The threat from AI is not that it will gain self-consciousness or a soul. The threat is that we will destroy ours—that, bewitched by the appearance of AI-consciousness and AI-souls, we will lose interest in cultivating and honoring our own. In doing so, we will cease to produce students capable even of cheating intelligently, let alone of writing coherent papers themselves. At the same time, we will cease to produce teachers capable of telling the difference....
To maximize the chance of not having their jobs replaced by machines, students need to learn how not to think like machines....
AI is not going to make human intelligence obsolete; on the contrary, it is going to make human intelligence more necessary than ever. But human beings might well let AI make them stupid enough to believe that their own intelligence is obsolete.
I am much more concerned about the decline of thinking people than I am about the rise of thinking machines.
The 20th-century computer scientist Edsger W. Dijkstra famously compared the question of "whether computers can think" to that of "whether submarines can swim." I find this analogy excellent: it does not deny that the machine can accomplish certain tasks more effectively than an unassisted human, nor that humans must learn to use the machine in order to succeed in a modern economy; but it also makes clear that the machine does not replace human ability. If your submarine breaks down in the middle of the water, you better hope you still know how to swim.
Today's computers are much more capable than those of Dijkstra's era, but his point remains equally valid. If this distinction between human thinking and "AI" eludes many people, perhaps it is because most people do very little thinking anyway.
[AI] is only using statistics and pattern recognition to generate predictable responses to my inputs. But isn't that what many people are doing? We go through our days responding to prompts, like robots, and much of what we say and do is as banal and formulaic as AI slop.
I recall observing, well before AI was a thing, that most of the work students produced for the "creative writing" workshops I attended could easily have been generated by a computer program. People weren't writing stories and poems so much as they were generating material that sounded like what a story or poem is supposed to sound like. If they had anything truly fresh to communicate, they were doing a very fine job of hiding it....
With AI cranking out predictable content free of charge and at lightning speed... I can't get away with thinking and writing like a robot. I must write like a real person. I must do the fun but difficult work of conducting fresh research, making my own connections, drawing my own conclusions, and articulating them in my own words. In 2022, I could get away with the occasional cliché or warmed-over observation. In 2025, that's an open invitation for the robots to eat me for dinner.
"AI"s are certainly artificial, having been made by human hands. But they are not intelligent. To call them "artificial intelligence" is to accept, not just a fiction, but a lie. It is to misconstrue both the nature of machines and of man. It is to give in to the ways in which chatbots threaten to atrophy our humanity, and, in extreme cases, even drive us to madness.
In lieu of "artificial intelligence," I propose a more accurate, ethical, and socially responsible name: "pattern engine."
There is, however, something qualitatively different about AI from the computers of yesteryear. In fact, with hindsight the computers Dijkstra was speaking about were less like submarines, and more like flippers and SCUBA tanks: rather than doing the swimming for us, they helped us to do our own swimming better. Today's AI models are the submarines, allowing us to travel leagues and leagues under the sea without ever touching the water or even learning to swim in the first place. Moving onto land, an analogous contrast is between a bicycle and a locomotive (or an automobile).
In 1981... Steve Jobs... read an article in Scientific American that compared the efficiency of locomotion across species.... Humans ranked about a third of the way down... But then someone had the insight to test a human on a bicycle, and the cyclist blew the condor away.... The computer, he said, was "a 21st century bicycle" for the mind....
With the launch of ChatGPT, gushed Microsoft CEO Satya Nadella in early 2023, "We went from the bicycle to the steam engine."... Bicycles and trains are both technologies that move us from place to place.... But the comparison falls apart when you consider their effects on the traveler. In terms of effort, a steam engine doesn't really "amplify an inherent ability." It replaces it. You sit back and the coal does the work....
If output is your only metric, then the steam engine really is just a better bicycle. Both get you from A to B. One gets you there faster with less effort. Case closed. The fact that you arrive having done nothing, learned nothing, built nothing—that's not a bug, that's the point. Effort is a cost to be minimized, not a value to be preserved.... [W]hat the output-only frame makes invisible... is what happens to the cyclist along the way. The increase in muscle strength, in cardiovascular fitness. The knowledge of the landscape. And, crucially, the capacity to do it again tomorrow, alone, if necessary.
The cognitive downsides of AI dependency are significant and urgent. The primary concern is that overuse will erode our agency and, therefore, our humanity.... it should go without saying that if a task essentially involves human nuance, judgment, or relationship-building, that's probably not a task we should be outsourcing to AI.
Every time we're inclined to use AI for something, we should ask ourselves whether doing so could potentially erode the mental muscles that make us who we are. For example, since engaging with other people's ideas is part of what I do for a living, I shouldn't be relying on AI summaries of books and journal articles. I could very easily miss something important, and my interpretative and analytical muscles could begin to atrophy. On the other hand, scanning through a dozen PDFs to find a reference isn't an essential skill for me; therefore, I should feel free to let AI do it instead.
The distinction between human and computer, and what we lose by outsourcing our thinking to AI, is especially stark in mathematics, because mathematics is the science of logical thinking. For two decades I have taught mathematics students who struggled to fake their way through math problems by searching the Internet for similar questions, throwing together words and equations that seem related, copying a number from the back of the book, and confidently drawing a box around it. But only now do I have a name for what they have been doing: behaving like an AI.
If we humans are to remain employable in the age of AI — and, more importantly, if we are to retain our humanity — we must learn how not to behave like an AI, how to behave instead like a human. This, then, should be the primary goal of a university education. But in fact, it always has been the goal of a university education. The only difference now is that we have a name for it, and an increased urgency to achieve it.
An old professor of mine, in my freshman year, once said something wise and important to a seminar I was in when one of my classmates observed that "I know what I think, I just can't get the words down on the page." My teacher responded: "Well, you don't actually know what you think, then. The act of writing the thing is the same thing as the thinking of it. If you can't write it, you haven't actually thought it."
Which is to say that writing and reasoning are effectively identical activities — and for many years now, writing has been the way we have taught young people how to reason....
Generative AI thus presents a double threat: first it tempts you by offering to unburden you of your need to reason, the tedium of organising your thoughts. But, almost as bad, it also scrapes the brain of human civilisation of all information and all learning ever and then, apparently, reduces it all to unvariegated mush, spitting out to you its tawdry imitation of thinking — insipid, composite accounts that I would say are written at a 10th-grade level but for their uncanny alien quality that is not quite like anything human at all.
The worry is that we, as a society, will become innumerate, not just illiterate. A.I. appears to be exacerbating an alarming trend in which our basic education is failing our young citizens. And that crisis is aimed at the most basic elements of that education: reading, writing and arithmetic.
My advice to young students today is to study language and mathematics. When you talk to a chatbot, you're using everyday language to talk to a mathematical system that, in turn, talks back to you. Technical skills won't be enough to deal with the unpredictable results in markets, so a broad-based knowledge of math and language will be the only way to adapt. And while jobs might disappear in one sector, we will always need humans who can make sense of A.I.
For all of these reasons, I design all the courses I teach to inspire, guide, and support students in learning how to behave like a human rather than an AI, in the context of mathematics.
Below I have collected some additional recommended readings on the subject of AI, with selected quotes from each.
Observing how these AI models often miss cultural contexts, overlook local knowledge, and frequently misalign with their target community has brought home to me just how much they encode existing biases and exclude marginalised knowledge.... The problem is far deeper than gaps in training data. By design, LLMs also tend to reproduce and reinforce the most statistically prevalent ideas, creating a feedback loop that narrows the scope of accessible human knowledge....
As AI-generated content has started to fill the internet, it adds another layer of amplification to ideas that are already popular online. The internet, as the primary source of knowledge for AI models, becomes recursively influenced by the very outputs those models generate.... As LLMs are trained on data shaped by previous AI outputs, underrepresented knowledge can become less visible.... Over time, this would result in a degenerative narrowing of the public knowledge base, driven not by censorship but convenience and algorithms.
The biggest problem with AI companions is not that they're sycophants. It's that the interactions they manufacture are one-sided.
As human beings, one of our fundamental motives is to matter. Mattering is not just about feeling valued by others — it's also about feeling that we add valueto others. We need to know that our actions make a difference. Extensive research shows that a sense of contribution is vital to our happiness, health, and success....
In healthy relationships, we give as much as we receive. In AI exchanges, we can receive endless streams of information and affirmation, but we have nothing to give back. No matter how good large language models become at simulating care, they'll never substitute for real relationships, because they have no needs to care for.
...as popular as vibecoding has become, it remains evident at all levels that the greater the user's experience and wisdom with code, the better the results will be. This is as true as ever in the age of AI. LLMs, left to their own devices, frequently take overly complex, inefficient, or ugly approaches to new problems, sometimes in subtle ways that are quite difficult to detect and debug. As more software initiatives use vibe-coded components, the demand for humans with the deepest knowledge of code will likely grow. Those are the engineers who can explain the problems clearly, plan effective strategies, break down tasks into components that the AI can execute cleanly, and verify, debug, and improve the results....
I don't think it will ever go out of style to deeply understand what you are doing, and why. We don't all need to be experts in, say, ARM assembly language to write code... of course, but we should know something about what's going on under the hood and how it influences performance. We don't all need to be experts in the fine points of the syntax of our favorite programming language, but we should be able to know when an AI has produced code that is misguided, hard to maintain, or just plain unsafe.
Authentic expression helps us understand truth. Not objective, capital-"T" truth, but subjective truth—the kind that reflects a person's interiority. Generated by systems lacking consciousness or lived experience, AI content, by definition, cannot be authentic.
Why should we care? Because authenticity is foundational to trust, the thread that ties human relationships together and allows us to work toward shared goals. Without the assumption of authenticity in communication, the small amount of trust that remains in our institutions is at risk of unraveling further, making everyday experience more clinical, cynical, and fake.
By punishing an AI when it does something we don't like, and rewarding it when it does something we do like, we are providing AIs with a set of data points about which actions are good or bad.... this kind of learning runs into another famous philosophical problem, "the underdetermination of theory by data." .... We can give an AI a billion cases of moral and immoral action, but AIs can learn practically any lesson from all of this training....
... As any parent knows, young kids are little psychopaths.... At a certain point, though, all the moral instruction kind of clicks, and they (for the most part) become decent people. What the underdetermination of theory by data teaches us is that this "click" is not the kind of thing that emerges mechanistically from reward and punishment. Rather, it's a product of the way this training works on human moral psychology.... Adult psychopaths lack this capacity, and thus never learn the right lesson.
AIs do not have a human moral psychology... because they don't have human brains.... so we have no reason at all to expect that AIs will—or can—learn the kinds of lessons that we hope for them to learn through alignment training. A psychopath cannot learn to care about others through a process of reward and punishment. And we have every reason to think that AIs are psychopaths, or perhaps something far more alien and far less disposed to human sympathy.
AI alignment is not something that works in theory but is difficult to put into practice. It's something that doesn't work in theory, and yet AI companies have decided to give it the old college try. I'm not opposed to trying—maybe the theory is wrong. But we're relying on the success of a theoretically impossible endeavor because the AI labs have already resolved to [build AIs with the capacity to kill us all]. So AI alignment has to work... or else we're doomed.
This is cartoonishly reckless.