AI Humans and the Real Limits of Machines

I read recently in The Atlantic Monthly that artificial intelligence firms are hiring and paying big salaries to philosophers to help them think through the issues facing the developers of AI.

They can send me a check, but I’ll give them some advice for free.

Limitations little discussed

First, anybody doing anything with the internet or artificial intelligence should read and discuss Hubert Dreyfus’s book On the Internet and Ragnar Fjelland’s article “Why general artificial intelligence will not be realized”.

Why those two? It is not, as one might assume, because either man is an “AI skeptic” or because either man warns against the dangers of AI. There are plenty of warnings from others. No, these works are useful because they describe in cogent ways the limitations of AI—limitations that will persist no matter how the technology develops.

What limitations would those be?

One obvious one is embodiment. AI does not have a body. AI cannot give you a hug if you need one. AI cannot breastfeed a baby while gazing into her eyes with love. AI does not have emotions, so AI will not be able to understand beings who do.

The heart of Dreyfus and Fjelland’s argument is that AI lacks the ability that comes with experience. It cannot gain the skills needed for the development of the kind of tacit knowledge that characterizes those who have mastered a practice. Fjelland writes:

An important part of expertise is tacit. The problem facing the development of expert systems, that is, systems that enable a computer to simulate expert performance (for example medical diagnostics) is that an important part of the expert knowledge is tacit. If experts try to articulate the knowledge they apply in their performance, they normally regress to a lower level. Therefore, according to Hubert and Stuart Dreyfus, expert systems are not able to capture the skills of an expert performer.

Put another way, experts can rarely reduce what makes them an expert into “rules” or algorithms, so it will be impossible to input this kind of expertise into machines.

Machines can simulate expertise to human beings who lack expertise, just as con men do all the time. As author Kelsey Piper notes in an article titled “The literary world is sleepwalking into an AI disaster,” the “prestigious” literary organization Commonwealth Foundation recently announced the regional winners of their 2026 Commonwealth Short Story Prize. Unfortunately, readers began suspecting that several of the winning stories were written by AI.

What tipped them off? Or, rather, “What tipped them off that didn’t tip off the readers at the Commonwealth Foundation?” According to Ms. Piper:

AI chatbots love metaphors and similes, and they often spit out ones that sound vaguely pleasing but are logically incoherent or ascribe properties to things that don’t make sense.” Thus, “The Serpent in the Grove” gave us, “The girl smiled like sunrise over a sink.” “The Bastion’s Shadow” says, “She carried it now in her bag, heavy as a charm.” “Mehendi Nights” describes something as “swaying against plaster like a warning bell.”

Another characteristic, it seems, is a “fondness for conjunctions that combine one thing that is abstract and/or incorporeal with another thing that is concrete and/or sensory.” Examples include passages such as “collect your griefs like stones in your pockets,” “connections between sorrow and the taste of metal,” “its presence a quiet pulse against her thigh.” Ms. Piper notes in “The Bastion’s Shadow” the phrase “stone had begun to carry the stories it couldn’t release,” and, in “The Serpent in the Grove,” these two descriptions: “A fact that felt like a small warm animal in her hands” and “her laugh is bright as zinc.”

“It’s not bad writing, necessarily,” says Ms. Piper gracefully. Perhaps. But one suspects the real issue is not merely that the prose was produced by AI, but that the prose produced by AI fooled the judges when others so easily sussed it out. When the standards for good writing, good art, and good architecture become as thin and untethered to reality as they currently have become, it will be as difficult for certain groups to distinguish between AI productions and actual human creations as it was in the Old West for people with no knowledge of medicine to distinguish between the snake oil salesman and the real doctor. Especially since the one promises so much more, so much more easily gained, than the other.

But consider now the example of a repairman, Mark, who used to fix things in my house—anything from plumbing to electrical to household appliances. It is not merely that an AI does not have a body or hands to do this work, or that the AI wouldn’t be able to carry on an interesting conversation with me about theology while it works (Mark’s father was a famous Calvinist theologian). It is this: I can’t tell you how many times Mark would have the wall opened up and exclaim: “Why would someone have done this?” Then, after about ten minutes, he would say: “Ohh, I get it.” And then he would explain to me the “workaround” some previous repairman did, and how he could either completely replace it or make another workaround that some other repairman would need to decipher decades hence. In many older houses, there are layers of workarounds needing to be deciphered before anything can be done.

For more examples, see Michael Crawford’s examples about motorcycle repair in Shop Class as Soul Craft. The tacit knowledge and judgment of the expert cannot be translated into algorithms. AI can help provide and sort through masses of information to help inform judgment, but it cannot make a wise judgment.

Here is my free advice to AI developers and users. The trick to AI is figuring out what it can do and do well and what it simply can’t do. AI can provide the resources to help people make better judgments. But since AI can’t really make wise judgments—especially the kind that depend on emotional connection and a sense of what other people need or the limits they can tolerate—don’t try to make AI do that or depend on it to do that.

AI will eliminate what is already artificial

To be honest, though, I’m less interested in AI than I am in humans. I am not of the camp that says “AI is coming; you cannot stop it.” I suppose, like nuclear bombs, we could restrict it if we chose to do so. That would probably be beneficial in some ways, and unwise in others.

What I want to warn about is not the “AI invasion” as much as the invasion that has already taken root. In one sense, it took root when Adam and Eve made a certain choice in the Garden. But that choice has taken on new forms with every advance in technology.

What I want to warn about is not specifically the “inhumanity” of AI. I want to warn about the inhumanity of humans dealing with other humans, especially in the workings of modern bureaucracy. Will machines take people’s jobs? If you write email messages with meaningless bureaucratic verbiage that sound like a machine wrote it then you will be replaced by a machine. If you are a “human resources” person who cares little about the actual humans you are meant to serve—if they are simply numbers on a spreadsheet to you, and you can never make sensible judgments about “special cases”—then you will be replaced by a machine. If you are acting like a machine; if you are a bureaucrat who acts without emotion, without wisdom and judgment, and carries out rules like an algorithm; if a machine can do what you do the way you are doing it; then you will be replaced by a machine.

In an episode of the television series Westworld, a man gets a call from a company where he has applied for a job. The voice on the phone goes through the normal patter about “Thanks, but we’ve decided to go with someone else.” The voice responds kindly to a few of his questions. But then, at a certain point, he senses something and asks: “I’m sorry, I don’t mean to be rude, but are you a human being?” The call hangs up.

Was it a robot or not? We never know. Would it have mattered? If robots replaced administrators, would anyone notice? A colleague suggests we get ChatGPT to write our assessment reports. Would they notice?

Since so much writing produced these days is robotic, much of it could be done by a robot. That’s sad, but when your educational system spends all its time training dutiful bureaucrats on spreadsheets and “business writing” and little or no time reading beautiful English prose, then this is the result: people without words mouthing gibberish that is largely empty of thought.

The great parody of the ChatGPT culture can be found in a recent episode of the sophomoric comedy South Park. Some students are using ChatGPT to write their papers. When their teacher finds out, instead of being angry, he is ecstatic because now he can get ChatGPT to write the comments on his students’ papers, so he doesn’t have to read them.

A robot is writing the papers, a robot is commenting on them, and no humans are involved. So why do it? Wouldn’t it be better not to perpetuate the illusion of education? And yet schools indulge this illusion all the time. So I suppose ChatGPT is the appropriate tool for the culture we’ve created.

Also in this episode, young Wendy complains to her boyfriend Stan that he doesn’t respond to her texts quickly and thoughtfully, in the way one of her friend’s boyfriends does. When Stan checks with this young man, “How do you do this?” he is told, “ChatGPT.” He just inputs the text into the system and posts the response. Stan is reticent at first, but he can’t resist the social pressure.

Soon, he is posting ChatGPT responses without even looking at his phone. Unfortunately, Wendy has been sharing serious things with him in texts he hasn’t read. ChatGPT has been responding, and since it is designed to please the person inputting the information, she has been getting very sympathetic messages about her problem, just not from him. He has no idea what she has been sharing, but he can’t let her know.

This is the odd, unspoken reality about ChatGPT: its role must remain hidden for it to work. The illusion of the human element must be preserved, or people are offended.

ChatGPT is a sad illusion of human communication that connects no one, just as modern social media produces an illusion of reasoned argument. Mutual brainless denunciation is not an argument. But given these recent developments, I am tempted to propose that we let ChatGPT produce all our reports, write and respond to most of our emails, and argue with other ChatGPT robots on social media. Thus liberated from foolish wastes of our time, we humans can have lunch together and talk about real things that matter.

Be that as it may, people who want to succeed in the coming AI-infused machine age will need to up their game, especially if they are members of that bloated class of bureaucratic “managers” and “administrators.” They will need to master skills—human skills—that machines can’t. Things like mastering the warmth of human connection; the ability to learn from others and from experience, and to respect traditions of expertise; the wisdom to deal with a host of people, each of whom deserves respect for their intelligence, their hopes, dreams, and goals. AI cannot develop virtues. Humans without them will have little of value to add.

When AI invades, we had better learn to use it well. And using it well will require that we become less “machine-like” and more, not less, human.