s analysis and becoming resentful—or constructing relationship-doomsday scenarios in my head. And what good would that do? As another friend of mine put it: I would only do this sort of analysis with someone I didn’t want to be friends with anymore.
Chatbots may tell us a lot. But perhaps, as with astrology, we should consider their assessments of our most important relationships not as determinative but as opportunities to be more thoughtful—and we should see their advice less as an edict than as a nudge. I mean, all of us could probably be better to our loved ones. In the end, I don’t need a chatbot to tell me what I already know.
When you buy a book using a link on this page, we receive a commission. Thank you for supporting The Atlantic.
Leave a Reply