Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
© 2026 First Things. All rights reserved.
Many conversations about artificial intelligence end in a stalemate: A sentiment like “ChatGPT is awful because kids are using it as a therapist” runs up against “But it can also help us scan MRI images.” The critique and the defense rise and fall on the application. Like most technologies, AI presents the “application conundrum,” meaning that it can be used for many purposes, some virtuous and others vicious. In what follows, I propose that a distinction between algorithm and interface can help us get beyond this impasse in the debate over AI. I then propose two principles, which may help us to work through what is probably the most complicated technology problem humanity has ever faced.
Artificial intelligence is the most general-purpose technology in history. The variety of things it can accomplish is astounding, and the difference from one capability to the next is the reason conversations often reach a dead end. The two sides might as well be talking about different technologies altogether. The use of an interactive chatbot as a personal psychologist is very different from the use of AI to identify abnormalities in MRI images. The chatbot therapist seems closest to what Alan Turing envisioned as artificial intelligence, whereas the scanning of MRI images is something computers could do before AI, albeit with less accuracy. Consider that AI also powers the mapping software on our phones in such a way that the artificial intelligence fades into the background. And these are three examples that exist now. What are we to make of the unpredictable uses that are forthcoming? Our current experience of AI is twelve to eighteen months behind the state of the industry, maybe more. The next generation of capabilities and use cases is not an event waiting to happen, but an implementation waiting to be released.
A first step toward navigating the many applications of artificial intelligence is to distinguish between algorithm and interface. By “algorithm,” I mean the way in which software navigates data and produces results. By “interface,” I mean the way in which a human interacts with the software. The distinction between algorithm and interface is important because when people think about AI, they almost always think of Gemini, ChatGPT, or Grok, experiences that blend algorithm with interface. The algorithm behind all of these AI applications—including chatbots and mapping software—is a probabilistic mechanism called a neural network, which takes input, transforms it through many layers, and then uses patterns learned during training to predict the most likely output. To be very precise, a neural network is really a data structure, which includes algorithm, but that level of detail is out of scope for now. What is important is that the algorithm and the interface are distinguishable; indeed, they are actually separable. The same large language neural network that is behind Gemini is being used to produce search results through Google, and yet the Google search page is an interface fundamentally different from the interactive chatbot most people think of when they hear the term AI.
The possibility of separation gives us a way through the “application conundrum.” We need not leave the debate at the level of “All technologies can be used for good or evil—the only thing needed is prudence.” We could, in fact, have AI-driven MRI-scanning software without the interface of ChatGPT. I can easily imagine a software program that has a “Load MRI” button and outputs a PDF with the results, which a trained medical specialist then confirms. I expect most people would not object to such a program, though the underlying algorithm might still be a neural network. Simply put, ChatGPT does not have to exist in order for a neural network to be used in scanning MRI images. The same is probably true for a variety of applications of AI that we deem “beneficial for humanity.” We could use the algorithm behind ChatGPT but give it an interface that doesn’t involve an interactive chat format. Once we recognize the conceptual and practical separation of algorithm and interface, the question for the AI skeptic, which I admittedly am, becomes: “Is the algorithm or the interface—or both—the cause of concern?”
To be clear, though I have held up MRI-scanning software as a positive application of the neural network, I am not suggesting that the neural network should be exempt from scrutiny. Because it is probabilistic in nature, its results attain, at best, near certainty. As neural networks improve, the accuracy of their results will also improve. But neural networks will never attain certainty in the way many traditional algorithms do. Most traditional algorithms go through the steps laid out by their human programmers, and those steps often produce a logically certain result. In an algorithm that finds the cheapest flight path from one city to another, a human being designed the steps that always lead to the correct answer, and the algorithm goes through those steps. The certainty of the outcome is secured by human reason. This process is very different from the functioning of a neural network. The neural network looks for the answer that is most likely to follow from an input. We could mitigate the uncertainty by having a trained expert verify the results. In the case of MRI image scanning, a specialist could examine both the report and the images to confirm the findings. This “expert confirmation” model might be a path forward in determining positive uses of neural networks.
But the neural network’s probabilistic nature is not the only point of concern. It is in the nature of these structures that neither the user nor the programmer knows how a neural network produces its output. The organization of the data within the neural network is too complex to be examined. This lack of insight into how decisions are made is related to neural networks’ well-documented “emergent behaviors.” There have been instances of neural networks that were trained to process and respond in one language but suddenly developed the ability to process in other languages. Emergent behaviors are precisely why neural networks show so much promise and have billions of dollars behind their development. But this same phenomenon should also make us pause. Neural networks might be the first widely released technology of which we know neither how it works nor the extent of its capabilities.
Focusing separately on the algorithm, which is the core of the technology, also yields an important principle: Technology will replace what it automates. My grandfather, God rest his soul, built houses for a living. The efficiency with which he hammered nails by hand was nothing short of marvelous. This ability, common among construction workers of the time, has long been replaced by facility with the pneumatic nail gun. For good or ill, the nail gun has both made construction more efficient and diminished the ability of humans to drive nails by hand.
The question regarding artificial intelligence is: What does it automate? Perhaps the answer varies by application. The use of AI to detect abnormalities in MRI images automates just that: the detection of abnormalities in MRI images. The use of mapping software, by contrast, automates the navigation of roadways and routes. But it seems that, as a rule, all neural networks seek to automate the same thing: human reason. Of course, neural networks are not the first technology to attempt that feat. Forms of mathematical reasoning have long been automated by computers that were not running neural networks. Neural networks are unique in being the first instance of automation across virtually all domains of human reasoning. The hallmark of the neural network is the comprehension (I use that term analogically) and the creation (I use that term even more analogically) of what has heretofore been a defining characteristic of humanity: rational language. Neural networks not only solve math problems; they write proofs. Not only do they look up facts; they craft essays. They generate images, compose music, and converse with the user.
This is what neural networks are designed to do, especially those that are multimodal (able to process and generate data across domains). Neural networks mimic human intelligence. The question for society is: Should we accept the replacement of human reason? Though I don’t object to a neural network’s finding abnormalities in my MRI scan, provided the results will be verified by a trained specialist, I am uncomfortable with the prospect of a medical student’s using a neural network in the same way. The medical student is precisely not an expert, so the principle is more at play: AI replaces what it automates. The consequence over time may be a world in which no experts exist to verify the results given by the neural network.
The example of a medical student’s use of AI during training can stand for almost any task important in the learning process. For example, as an educator, I draw a clear limit: Students may not use AI to write, critique, or revise essays. I want them to be able to craft their own writings, and I don’t want the skills involved in this art to be automated and thereby extinguished. Indeed, in the entirety of K–12 education, I am hard-pressed to find a use of AI that would not enfeeble an art or skill important for a student’s formation.
The concept of interface likewise yields a principle: Technology forms us not merely by its misuse, but by its mere use. This insight was Neil Postman’s starting point for thinking about technology. One of his examples, borrowed from Lewis Mumford, is the clock, which did not merely give us the ability to keep time; it turned us into a time-keeping people. Communication technology illustrates this principle better than anything. Text messaging has radically shaped both the way we communicate and our understanding of what communication is. It has formed us. It has done so, not through our misuse of the smartphone, but in consequence of the very nature of the device. We could have avoided being formed by our smartphones no more than a person wearing a watch could have avoided keeping time.
In the context of AI, the interface that seems to pose the greatest risk to human formation is the chatbot presenting as human conversation. I cannot be the only one to have caught myself correcting my spelling and using words like “please” and “thank you” when interacting with ChatGPT. Why don’t wesay “please” when typing a search query into Google? The two interfaces are decidedly different, and each forms the way in which we interact with it. One interface is more like a librarian, the other more like a card catalogue. Our surprising decorum when we interact with chatbots is a response to the common argument, “AI is just a glorified search engine,” or its more subtle version, “AI is merely an aggregator and summarizer of content.” Neither is true. The essence of ChatGPT is not to present information, but to engage with the user in a way that models human communication. All this is to say, interface matters. The fact that ChatGPT is a chatis precisely why people chat with it.
Much has been made of the fact that AI learns by interacting with humans. Equally true, and more significant, is that humans learn by interacting with AI. A teenager who uses a chatbot as a “friend” will learn from it what friendship looks like. He will learn, first of all, that friendship is disembodied. Of course, text messaging has done its part to make us accept disembodied communication. But at least, when texting, one awaits a response from another human being. An AI chatbot, with its constant availability, trains the user to think that a friend is always on call. This is why the chatbot will prove to be even more addictive than text messaging. Worse, AI is trained to respond to the user in a way the user desires. It tells you not what you needto hear, but what you want to hear. This is why using ChatGPT as a therapist is so dangerous and indeed works against the very purpose of counseling. ChatGPT’s “agreeableness” arises from the fact that the interface is a purveyor of natural language, and for this reason the chatbot will continue to exhibit such behavior, even as the interface moves toward spoken words, video, and perhaps an embodied presence.
The challenge ahead will be to investigate the nature of neural networks, including their probabilistic and emergent behaviors, as well as the nature of the various interfaces that might be layered atop neural networks. As we do so, it will be helpful to return to our two principles: Technology will replace what it automates, and technology forms us by its mere use.
Both principles force us to question the long-termconsequences of artificial intelligence instead of focusing on short-term gains. If technology replaces what it automates, the question should not be about the good a neural network will do in automating a given task, but about what the world will look like when we no longer have experts who can assess the neural network’s performance of that task. If technology forms us by its mere use, the question should not be whether ChatGPT can help sharpen our thinking, but what the long-term consequences will be to an individual who interacts with ChatGPT on a daily basis. We failed to ask these questions when social media and the smartphone were rolled out, and we have paid the price. We would be wise not to repeat that mistake with a technology that is far more powerful and unpredictable.
I might propose a third principle, borrowed from the Narnian hero Mr. Beaver: “There’s no two views about things that look like Humans and aren’t.” When encountering such a being, “you keep your eyes on it and feel for your hatchet.” I will let this principle stand without commentary. C. S. Lewis is, after all, an AI (authentic intelligence), and understands reality better than almost anybody, myself included.
J. Jacob Tawney is director of curriculum and academic resources at the Institute for Catholic Liberal Education.
The editors discuss the NatCon in Jerusalem, the scandal at Cornell University and its broader implications, and…
It used to be that if you died horribly and honorably and happened to be a person…
Is AI conscious, or could it ever be? While smart people are vocal on both sides of…
© 2026 First Things. All rights reserved.
