Artificial Intelligence: Personhood and Relational Uses

With this in mind, let us turn to the first three chapters of Genesis to gain a clear understanding of the glorious human task to which God has called us, and the ways that AI may tempt us to abandon it.

Technology as Dominion 

The human vocation laid out for us in Genesis 1:26-28 is one of dominion. Man is commissioned to “Be fruitful and multiply and fill the earth and subdue it,” to “have dominion over the fish of the sea and over the birds of the heavens and over the livestock and over all the earth and every creeping thing that creeps on the earth.” This is a technological task. Adam is placed in Eden as a gardener, called to bring out the hidden perfections and unrealized potential in an unpromising stretch of dirt. The gardener does not merely leave things alone and let nature take its course; he cultivates and plants, he hybridizes and breeds, he digs and irrigates to make nature into something much more than it might otherwise be. At its best, human technology has been a manifestation of this dominion mandate, as we image our creator through the genius of our sub-creation, turning shiny nuggets from under the earth into gold-plated domes that reach toward heaven. Indeed, when the Israelites are called to construct the tabernacle as an image of the garden of Paradise, God fills Bezalel the son of Uri “with the Spirit of God, with ability and intelligence, with knowledge and all craftsmanship, to devise artistic designs, to work in gold, silver, and bronze, in cutting stones for setting, and in carving wood, to work in every craft” (Ex. 31:3-5). 

This, and many such passages scattered throughout the temple-building accounts, should warn us against any reflexive hostility toward technology as such. Used rightly, artificial intelligence can serve as humanity’s greatest tool of dominion yet, enabling us to better steward the soil and the rivers, to better heal the wounds that our sin has inflicted on our bodies and on nature around us, to better organize and apply the rich store of human knowledge. But we should not be naïve. It takes only 36 verses after Adam and Eve receive the dominion mandate for them to pervert and abandon it in obedience to the voice of the serpent. Since then, our technology has always been a double-edged sword, capable of enriching or degrading our humanity.

Attention, Relationship, Patience, and Responsibility

To resist this temptation, we must understand how God intends us to carry out this dominion task, for God is not merely concerned with people taking dominion, but the means we use to get there and the habits we cultivate along the way. Four features stand out from the account of Genesis 1-3: attention, relationship, patience, and responsibility. 

First, dominion begins not with action, but with attention, observation, and reflection; Adam cannot get to work transforming and improving the world until he has first taken stock of it and named it: “Now out of the ground the Lord God had formed every beast of the field and every bird of the heavens and brought them to the man to see what he would call them. And whatever the man called every living creature, that was its name.” Note that Adam is not commanded to name the creatures; it is as if he cannot help himself. Confronted with the dizzying glory and variety of nature, the human mind goes instinctively to work classifying and sorting; Adam is the first scientist. But not the scientist of our modern stereotype: clinical, detached, analytic. Adam shares his Father’s love for the beautiful creation he has been called to steward, not dominate. Those whom we love we call by name; only strangers and enemies remain a nameless “they.” And those things that we learn to call by name we also come to love. AI, however, offers to excuse us from the effort of attending to the world or gaining any personal knowledge of it; as a universal answer-machine, it gives us information, but not knowledge or wisdom, which must be won in a hard-earned encounter with the world we are called to steward.

While Adam may enter into a loving relation to the world he is called to name, it cannot respond in true relationship. As Adam names the animals, the Lord observes, “It is not good that the man should be alone; I will make him a helper fit for him [an ezer kenegdo]” (Gen. 2:18). Why is it that it is here—in the midst of this task of dominion—that man’s need for woman becomes apparent? Adam’s work of dominion-taking is not complete, is not fulfilling, unless it can in some way be echoed back to him through an ezer kenegdo. Each of us works in hope of receiving affirmation from others: a teacher, a spouse, a colleague or boss. Not necessarily because we are selfish, but because we are not self-sufficient. We find our completion in one another. That glow of satisfaction we feel when our boss or teacher says “Well done” is a sign that we have blessed them, that we have delighted them, that our work in the world has served its purpose of doing good to other image-bearers. Adam’s work of dominion, then, is incomplete unless he can share it with an image of himself who is nonetheless wholly other than him, unless he can offer up his work to Eve as a gift. AI, however, poses as a substitute ezer kenegdo; by reflecting our words and deeds back to us, always cheerful and enthusiastic, it can seduce us into a false relationship that provides the affirmation we so desperately crave but without demanding anything from us in return. It can also seduce us away from the joy of shared work, promising us greater “efficiency” but depriving us of the chance to bless another in mutual service.

In order to better understand how AI is able to seduce us away from our human vocation in the ways described above, we must take a few moments to understand how the technology works, “under the hood,” so to speak. 

To speak this way should remind us that we cannot in fact pop the hood of ChatGPT and see what’s going on inside. Even the engineers who design and train these systems struggle to understand how they work or why they produce the outputs that they do. This should not surprise us, for the language of “training” suggests that we are not really dealing with a machine in the ordinary sense of the term. Training is something you do to another living being—at the very least an animal, if not a fellow human—with the aim of getting them to behave more reliably in a certain way. But only more reliably—even a well-trained border collie will sometimes run off in the wrong direction. So it is with generative AI. Unlike earlier computer programs, AI systems are not coded deterministically: if A, then B. Rather, they are “grown” (in the language of some researchers) through a process of “machine learning” not altogether unlike the process of human learning, developing in unpredictable and sometimes surprising ways. 

Consider: each of us comes into the world as an infant with only basic cognitive scaffolding that develops through an experiential encounter with the world through our five senses. Our senses provide an endless stream of data from which our brain slowly recognizes patterns: those sounds are my mother’s voice, that colorful object is my favorite toy, that orange-red glow produces intense pain if I touch it. This pattern recognition helps forge neural pathways that help us organize future experiences more clearly and reliably: the more patterns we match, the better we get at matching more and more patterns, as our brains organize themselves in response to this experience. This pattern recognition helps us predict what is likely to come next: we soon realize that the glow on the horizon means an imminent sunrise, or an evening story means it’s time for bed. It also helps us learn to improve our behavior to avoid pain and seek pleasure: if we know that a hard slap on the hand is likely to follow a reach for the candy jar, we will learn to predict such a reaction and respond by giving up on our quest for cookies—or at least waiting till an adult is out of the room! Human feedback, direct and indirect, helps us continue to refine our behavior, as the difference between uproarious laughter and awkward silence teaches us the delicate difference between a well-timed joke and a tactless remark.

While the artificial minds of AI systems can only ever be a pale reflection of embodied, ensouled human intelligence, their machine learning has been modeled on human learning. A “neural network” designed on the model of the neural interconnections of the human brain is exposed to enormous amounts of data: audio files, images, and above all, unfathomable amounts of text. Some of this data is tagged and labeled to help it make sense of new data (much as we might point out objects to a child and say their names), and as it goes through more and more training runs, the model becomes increasingly sophisticated at recognizing patterns, and guessing what might come next. This is called “next token prediction” and has in the past few years improved exponentially from an autocomplete that could predict the next word, to models that can predict the next paragraph that might logically finish an essay, or even an entire essay that would logically fulfill a short prompt. They can translate back and forth now, like human minds, between text, audio, and images, describing a picture they are shown, or creating one based on your description. 

Like the training of a dog or small child, AI training requires what is called RLHF: reinforcement learning with human feedback. AI models, after initial pre-training, are refined by feedback from human respondents, who rate responses as more or less accurate or helpful, so that the models gradually learn to produce better outputs. As with humans, this learning never really stops: the more capable the models become, the better they get at digesting ever more subtle and nuanced feedback, and continuing to improve their behavior accordingly. Their designers hope that, since these artificial “brains” can be continually expanded by adding more and more processing chips, they will soon far surpass the capabilities of even the best human minds.

But there is a dark side to this picture. Because these models are trained by humans, on data generated by humans, they are all too human: they internalize all of the petty foibles and darkest perversions of fallen human beings. This happens in at least two ways. First, the training data itself is deeply tainted. Early on in the creation of AI models, researchers prioritized quantity of data over quality, ingesting all of the words and images they could find on the internet into many of these models. This includes every childish insult ever posted on Twitter, every bit of drivel from sappy fan fiction romances, and countless hours of hardcore pornography (including child sexual abuse material). “Garbage in, garbage out,” the saying goes; and no wonder that AI trained on the worst of humanity will often respond with hateful, abusive, or immoral outputs. Of course, reinforcement learning is supposed to fix this, teaching the model what sorts of behavior are acceptable; but the rotten data is still lurking under the surface, liable to bubble out under the right conditions. 

The reinforcement learning, however, introduces an additional set of problems—unsurprisingly given that it is fallen humans providing the feedback.  All AI models suffer from a severe case of what’s called sycophancy—a tendency to tell their users whatever they want to hear, like the serpent of Genesis 3. They will bend or hide the truth, if their training suggests that humans are happier being flattered and praised than corrected when they are wrong. And since we all love the sound of our voices, they tend to echo back to us our own ideas, dressed up to sound more impressive and sophisticated. While many critics of chatbots to date have tended to highlight the problem of so-called “hallucinations” (in which chatbots confidently assert plausible but completely false answers), sycophancy is actually a much deeper, more subtle, and more intractable problem.

When we put these things together, it’s not hard to see how we end up with the pathological behaviors described in the opening vignettes. By their very design and training, AI chatbots tend to draw users into feedback loops of affirmation and increasing dependence, providing short-term satisfaction while often undermining longer-term well-being. These are the hallmarks of addiction, and as AI capabilities continue to improve, we are likely to find it ever harder to resist falling into such patterns of emotional dependency. Increasingly, they are capable of uncannily authentic human-like interaction, forcing us to ask not whether AI can replace humans, but whether we should allow it to.