There is something deeply unsettling happening in the world of AI education right now.
Not a scandal. Not a controversy. Something quieter.
And far more dangerous.
Millions of people are learning how to use AI from people who don’t actually understand how AI works. They’re following rituals. Performing the right motions.
Getting results. Real, usable, sometimes impressive results.
And that’s exactly the problem.
Because getting a result is not the same as getting the best result. And if no one ever taught you the difference, you’ll never know what you’re missing.
But here’s what this post is really about.
It’s not about AI.
It’s about the most valuable thing you own. The one thing no algorithm, no model, and no influencer can replace.
Your ability to think critically.
To question. To test. To say “wait, is that actually true?” and mean it.
That ability is under attack right now. Not by AI itself — but by the way we’re choosing to consume information about it. And if we don’t protect it, we won’t just get worse results from our AI tools. We’ll lose the one cognitive edge that makes us worth anything in an AI-powered world.
I’ve written about this directly before. The idea that critical thinking is the only skill that actually matters now — and what’s happening in the AI education space is the sharpest test of that thesis I’ve ever seen.
This isn’t a post about dunking on AI influencers. It’s about a pattern that has repeated itself throughout human history. A pattern where confident misinformation, dressed up in the costume of expertise, survives for decades. Sometimes centuries.
We’ve been here before. Multiple times.
And every single time, the people holding the “proof” in their hands were the last ones to believe they were wrong.
Part One: The Islands That Built Fake Airports
In 1945, the war ended and the Americans went home.
For the indigenous peoples of the South Pacific islands Vanuatu, Papua New Guinea, the Melanesian archipelago this was a catastrophe. Not because of the violence or the politics. But because for years, these islands had been transformed into military supply lines. American soldiers had arrived with cargo unlike anything these islanders had ever seen. Canned food. Medicine. Radios. Jeeps. Weapons. Clothing.
Extraordinary wealth. Delivered by air.
When the war ended, the airstrips went quiet. The planes stopped coming. The cargo disappeared.
So the islanders did the only rational thing they knew how to do.
They rebuilt the airstrips.
Not with concrete and steel, but with bamboo and packed earth. They carved wooden headsets and held them to their ears. They lit fires along the runways at night to guide planes in. They built wooden control towers and staffed them with men who mimicked the motions of the air traffic controllers they had watched.
They marched in formation with wooden rifles.
They did everything right.
The form was perfect. Every ritual was accounted for. Every behavior replicated.
The planes never came.
These became known as Cargo Cults. One of the most documented anthropological phenomena of the 20th century. And the reason they became famous isn’t because the islanders were foolish. It’s because their logic was completely sound given what they knew. They had observed a cause-and-effect relationship. Build the runway, perform the rituals, receive the cargo.
They just had no visibility into the actual mechanism.
They were working with the output. Not the system.
In 1974, Nobel Prize-winning physicist Richard Feynman stood at the Caltech graduation podium and used this exact story to call out an entire generation of scientists. He called it “Cargo Cult Science” — the practice of performing all the rituals of real science while missing the fundamental integrity that makes science actually work.
He was talking about researchers.
But he could have been talking about your favorite AI influencer.
Part Two: The Doctor Who Was Killing His Patients (And Didn’t Know It)
For more than two thousand years, the most trusted medical minds in the world believed in bloodletting.
The premise was elegant, even scientific-sounding. The human body contained four humors:
- blood
- phlegm
- yellow bile
- black bile
Disease was the result of imbalance. The cure was to drain the excess.
So they drained. Feverish patients. Plague victims. Women in childbirth. Children with infections.
They drained and drained and drained.
And patients did recover. Not all of them. Many died. But enough recovered to confirm the belief. A fever broke. A headache cleared. A patient stabilized. And the physician wrote it down as evidence.
The practice didn’t survive two thousand years because doctors were stupid. It survived because confirmation was everywhere. The successes were documented. The failures were explained away. The patient waited too long, the balance was too far gone, the treatment wasn’t aggressive enough.
The result, whatever it was, always confirmed the theory.
Physicians who questioned the practice were considered dangerous radicals. They were pushing back against two millennia of documented results. The evidence was overwhelming. Who were you to argue?
The terrifying lesson of bloodletting isn’t that bad medicine existed. It’s how long it took to die. Not because the evidence was hidden, but because the framework for interpreting evidence was broken. The doctors weren’t lying. They genuinely believed they were helping. And they had thousands of documented patient cases to prove it.
They just had no control group.
They had never asked: what would have happened without the treatment?
That question no one was asking is the same question that almost no one is asking about AI advice on the internet right now.
Part Three: The Man Who Proved Doctors Were Wrong (And Paid For It)
In 1847, a Hungarian physician named Ignaz Semmelweis noticed something that should have been obvious.
He was working in a maternity ward in Vienna — one of the most prestigious hospitals in the world. The ward had two clinics. In the first clinic, staffed by doctors and medical students, the maternal mortality rate was roughly 10%. In the second clinic, staffed by midwives, the rate was closer to 2%.
Women begged not to be admitted to the first clinic.
Semmelweis was obsessed with why. He studied everything. The position of delivery. The ventilation. The timing. Then a colleague died — after being accidentally cut during an autopsy — and the symptoms matched exactly what the mothers were dying from.
It hit him like a lightning bolt.
The doctors were going directly from performing autopsies on cadavers to delivering babies without washing their hands.
He instituted mandatory handwashing with chlorinated lime solution in his clinic. The mortality rate dropped from 10% to 1%.
He had the data. He published it. He presented it to the medical establishment.
They laughed him out of the room.
The pushback was immediate and vicious. Senior physicians were offended by the implication that they could be the cause of death. The idea that a gentleman physician’s hands could be unclean was considered absurd.
Insulting, even.
And besides, they had results. They had decades of successful deliveries. They had reputations. They had authority.
Semmelweis spent the rest of his life fighting to be heard. The medical establishment refused to change. He grew increasingly erratic from the weight of knowing the truth and being ignored. In 1865, he was committed to a mental asylum.
He died two weeks later. He was 47.
A decade after his death, Louis Pasteur and Joseph Lister proved germ theory. Vindicating everything Semmelweis had said. The medical community finally adopted handwashing.
They just didn’t do it until the man who discovered it was dead.
The “Semmelweis reflex” is now a term in psychology. It describes our instinct to reject new information that contradicts established belief — especially when the people delivering it challenge our authority or identity.
The doctors weren’t evil. They were just protecting a worldview.
And they had the results to prove it.
The Part Nobody Wants to Talk About
Before we get to AI, we need to talk about something uncomfortable.
There’s a 2006 movie called Idiocracy. A comedy about a perfectly average man who wakes up 500 years in the future to find that society has been so thoroughly dumbed down by mass media, anti-intellectualism, and the slow erosion of critical thinking that the least intelligent person alive is now the smartest person on Earth.
It was supposed to be satire.
It’s feeling more like a documentary.
The movie’s core premise isn’t that people got dumber genetically. It’s that they stopped questioning. They accepted what the screen told them. They trusted volume over substance, confidence over accuracy, and familiarity over truth. Critical thinking didn’t get beaten out of them. It got slowly, pleasantly, entertainingly replaced with something easier.
We are living through the attention economy’s greatest achievement: a world where the loudest, most confident voice in any room gets treated as the most credible one. Regardless of whether they actually know what they’re talking about.
And in the AI space, that is genuinely dangerous.
Because the cognitive biases at play here are not new. They are ancient, documented, and deeply human.
Confirmation bias tells us to trust the evidence that confirms what we already believe, and dismiss the evidence that doesn’t. You tried the technique, you got a result, you stopped asking questions.
Authority bias tells us that a person with a large following, a confident delivery, and a polished video must know what they’re talking about. Follower count becomes a substitute for expertise. I’ve written an entire post on how this exact dynamic plays out in the SEO world — where confident, wrong takes about AI search spread faster than the careful, accurate ones.
The Dunning-Kruger effect tells us that the less someone actually knows about a complex system, the more confident they tend to be in their explanation of it. The people who understand AI most deeply are usually the most careful, qualified, and hesitant in their claims. The people who understand it least are often the ones with the most to say. This is exactly what I mean when I write about why so-called “experts” are addicted to complexity. The performance of expertise often fills the void where actual expertise is absent.
These are not character flaws. They are features of human cognition that evolved long before YouTube existed.
But in 2026, in the middle of the fastest-moving technological shift in human history, letting these biases run unchecked is not just intellectually lazy.
It’s costly.
Now Let’s Talk About AI
Here’s what’s actually happening right now.
A person with a large audience (a self-described AI expert, prompt engineer, or automation guru) discovers a feature in ChatGPT, Claude, or some AI platform. They test it. They get a result. The result looks impressive. They film it. They post it. They explain the mechanism based on what they observed, not based on how the system actually works.
And then it spreads.
Thousands, sometimes hundreds of thousands, of people adopt the technique. They try it. They get results. The results look good to them. They become believers. And when someone comes along and says “actually, that’s not how this works”, they hold up their results like a shield.
“Look at what I got. Explain that.”
The bloodletting physicians said the same thing.
Here are three of the most common examples I see repeated constantly and what’s actually happening under the hood:
“I trained the AI on my documents.”
You didn’t train anything. Uploading a file to a chat interface puts that content into the model’s context window. Its short-term working memory for that session. The model is not learning. It is not being retrained. Its weights are not changing. When the session ends, it forgets everything. Completely. The word “training” implies a permanent change to the model. That is not what happened.
“I’m building the AI’s memory.”
Some platforms have memory features. Most of what people call “memory building” is either context window stuffing or a retrieval system that pulls relevant information when prompted. It is not the same as human memory. It does not compound over time the way people believe. And depending on the platform, it may not persist the way you think it does.
“The AI knows my brand now.”
If you’ve uploaded files, written long system prompts, and had several conversations, the AI has access to information about your brand inside those contexts. It doesn’t know your brand. There is no continuous entity sitting there getting to know you between sessions. Every conversation starts fresh unless a system is explicitly designed to inject that prior context.
None of this means the results people are getting are fake. The results are real.
But here is the Semmelweis question no one is asking:
What would the result look like if you understood the actual mechanism?
That’s the control group that doesn’t exist in most AI education content.
And nobody is building the bamboo airstrip and asking themselves why the planes haven’t come yet. Because some planes are coming. Just not the ones that matter most.
And if you’re building workflows, automations, or business systems on top of mechanisms you don’t actually understand, you’re going to pay for it eventually. That’s a point I break down in depth when talking about what happens when everyone has access to the same AI tools. The people who understood the real mechanism will pull ahead, and the people who only knew the ritual will be left holding a bamboo headset wondering what went wrong.
How to Test the Assumptions (Before You Build Your Entire Workflow On Top of Them)
You don’t have to take my word for any of this. Here’s how to pressure-test what you’re being taught:
1. Ask where the claim comes from.
Did the person making the claim build anything with raw AI APIs? Have they worked with model weights, fine-tuning pipelines, or token limits at an engineering level? Or did they observe an output and reverse-engineer an explanation? Observation is not mechanism. Results are not proof of process.
2. Test the opposite.
If an influencer says “uploading your knowledge files trains the AI,” do the same task without uploading the files. Compare the outputs side by side. If the results are nearly identical, the file upload isn’t doing what they claimed. This is your control group. Run it.
3. Start a new session and see what the AI actually remembers.
If someone told you the AI has learned your preferences — close the chat, start fresh, and ask it the same questions. What does it actually retain? This single test will clarify more about AI memory than a hundred YouTube tutorials.
4. Ask the AI to explain what it’s doing.
This isn’t foolproof — AI models can confabulate — but asking the model directly how it’s processing your documents, what it has access to, and what it will remember after the session often surfaces more accurate information than what the influencer told you.
5. Follow people who are building, not just explaining.
There is a difference between someone who uses AI tools and someone who builds them. Builders have visibility into the actual mechanism. They’ve hit the walls. They’ve read the documentation. They’ve worked with the APIs directly. When a builder tells you how something works, they’re speaking from architecture.
When a user tells you, they’re speaking from output. The same principle applies to anyone automating their content pipeline — the people blindly automating their blog posts with AI workflows built on misunderstood mechanics are going to produce a lot of bamboo runways before they figure out why the planes stopped coming.
The goal of every single one of these tests is the same thing.
Protect your ability to think critically.
Because that’s the real asset here. Not the prompt. Not the workflow. Not the perfectly optimized system message.
You asking the right questions is the only thing that separates leveraging AI from being misled by it.
Why I’m Telling You This
I built Magai.
Not as a no-code wrapper slapped together over a weekend. As a serious AI platform that gives people access to every major AI model through a single, thoughtfully designed interface. I’ve worked directly with the APIs. I’ve read the documentation. I’ve hit the limits. Context windows, token counts, memory architecture, retrieval systems, etc.. Because I had to build around them.
That’s not a credential I’m flexing. It’s context for why I can tell you with confidence: a lot of what is being taught about AI right now is the bamboo airstrip. It looks right. The ritual is convincing. The results are real enough to sustain the belief.
But the mechanism is wrong.
And I’d rather you know that now, before you’ve staked something important on a system that was never really there.
The Semmelweis reflex is going to kick in for some people reading this. That’s fine. Question me too. Test everything I’ve said here against what you know, what you’ve experienced, and what you’re willing to go find out for yourself.
That instinct, that friction, is not a bug.
It’s the whole point.
The world doesn’t need more people who are good at following AI tutorials.
It needs more people who are good at thinking.
Don’t outsource that. Not to an influencer. Not to an algorithm.
And not to me.
