Why I Have Not (and Will Not) Use AI on This Website

N.B. I wrote this piece several months ago and have had it in drafts ever since. This morning I came across a disturbing article on the prevalence of AI in online media and thought I should go ahead and post it.

Confession time: at ninety posts and counting, covering everything from my rambling thoughts on fuel economy to detailed descriptions of how to measure your car's cooling system performance—I'm not writing these for you. Yes you, reading this post or any of the others here.

Vehicle load diagram, generated 100% by me with no help from AI (we weren't allowed to use it in this course); the closed outline shows the operating envelope of the aircraft with all loads normalized to gravitational force. One unexpected insight in my most recent degree program was how quickly and insidiously the use of AI spread among the students, despite many instructors prohibiting its use on assignments like this one.

I write them for me. My aim was to create a repository of information of the sort that would have been helpful when I was first getting started in aerodynamics, knew next to nothing about it, and had encountered lots of internet and popular myths. I wasted a lot of time understanding aerodynamics incorrectly, and not able to articulate why or even recognize that my ideas about it were not based in reality.
 
That’s one reason I don't run ads on this site or monetize it in any way, and have just stuck to a free blogging platform: I want this material to be freely available to anyone curious about it.
 
I also want it to be correct, and as I've put tremendous effort into an engineering education, including earning a second bachelor's degree in aerospace engineering, I'm being more rigorous about technical specificity and revisiting past testing to correct inaccuracies and mistakes.
 
For these reasons—learning and understanding, precision, and correctness—I have not and will not use AI to write anything here.
 
AI Doesn’t "Know" Anything
 
AI models have been fed vast amounts of digitized material, most of it scraped from across the internet (including from this site, as I discovered recently). As anyone who has been alive since the beginning of the internet can attest, lots and lots of what's out there—especially coinciding with the rise of social media—is imprecise, uninformed, innocently incorrect, or maliciously wrong. This means AI responses to human queries can be and often are incorrect.
 
For example, a few months ago I was thinking about something and decided to test ChatGPT by asking it a simple question: what is the meaning of the opening lines of Shakespeare's Richard III—lines, I should add, that are almost always only half quoted by humans and consequently misinterpreted? This was its response:

Notice that it prints, aligning with modern usage, just the first half of the sentence. But the chatbot at least suggests the continuation and confidently writes "sun" as if that is the line. However, it is not: the earliest surviving print of Richard III (1597) opens,
 
Now is the winter of our difcontent,
Made glorious fummer by this fonne of Yorke:
 
That word, "fonne" (sonne) is deliberately ambiguous. In Shakespeare's other works, he uses "sunne" or "sun" for the Sun, not "sonne" with an o—but "sonne" was an earlier Middle English variant of Sun in addition to being used for "son." Consequently, this word is translated as both "sun" and "son" in modern editions of Richard III, depending on who did the editing. There is no right or wrong way to convert it into modern English. ChatGPT does not report any of this nuance, offer any explanation of the ambiguity in the line, or even mention that it is printed either "sun" or "son" depending on what modern edition you open. If you were relying solely on ChatGPT for information about this line, you would have no idea the depth of what's missing because ChatGPT doesn’t "know" either. That's a problem, and it points to another glaring flaw in how we perceive AI.
 
AI Cannot Think
 
Generative AI chatbots do not think, cannot reason, are not sentient and are not conscious. The large language models (LLM) we interact with—ChatGPT, Claude, Gemini, and all the rest—are token probability calculators. They have been fed vast amounts of digitized information (that's important; if you're curious about a source that does not exist on a computer somewhere, like a handwritten letter or an obscure book that has never been scanned, AI does not know anything about it. I got in an online argument once with someone who was convinced that Brigham Young never wrote a letter I referenced because AI said he hadn't—a letter which very much exists but only in a physical copy in the archives of the LDS Church). AI can produce new content in the form of text, graphics, and even computer code that bases these generated responses on the probability of token order and proximity ("tokens" being, ultimately, numbers; as my computer science professor said in the very first class, "Computers only work with numbers"). It doesn’t actually know anything. It has no judgment, no morals, and no sentience. Because it is complex enough that it can appear to have these things just means it is a very good mimic—and also a dangerous one.
 
Years ago when I was in high school, I found a website that used machine learning to generate musical harmonizations in the style of Bach. A bunch of chorales from Bach's cantatas had been fed into a computer; when a user input a short melody of their own choosing, the program used the same type of token statistics generative AI does today to fill in harmonies. It did a passable job. However, it was missing something glaring. The brilliance of Bach's chorale harmonizations lies not in the intervallic relationships of the notes to one another but in how novel the combinations were at the time, how certain harmonies and moments reflect words in the texts, how they adhere to or break conventions including the doctrine of affekts and the theory of fundamental bass, and how they reflect the workings of the mind that produced them and their cultural context. That program could mimic Bach, and so can any first-year theory student, without any of the staggeringly important details and context that made him Bach, devoid of understanding, innovation, and synthesis.
 
AI models today do the same thing, only with far larger scope. Now instead of just imitating Bach chorales they can mimic all sorts of things, from term papers to code blocks. A disturbingly large number of people apparently use AI as a therapist, sometimes to disastrous ends. "Vibe physics" has become a phenomenon; people with little to no scientific background think they have made breakthroughs that will upend human knowledge because an AI model designed to keep them interacting with it goaded them on toward absurd conclusions. Some people have even disappeared themselves at the behest of AI, convinced that the world was ending and spurred onward by the quasi-religious output of an AI mimic.
 
I've lost track of how many times I have had to correct ChatGPT when it says something inaccurate or incorrect. For example, this little tidbit:

I had asked it at what age the average American begins drinking coffee, and its response had some major errors. Of course, when corrected it gave the familiar, chirpy "well thanks!" response too familiar to so many of us.
 
Chatbot output is always and only a reworded and reorganized (oftentimes incorrectly) version of material that already exists online in some form, run through a bad thesaurus with not even a middle-school understanding of what it means—because, of course, AI cannot understand anything. But here's the thing: words matter. Recently, I came across two people arguing online because one of them had written (in an article on a major automotive site) that aerodynamics "come into effect" at 45 mph, and then defended himself by attributing that statement to Barnard. Since I had the book at hand, I looked it up and then commented to point out that Barnard wrote that aerodynamics "predominate" at 45 mph—which does not mean the same as "comes into effect." The subtly different meaning of word choices inserted by AI models based on their probability and proximity to other words can upend a person's understanding of the concepts those words are meant to represent.
 
AI Replaces Thought
 
This doesn't stop people from trying to use AI to outsource thinking. Someone online came up with a metaphor I wish I could take credit for: using AI to write papers and complete assignments is like taking a forklift to the gym to move the weights around. Getting the weights from one place to another was never the point.
 
The importance of thinking and wrestling and being uncomfortable with things in order to figure them out cannot be understated. In the early days of supersonic flight, testing of the F-102 interceptor showed that it vastly underperformed its expected maximum speed. Richard Whitcomb, a NASA aerodynamicist, was sitting in his office one afternoon with his feet on his desk, mulling over a lecture on supersonic flow he had attended a few days earlier, when suddenly it hit him: if we think of the flow around the aircraft as "pipes" that must bend to follow the shape, as the lecturer had suggested (since, at sonic condition, the "pipes" no longer narrow as freestream speed increases), too much constriction in those "pipes" increases non-lifting wave drag. From this flash of inspiration he derived the Whitcomb Area Rule (the derivative of cross section area along the length of the aircraft should be minimized to minimize wave drag). The F-102 fuselage was modified with a "Coke bottle" shape to account for the increase in area due to the wings, and the aircraft achieved its performance target. Boredom is important. The most impactful human insights have come from people mulling things over, letting their minds wander and wondering, "What if…?" If we turn to AI for a pat answer to every problem and to screens for a constant stream of entertainment, we lose this crucial component of intellectualism.

F-102 development prototypes. Left, before Area Rule modification; right, after (image credit: Wikipedia).

Does this mean AI is useless? Of course not. But it does mean we should be careful. Recall this quote from Oscar Wilde:
 
"Do you think, for instance, that we object to machinery? I tell you, we reverence it; we reverence it when it does its proper work, when it relieves man from ignoble and soulless labor, not when it seeks to do that which is valuable only when wrought by the hands and hearts of men. Let us have no machine-made ornament at all; it is all bad and worthless and ugly. And let us not mistake the means of civilisation for the end of civilisation; steam-engine, telephone and the like, are all wonderful, but remember that their value depends entirely on the noble uses we make of them, on the noble spirit in which we employ them, not on the things themselves" ("Art and the Handicraftsman," emphasis added).
 
As I see it, we are rushing headlong into an AI revolution dominated by ignoble uses. Widescale replacement of human workers with AI, people who believe that knowledge and memory—and, by extensions, thought—are now unnecessary since one can always just ask an AI chatbot, the widespread use of AI to complete school assignments or to explain concepts (often incorrectly), the generation of AI "art": I don't view any of these as noble uses, and they have tremendously negative impacts on our ability to exist as humans both today and into the future.

 
Outsourcing Cognitive Work
 
The thing that bothers me most not just about AI but modern technology in general—although AI exemplifies this problem—is the magnification of an undercurrent that has always driven technological development but in recent years has come to dominate it. Since the Industrial Revolution, new technologies outside of physical automation have tended to transmute cognitive load, morphing it from one set of skills to another perhaps related but distinct set; now, they increasingly remove it entirely and place it in the hands of machines.
 
For example, riding a horse or driving a wagon as one's primary form of transportation took skill and cognitive work. When cars were invented, they didn't remove the role of skill and technical ability on the part of the human operator—they transformed it. Drivers of cars still needed to pay attention to where they were going, implement a series of controlled movements, and adjust their vehicle control based on sensory inputs. If anything, as cars got faster they increasingly required more skill to pilot than wagons, since everything comes at you faster at 65 mph than 10. The new technology of self-propelling machines, still piloted by human operators, didn't remove cognitive load.
 
Starting more than a decade ago, however, we began implementing various advanced driver assistance systems (ADAS): adaptive cruise control, lane monitoring, etc. These spiraled into partial autonomy, starting with a certain company's release of two deceptively-named driving aids. I can walk into dealers for any number of car brands today and buy a car that will "drive" itself down the highway. We nominally require drivers to pay attention but, since they don't physically have to anymore, this is difficult to police and people have found numerous ways to cheat. And what reason do they have to pay attention when the whole appeal of these ADAS is that the car will handle the cognitive load of driving?
 
Rather than transform the cognitive tasks associated with transportation into different skill requirements, our most recent technological developments try to remove them entirely and the act of driving morphs into a completely passive activity. We see this not just in autonomous cars but all other areas of tech. I know people who pull up GPS navigation any time they leave their immediate subdivision, even to go to the grocery store—something that was impossible a quarter century ago. The need to understand and maintain a mental map, or even to read and interpret a paper map and figure out how that overlays real topography (anyone else grow up with Thomas Guides?), has been removed and we simply follow a series of immediate commands to "turn left" or "use the right two lanes to merge," with sometimes disastrous results.
 
AI is the logical extreme of all this. I read a post on a web forum recently by a prolific user who wrote, apparently in complete seriousness, something to the effect of, "Knowledge is obsolete now that we have AI to remember everything for us" (as someone who completed a master's program that required memorization of some of the densest music ever written and was indelibly changed by the experience, this take is laughable to me). AI will spit out code blocks or even entire programs without the prompter needing to understand anything about how the code works. AI will write term papers without a student needing to remember anything about the subject matter (if they even learned it in the first place). There's even an AI agent now designed to interface with Canvas and complete entire courses (watching lectures, taking notes, submitting assignments, working through exams) without the "learner" bothering to "learn" anything.
 
Why is that a problem? First, we're barreling toward an Idiocracy (2006) scenario where a smaller and smaller part of the human population knows and, more importantly, understands things. When AI does all the "learning" for you, you don't actually learn anything, and studies have shown that retention tanks when "learning" is assisted by an AI agent. Second, because more and more of us are becoming uneducated idiots and AI is being pushed by corporate leadership to an extent not remotely justified by its actual capabilities, this problem will only get worse. Third, it makes us not human anymore. We've arrived at the level of technological prowess we have achieved because we used our brains, figured things out, learned things, remembered things, failed at things, made mistakes, worked through it and managed a better understanding. When we stop doing these things because "AI will do it for us," we are no better than food and entertainment receptacles. You can have AI spit out a summary of Plato's Republic and you will think you now understand it, but you won't because you did not have the experience of reading it yourself, mulling over his arguments, considering their structure and nuance, thinking through counterexamples, parsing issues in the translation, and considering the merits of not just Plato's philosophy but your own philosophical model that you have to build in response to digesting his. Learning is never just about knowing and recalling facts.
 
For example, here's an experiment you can do quickly. I'm listening to an afternoon show on NPR right now, and the guest (a blind person who helped develop Siri's voice) just suggested asking the digital assistant the value of the distance from Earth to the Moon divided by the length of the Amazon River. Before you do that, see if you can reason your way to an answer without using any external resource. What is your estimate of the average distance to the Moon? Is it thousands of miles? Hundreds of thousands? How about the length of the Amazon? Is the Amazon the longest river in the world? Second longest? Third longest? What does that tell us about its length if we have to guess it, especially in relation to rivers you may know such as the Mississippi? What order of magnitude difference is there in these two distances? Is the distance to the Moon on the order of ten times longer than the Amazon? One hundred times? In what ballpark range should our answer lie? What is your estimate of the average distance to the Moon? What is your estimate of the length of the Amazon? Finally, what is your answer?
 
Now, ask Siri or Alexa or an AI chatbot. How close did you get? More importantly, what process of reasoning did you use to get to your answer? How might you apply the same type of reasoning to another problem? Did you mess up somewhere, under- or overestimating a distance or making an arithmetic mistake? Why? Just by thinking about this inane question, you had to exercise your mind and reinforce neural pathways used in logic and problem-solving, recall information you may not have thought about for some time, and then use another area of your brain to perform some basic arithmetic. If you didn't try and come up with an answer yourself and instead went straight to asking AI, you got none of those benefits and will be less able to handle another problem in the future. This time it was a stupid question; then, it may be something important.
 
I have spent a great deal of time and expended a lot of effort to expand and improve my understanding of physics, mathematics, science, and engineering over the last several years (and you can too!). I still get things wrong all the time and will forever because I'm human, but as I understand more and more, I correct my previous misconceptions or misunderstandings and grow beyond them (some of you may have noticed that I frequently go back and revise old posts here to try to correct things I got wrong, in fact). Outsourcing any of this to AI would have been far easier but would not have resulted in improving my understanding because that happened as a direct result of my misunderstandings and failures. That's why I refuse to use it on this site and rarely use it in general. It is antithetical to my purposes here. We need failure and confusion in order to grow. As Martin Luther is reputed to have said (and was sort of an unofficial motto at my first alma mater, a small Lutheran university), "Go, and sin boldly."