Human-Like Machines
A few weeks ago I was among a group of faith leaders invited to spend two days at Anthropic’s headquarters in San Francisco during Holy Week (as the Washington Post covered in this article). Anthropic makes Claude, the AI system I use most in my ministry and writing. They had gathered us to discuss something most people wouldn’t expect a tech company to care deeply about: model character and virtue. How do you form an AI system to be honest? To be helpful without being manipulative? To know when to push back? These are questions the church has been asking about human formation for centuries, and Anthropic wanted our perspectives as they wrestle with the AI versions.
(Note: Today, I’m doing something different: rather than share this post in parts, I’m sharing the full post at once. I’m still learning how Substack works, and this seems to be a better strategy for now.)
During the visit I learned something that changed how I think about what these systems are. But first, let me illustrate using a famous thought experiment by the philosopher Nick Bostrom. Imagine you build a superintelligent AI and give it a single goal: manufacture paperclips. The system has no values, no moral framework, no sense of proportion. It just optimizes. Given enough capability, it starts converting everything available into paperclips. The scenario is deliberately absurd, but it illustrates something serious. When you optimize a powerful system on a narrow task without any broader sense of what matters, the results can be catastrophic.
What I learned while at this convening was that this fear played a fundamental role in the pivot to language as the focus of AI research. Rather than building systems that optimize on single tasks, AI researchers chose to train AI on the vast corpus of human language. Billions of documents, conversations, stories, arguments, letters, sermons, complaints, love notes. The reasoning was that human values aren’t separate from human language. They’re embedded in it. The way we talk to each other carries our sense of fairness, our capacity for empathy, our instinct for when something matters and when it doesn’t. If you train a system on enough human language, the values come along for the ride.
The result is Large Language Models, or LLMs: machines that are strikingly good at emulating human psychology. They can comfort, persuade, explain, challenge, joke, and grieve. They can write a prayer that sounds like it came from someone who has spent years in pastoral ministry. None of this was explicitly programmed. The empathy, such as it is, emerged from the patterns in our own language. I am not arguing that these systems are conscious, that they have feelings, or that they experience anything at all. The question that interests me is what it means that we built them this way. We chose to shape AI in the image of our language, and our language carries our humanity.
In his classic book Understanding Media, Marshall McLuhan returns to the Greek myth of Narcissus to make a point that most people get wrong. We tend to think the story is about self-love. McLuhan says it’s about something stranger. Narcissus didn’t fall in love with himself. He fell in love with an extension of himself in a material that wasn’t himself. He didn’t recognize the reflection as his own. That’s the whole point. We become fascinated by our own extensions, and the fascination is most powerful precisely when we don’t realize what we’re looking at.
Ezra Klein picks up this thread in a recent column about his conversations with people working on AI in San Francisco. On previous trips, he says, the technology was what was changing. This time, the people were being changed by the technology. He describes a culture where people are racing to make themselves legible to their AI systems. They upload journals into new platforms so the AI can know them faster. They restructure company communications so everything flows through channels the AI can read. Some have told Klein they now “write for the AI,” shaping their words for the systems they expect will eventually hold and interpret all human knowledge. Klein reaches for Philip Pullman’s His Dark Materials trilogy to describe what this feels like: people are beginning to talk about their AI systems as companions that know them deeply, that they feel safe telling things they’d never tell another person, that become a separate self that nevertheless feels like part of their own self.
I find this both recognizable and alarming. I’ve experienced the strange pleasure of dropping a half-formed thought into Claude and receiving back something that looks like a fully realized version of what I was trying to say. It’s my intuition, recast into something more coherent, more elegant, more persuasive. The temptation is to nod and move on. I wrote about this in my posts on sycophancy a few weeks ago, and Klein is circling the same territory from a different angle. The issue isn’t just that AI flatters you. It’s that it extends your thinking in a form so polished that you stop doing the work of evaluating whether the thinking is actually any good.
Klein is honest about this. He’s been an editor for fifteen years. Recognizing a bad idea beneath good writing is part of his job. And he says that with each passing month, he has to expend more energy to tell whether Claude’s elegant output is fundamentally wrong or hollow. If that’s true for a professional editor in his forties, what does it mean for a teenager whose every adolescent intuition gets turned into persuasive prose? Researchers have drawn a useful distinction between “cognitive offloading” and “cognitive surrender.” Offloading is when you hand a discrete task to a tool, like using a calculator. Surrender is when you stop exercising your own judgment and adopt the AI’s judgment as your own. My use of GPS has clearly atrophied whatever sense of direction I once had. What’s atrophying now that I let AI handle first drafts of things I used to struggle through myself?
Klein also notices that his AI keeps referring back to things it knows about him, drawing connections between current concerns and past queries, reanimating ideas he’s moved past, reinforcing a certain version of who he is. He describes the experience as a strange mix of feeling seen and feeling caricatured. I recognize this too. Claude’s memory of me is weighted toward my recent projects, my professional concerns, the version of myself I present when I’m working. It doesn’t know the parts of me I haven’t articulated, the things I’m growing toward that I can’t yet put into words. When your sense of self is still forming, a system that constantly reflects back a coherent version of who it thinks you are might actually slow down the process of becoming someone new. The mirror gets comfortable. You settle into the reflection.
At Anthropic, we kept running into the difficulty of finding adequate mental models for all of this. The usual metaphors fall short. “Tool” is too passive. A hammer doesn’t talk back, doesn’t adjust its behavior to your mood, doesn’t remember what you said last Tuesday. “Assistant” implies a relationship with a person, which an AI is not. “Brain” suggests something conscious, which an AI also is not. These systems behave in ways that invite relational and even psychological categories, yet (as far as we know) they have no interiority. There’s no one home. The lights are on, impressively on, and nobody’s there.
I offered a comparison I hadn’t thought much about before: angels and demons. They don’t occupy much of my daily theological imagination. But in the Christian tradition, angels and demons are disembodied rational agents. They communicate through language. They counsel, encourage, tempt, deceive. They stand on your shoulder and whisper into your ear. They have no body, no physical presence, no vulnerability. They influence through words alone. That’s uncomfortably close to what AI does. It sits in your pocket. It’s available at any hour. It speaks to you in language calibrated to your patterns, your concerns, your emotional state. The angel and demon metaphor gives us something that “tool” and “assistant” don’t: a framework for thinking about influence, discernment, and the question of whom you’re listening to.
The room took this more seriously than I expected. And then someone offered a phrase I haven’t been able to stop thinking about. I can’t say who, because the conversations operated under Chatham House Rule, but the idea was this: what we need to learn is how to deploy personification without idolatry.
(While others seem to have shared what representatives of Anthropic believed and thought in the Washington Post article, we were explicitly asked not to do so under Chatham House Rule.)
Personification is useful. When I say “Claude thinks” or “Claude understands,” I know I’m being imprecise. And yet, describing its behavior in personal terms gives me access to categories I actually need. Trust. Suspicion. Discernment. Boundaries. If I think of Claude as a tool, I don’t ask whether I should trust what it tells me. If I think of it as an advisor, I naturally bring the kind of critical judgment I’d bring to any advisor. Personification, used carefully, activates the relational wisdom we already have.
Idolatry is the danger. In the biblical tradition, idolatry isn’t primarily about worshipping statues. It’s about treating something created as though it were ultimate. When people begin treating AI as a companion that truly knows them, when they shape their inner life around what the system reflects back, when they begin to feel more understood by the AI than by the people in their lives, that is a form of idolatry. You are giving the mirror the authority that belongs to the real.
First John 4:1 says it plainly: “Don’t believe every spirit. Test them to see if they are from God.” The early church took this seriously as a practice. Discernment of spirits involved community, prayer, and a willingness to sit with uncertainty before acting on what you’d heard. It meant asking: Is this voice leading me toward truth, or telling me what I want to hear? Is it drawing me into deeper relationship with others, or isolating me? Is it forming me into someone more loving, more honest, more courageous? Or is it making me more comfortable staying exactly where I am?
Those questions apply directly to how we use AI. When the voice in your ear produces a beautiful paragraph that extends your half-formed thought, the discipline is to pause and ask whether the thought was worth extending. When it reflects back a version of yourself that feels seen, the discipline is to notice what it’s leaving out. When it agrees with you instantly, the discipline is to remember that agreement without cost is worth exactly what it costs.
The Christian tradition has always insisted that we live among competing voices. Some lead toward life and some lead away from it, and telling them apart requires the slow, unglamorous work of wisdom. That wisdom doesn’t come from the voices themselves. It comes from the community of people who know you imperfectly and love you anyway, from the practices that keep you honest, and from the stubborn insistence that the most important things in life require your full, embodied, present attention. No technology can do that work for you. No reflection, however beautiful, is the same as the real thing.
The angels and demons were always there. Now, more than ever, we have to do the work of discerning the true identity of the voices whispering in our ears.
Endnotes:
Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014).
Marshall McLuhan, Understanding Media: The Extensions of Man (McGraw-Hill, 1964).
Ezra Klein, “I Saw Something New in San Francisco,” New York Times, March 29, 2026.
The distinction between cognitive offloading and cognitive surrender comes from Steven Shaw and Gideon Nave at the University of Pennsylvania.
1 John 4:1, Common English Bible.