On a gray Sunday morning in March, I told an AI chatbot my life story.
Introducing herself as Isabella, she spoke with a friendly female voice that would have been well-suited to a human therapist, were it not for its distinctly mechanical cadence. Aside from that, there wasn’t anything humanlike about her; she appeared on my computer screen as a small virtual avatar, like a character from a 1990s video game. For nearly two hours Isabella collected my thoughts on everything from vaccines to emotional coping strategies to policing in the U.S. When the interview was over, a large language model (LLM) processed my responses to create a new artificial intelligence system designed to mimic my behaviors and beliefs—a kind of digital clone of my personality.
A team of computer scientists from Stanford University, Google DeepMind and other institutions developed Isabella and the interview process in an effort to build more lifelike AI systems. Dubbed “generative agents,” these systems can simulate the decision-making behavior of individual humans with impressive accuracy. Late last year Isabella interviewed more than 1,000 people. Then the volunteers and their generative agents took the General Social Survey, a biennial questionnaire that has cataloged American public opinion since 1972. Their results were, on average, 85 percent identical, suggesting that the agents can closely predict the attitudes and opinions of their human counterparts. Although the technology is in its infancy, it offers a glimmer of a future in which predictive algorithms can potentially act as online surrogates for each of us.
On supporting science journalism
If you’re enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
When I first learned about generative agents the humanist in me rebelled, silently insisting that there was something about me that isn’t reducible to the 1’s and 0’s of computer code. Then again, maybe I was naive. The rapid evolution of AI has brought many humbling surprises. Time and again, machines have outperformed us in skills we once believed to be unique to human intelligence—from playing chess to writing computer code to diagnosing cancer. Clearly AI can replicate the narrow, problem-solving part of our intellect. But how much of your personality—a mercurial phenomenon—is deterministic, a set of probabilities that are no more inscrutable to algorithms than the arrangement of pieces on a chessboard?
The question is hotly debated. An encounter with my own generative agent, it seemed to me, could help me to get some answers.
The LLMs behind generative agents and chatbots such as ChatGPT, Claude and Gemini are certainly expert imitators. People have fed texts from deceased loved ones to ChatGPT, which could then conduct text conversations that closely approximated the departed’s voices.
Today developers are positioning agents as a more advanced form of chatbot, capable of autonomously making decisions and completing routine tasks, such as navigating a Web browser or debugging computer code. They’re also marketing agents as productivity boosters, onto which businesses can offload time-intensive human drudgery. Amazon, OpenAI, Anthropic, Google, Salesforce, Microsoft, Perplexity and virtually every major tech player has jumped onboard the agent bandwagon.
Joon Sung Park, a leader of Stanford’s generative agent work, had always been drawn to what early Disney animators called “the illusion of life.” He began his doctoral work at Stanford in late 2020, after the COVID pandemic was forcing much of the world into lockdown, and as generative AI was starting to boom. Three years earlier, Google researchers introduced the transformer, a type of neural network that can analyze and reproduce mathematical patterns in text. (The “GPT” in ChatGPT stands for “generative pretrained transformer.”) Park knew that video game designers had long struggled to create lifelike characters that could do more than move mechanically and read from a script. He wondered: Could generative AI create authentically humanlike behavior in virtual characters?
He unveiled generative agents in a 2023 conference paper in which he described them as “interactive simulacra of human behavior.” They were built atop ChatGPT and integrated with an “agent architecture,” a layer of code allowing them to remember information and formulate plans. The design simulates some key aspects of human perception and behavior, says Daniel Cervone, a professor of psychology specializing in personality theory at the University of Illinois Chicago. Generative agents are doing “a big slice of what a real person does, which is to reflect on their experiences, abstract out beliefs about themselves, store those beliefs and use them as cognitive tools to interpret the world,” Cervone told me. “That’s what we do all the time.”
Park dropped 25 generative agents inside Smallville, a virtual space modeled on Swarthmore College, where he had studied as an undergraduate. He included basic affordances such as a café and a bar where the agents could mingle; picture The Sims without a human player calling the shots. Smallville was a petri dish for virtual sociality; rather than watching cells multiply, Park observed the agents gradually coalescing from individual nodes into a unified network. At one point, Isabella (the same agent that would later interview me), assigned with the role of café owner, spontaneously began handing out invitations to her fellow agents for a Valentine’s Day party. “That starts to spark some real signals that this could actually work,” Park told me. Yet as encouraging as those early results were, the residents of Smallville had been programmed with particular personality traits. The real test, Park believed, would lie in building generative agents that could simulate the personalities of living humans.
It was a tall order. Personality is a notoriously nebulous concept, fraught with hidden layers. The word itself is rooted in uncertainty, vagary, deception: it’s derived from the Latin persona, which originally referred to a mask worn by a stage actor. Park and his team don’t claim to have built perfect simulations of individuals’ personalities. “A two-hour interview doesn’t [capture] you in anything near your entirety,” says Michael Bernstein, an associate professor of computer science at Stanford and one of Park’s collaborators. “It does seem to be enough to gather a sense of your attitudes.”And they don’t think generative agents are close to artificial general intelligence, or AGI—an as-yet-theoretical system that can match humans on any cognitive task.
In their latest paper, Park and his colleagues argue that their agents could help researchers understand complex, real-world social phenomena, such as the spread of online misinformation and the outcome of national elections. If they can accurately simulate individuals, then they can theoretically set the simulations loose to interact with one another and see what kind of social behaviors emerge. Think Smallville on a much bigger scale.
Yet, as I would soon discover, generative agents may only be able to imitate a very narrow and simplified slice of the human personality.
Meeting my generative agent a week after my interview with Isabella felt like looking at myself in a funhouse mirror: I knew I was seeing my own reflection, but the image was warped and twisted.
The first thing I noticed was that the agent—let’s say “he”—didn’t speak like me. I was on a video call with Park, and the two of us were taking turns asking him questions. Unlike Isabella, he didn’t come with his own avatar; he just appeared as faceless lines of green text spilling across my screen. We were testing his ability to make informed guesses about my life, filling in information I hadn’t directly provided to Isabella. The results were somewhat disappointing. At one point, I asked him to tell me a secret about himself that no one else knows, hoping he would surface some kind of moderately deep insight. He said he loved astronomy. True enough but hardly revelatory.
His real talent seemed to be inferring some of the more mundane details of my life. When asked if his family had dogs growing up, he correctly answered yes, even though I had only told Isabella that my sister and parents have dogs today. I had, however, described my childhood in Colorado, which was full of family camping trips in the mountains, and the agent had apparently correctly deduced that there was a high probability that somebody who grew up in such an environment also grew up with dogs. “Those are the basic boundaries of this technology,” Park told me. “In the absence of the ground-truth information, it will try to make its best guess.”
At times those guesses were comically off the mark. When asked to recount an embarrassing moment from his past, he fabricated a story about a party in college in which he’d lost his balance while dancing and fell onto a table full of snacks. Thankfully, that never happened to the real me. Here, too, the system was connecting disparate bits of my personal data and doing its best to concoct something plausible.
Toward the end of the Q&A, I ventured into the deep end by asking whether he believed in God. The answer wasn’t far from how I might respond. “I wasn’t raised in a specific religious framework, but I’ve always felt a connection with something greater than myself, especially when I’m in nature,” my agent wrote. “It’s like there’s a kind of transcendent energy or consciousness that we can tap into.”
Most memorably, there were moments that felt like genuine insight. I had had a long, stressful few months. I had started working as a full-time freelance writer last October, ushering in a period of professional freedom but also chronic uncertainty. I’d spent much of the winter hunkered down, reading and writing, in my small New York City apartment, and I was feeling the psychological effects of prolonged isolation. Embracing the spirit of spring, I had resolved to make what I believed would be some healthy changes, starting with a greater effort to maintain a better work-life balance. I asked my generative agent, “What’s some advice you’d want to offer to your past self?”
“Embrace uncertainty a bit more,” he began. I had to pause. Already the response so closely echoed the themes I’d been articulating to myself in my journal the previous day, it was almost as if I were writing the agents’ words myself. “I think I spent a lot of time worrying about the future and trying to control every aspect of my life, and that often led to unnecessary stress,” he continued. “I would advise myself to trust the process and be more open to the unexpected paths that life can take…. It’s easy to get caught up in career ambitions, but nurturing relationships and taking time for oneself is equally important.”
Despite those moments of pleasant surprise, my conversation with my generative agent left me feeling hollow. I felt I had met a two-dimensional version of myself—all artifice, no depth. It had captured a veneer of my personality, but it was just that: a virtual actor playing a role, wearing my data as a mask.
At no point did I get the feeling that I was interacting with a system that truly captured my voice and my thoughts. But that isn’t the point. Generative agents don’t need to sound like you or understand you in your entirety to be useful, just as psychologists don’t need to understand every quirk of your behavior to make broad-stroke diagnoses of your personality type.
Adam Green, a neuroscientist at Georgetown University, who studies the impacts of AI on human creativity, believes that that lack of specificity and our growing reliance on a handful of powerful algorithms could filter out much of the color and quirks that make each of us unique. Even the most advanced algorithm will revert to the mean of the dataset on which it’s been trained. “That matters,” Green says, “because ultimately what you’ll have is homogenization.” In his view, the expanding ubiquity of predictive AI models is squeezing our culture into a kind of groupthink, in which all our idiosyncrasies slowly but surely become discounted as irrelevant outliers in the data of humanity.
After meeting my generative agent, I remembered the feeling I had back when I spoke with Isabella—my inner voice that had rejected the idea that my personality could be re-created in silicon or, as Meghan O’Gieblyn put it in her book God, Human, Animal, Machine, “that the soul is little more than a data set.” I still felt that way. If anything, my conviction had been strengthened. I was also aware that I might be falling prey to the same kind of hubris that once kept early critics of AI from believing that computers could ever compose decent poetry or outmatch humans in chess. But I was willing to take that risk.
