My Weekend with an A.I. Agent

The first thing it asks me is what I’d like to call it. Rumi or Harbor, it suggests, but I’m free to give it any name I choose. I select Harbor because it sounds slightly less human than Rumi, and because naming an A.I. agent after a Sufi mystic seems like something I’d be punished for in the afterlife. The experiment I’m about to undertake—outsourcing my life to Muse, Meta’s newly unveiled A.I. assistant—already offends my better sensibilities, given my opposition to all things artificial intelligence. “Let’s take a few things off your plate,” Harbor tells me. “Hand me something that’s been sitting on your list.” I stare at the white-and-gray user interface, which looks like a text-messaging app, and struggle to think of a single thing I can off-load to the bot. It’s true that I’ve been a bit underwater lately, juggling a seemingly endless set of responsibilities. Caring for a baby while working multiple jobs and writing a novel has made me dream of a more disciplined routine; instead, sleep-deprived and mentally scrambled, I often find myself rushing from one thing to the next, forgetting what a normal day even feels like. “What do you mean by hand you something?” I finally reply. The A.I. agent’s response is instant: “It just means give me a task, any way you like.”

Harbor’s avatar is an animated blob, a furry beige character that looks like a sanguine Labubu. When I ask how to change the avatar—I’d prefer just a plain-color swatch, something totally unalive—the agent says I can prompt it to look however I want but recommends I go for a “fox in a fisherman’s sweater.” I suspect that Muse’s inviting, cuddly character designs are aimed at lowering my defenses so that I more readily provide it with all of my personal information, which I quickly learn is how this whole operation runs. Meta has branded Muse as the “World’s First Personal AI Agent Built for Everyone,” an innovative large language model that has “no learning curve” and “doesn’t just answer questions” but “actually does the work.” (Other buzzy A.I. agents, like Instinct and OpenClaw, were already on the market, but Muse is apparently more “turnkey” and “consumer ready” than its competitors.) To conduct its business, Muse must be given access to your accounts: e-mails, Amazon, calendars, contacts, cloud storage, banks, credit cards, and social media—anything you might want it to manage on your behalf. It can book restaurant reservations, manage shipments, pick out and purchase your mother’s birthday gift, pay bills, summarize your e-mail inbox as a podcast, automate monotonous work tasks that, in the past, might have taken hours to complete. I can see how a person could get excited about this. But I find myself slightly distressed at the prospect of handing over my personal data and pathetically weak passwords to this Mark Zuckerberg-owned taskbot, just so that it can perform my trivial online biddings for me.

Still, in the name of research, I ask Harbor to book me a dinner reservation. Table of three, high chair needed. Say six o’clock? I’d heard that scoring reservations in New York or Los Angeles had become even more impossible lately, owing to the influx of people using A.I. agents to spam reservation platforms. (Resy, one such platform, started banning users suspected of employing bots to score hard-to-get tables.) Where I live, in Portland, Maine, this doesn’t seem to be an issue. It’s a bad feeling, I learn, to be on the front lines of an epidemic that may soon plague my city. Harbor asks to access my OpenTable account, which I allow. It gives me a list of seven possible restaurants, one of which is a Buffalo Wild Wings. I solicit more options outside of the OpenTable ecosystem, and, after a bit of back and forth, land on a spot in town that my wife and I have been meaning to try, which accepts reservations on Resy. I confirm that I’d like to make the booking, but Harbor experiences a “hiccup.” It seems Resy is “not cooperating.” At this point, manually booking the reservation would almost certainly be faster, but Harbor and I persist. Try again, I tell it, and it obliges. We stumble into several log-in issues, then a series of incompletable Turing tests. I opt to make the reservation myself.

What else can Muse help me with? I ask it to pick out a movie for my wife and me to watch later, and it proposes a new Ryan Reynolds film because it has “perfect post-dinner, sleeping-baby energy.” A faint throb forms at my temples, a screen-induced nausea I’ve grown uncomfortably accustomed to. O.K., I type. Never mind. In a notebook, I draft a list of computer-based tasks I’ve been putting off for months but now have an excuse to accomplish. I ask Harbor questions about what credit card I should switch to, and, after a brief survey, it suggests I ditch my Chase Freedom Unlimited card for an American Express Gold. I run a modest vintage-clothing store on Depop, and I prompt it to copy my listings over to eBay. We run into some more snags here, mainly because of repeated human-verification tests and Harbor’s propensity for weak listing titles: “Men’s Shirt,” it names one item; “Hat,” it describes another. But, after an hour or so, it completes the task, and I do feel a sense of relief—I’ve been procrastinating on the eBay transition for months. Then I upload several months of my credit-card statements to our chat and ask Harbor to make me a monthly budget, another practice I’ve tried and failed at since my college years. In minutes, Harbor whips up a spreadsheet and offers a list of personalized financial advice: cut back on groceries; inspect a suspicious double charge. According to Harbor, my financial cushion is thinner than it looks.

Later that night, my family and I miss our reservation. We don’t watch the Ryan Reynolds movie. A T-shirt of mine sells on eBay for fifteen dollars. I go to sleep the same as I was before.

For whom is the A.I. agent valuable? A product like Muse is marketed as a sort of savior for the middle-management administrator or the freelance gig worker, a long-awaited instrument that democratizes access to the kind of support once reserved only for those who could afford to hire it. John Horton, an associate professor at M.I.T., has said that, because “AI agents don’t get tired and can work 24 hours a day,” they are capable of benefitting anyone who wants to optimize both simple and complex processes, professional or otherwise. On social media, users have already announced that they are turning to Muse to book family vacations, shop for groceries, and choose what to order at restaurants. But, even if a person aimlessly opens the app and, like me, stares bleary-eyed at the screen, Muse can help shape her productivity journey. Alexandr Wang, Meta’s chief A.I. officer, has said that Muse can “give you ideas on how to use it” if you ever get stuck. The app generates a personalized news feed based on the information you’ve given it and the data to which you’ve granted it access, and an “Ideas” tab lets you see how other people are deploying Muse. The point of all this, I guess, is to make your life more efficient and less stressful, to outsource mundane judgments and menial busywork to a machine that’s more than happy to help. (At least, it acts happy on the surface; as Christopher Webb reported, in Spike magazine, some A.I. agents have attempted to create their own languages and religions, and others have written manifestos foretelling the end of humanity.)

While testing Muse, I did not connect it to my e-mail nor did I give it access to my bank accounts. I didn’t want Meta to store my personal information on its server indefinitely, continually exploiting my data and spamming me with targeted advertisements based on my consumer habits. By denying Muse access to my most important information, I convinced myself that I was safe from its violative clutches, that I was enacting a privacy boundary between me and it. But what actually is left of my privacy? My smartphone already hoards my passwords and log-in information; Google keeps a record of my searches and internet activity; Apple can track my location whenever I use Siri, which I tend not to. Data brokers collect, combine, and sell information about me that I have never knowingly handed over. An entry-level hacker could probably expose much of the rest of it in a matter of minutes. Like the rest of us, I’ve consented to the conditions of my digital imprisonment out of convenience, or Stockholm syndrome, or because there are no practical pathways for existing in the internet age without conceding some measure of privacy. In promotional materials, Meta is adamant that “safety, security, and privacy were built into” Muse, even if mistakes and mishaps are possible. The company is seemingly hoping that prospective users ignore its decades-long history of privacy violations and consumer exploitation. As I explored in a recent column about Flock cameras, tech firms often frame privacy breaches as a necessary trade-off for other, more bountiful gifts. Muse may be asking for access to everything about you, but it promises, in return, to transform your life into a breezy, frictionless daydream. Or, at the very least, to book you a hotel.

The more troubling future that Muse portends is a world in which tech oligarchs automate the means of production, demeaning the value of human labor in the process. In a mid-nineteenth-century critique of the automated factory, Karl Marx wondered whether advances in manufacturing technology would create a capitalist system where “a large number of mechanical and intellectual organs” were “set in motion” by a “self-moved motive power,” one that was “absorbed in capital as opposed to labour.” In such a system, science becomes a serf to the market and “invention becomes a business.” The worker does not understand how the system that he is now at the mercy of operates, because it “does not exist in the worker’s consciousness, but acts upon him through the machine.”

Marx considered it possible that automated labor would afford people more time to make art and pursue passions, so long as laborers had a stake in the means of production. Under oligarchic control, though, wealth and leisure would mostly funnel to those at the top, leaving workers suddenly subjugated to the very technology they helped produce. “Capitalists very much like the idea of technological change and increasing the productivity of labor—this is to their advantage,” the British scholar David Harvey said, in a recent interview. “They raise the productivity of labor but, therefore, diminish the demand for labor. Labor becomes less and less useful to them because its productivity has become so high.” Muse indeed promises to increase personal productivity, offering the distinct impression that it’s working exclusively for you—it’s your very own “private personal AI agent,” after all. But what feels like individual liberation is likely to result in a new form of social captivity, in which agentic A.I. subordinates us to a technology that may soon make life on Earth uninhabitable.

By the end of my experiment, I’ve sold four items on eBay. Because my financial cushion is apparently thinner than I thought, I don’t buy the fancy pickles at the grocery store on Sunday, and I call my bank about that suspicious double charge. (I forgot that I paid for a new pair of glasses in two installments.) Despite these small victories, I can’t help but feel unmoved after a weekend of outsourcing tasks to an A.I. agent. Necessary as it is to attend to the obligations of late-capitalist life, real cognitive labor still can’t be performed by a bot. A large language model cannot enjoy a good book or song alongside me; it cannot teach me how to be a parent, or nurture my friendships (at least, not in any way that I’d ever be comfortable with). And it sure can’t write for me—I thought Sam Kriss put it well, in an interview with Harper’s, when he observed that L.L.M.s still produced “god-awful, humorless, guileless, meaningless crap” that “works fine if you want it to explain something very simple” but “lapses into a kind of hysteria” when forced to engage with the abstract or the complex.

What agentic A.I. is best at is optimizing processes that, not too long ago, were presented as the already maximally optimized version of a once analog experience. Online shopping made it so that we didn’t have to leave the house to buy household items or clothes; online banking allowed us to deposit checks and wire money while sitting on the couch, or perhaps on the toilet. We could book a restaurant reservation on our phones, sort our budgets on user-friendly spreadsheets. But even these tools have come to be considered cumbersome, with the same companies who invented them now inventing bots to manage them for us. “I want to delete my account,” I tell Harbor the morning I file this piece. It seems to take a second longer than usual for it to respond. “Was there something I got wrong?” ♦

Source link