AI Myths a Lot of Smart People Still Believe

A friend of mine runs a small agency. Sharp guy. Last month he mentioned, almost proudly, that he'd stopped double-checking anything the AI wrote for him. "It's basically always right now," he said.

A couple of weeks later it handed him a neat little statistic that made its way into a client deck. The client asked where the number came from. It came from nowhere. The model made it up, in the same calm tone it uses for everything.

He's not careless. He just believed a few things about AI that a lot of us quietly believe, and they're the kind of beliefs that don't hurt you until, one day, they do. Here are the ones I'd unlearn first.

"It understands what I'm asking"

This is the big one, and it's completely understandable. The thing writes like a person, so your brain files it as a person. But under the hood it isn't reasoning about your question the way you'd assume. It's predicting what words tend to follow the words you gave it, based on a mountain of text it was trained on.

Most of the time that produces something genuinely useful, because a lot of human writing is correct and sensible. But "predicts plausible text" and "understands and knows the truth" are different jobs that happen to overlap a lot, and the gap between them is exactly where the made-up statistics live. If you want the honest, no-magic version of what's actually going on, I wrote it up in how large language models work.

"The confident answer is the correct one"

We're wired to trust confidence. The person who sounds sure gets believed. AI has, by accident, weaponized this. It delivers a wrong answer with precisely the same steady assurance as a right one. There's no wobble in its voice when it's guessing.

So the instinct to build is almost the opposite of normal: when it sounds most certain about something you can't easily check, that's your cue to slow down, not speed up. I go deeper on that reflex in using AI without losing your own judgment, because it's the single habit that separates people who get burned from people who don't.

"Bigger and newer always means better"

Every few months there's a new model and a wave of "this changes everything" posts. Sometimes the new one really is better. Often, for your actual task, you wouldn't notice the difference in a blind test.

Chasing the latest model is a bit like buying a faster car to sit in the same traffic. The thing holding back your results usually isn't the model's raw power: it's a vague prompt, a task the tool is genuinely bad at, or the fact that you needed a fact and asked a pattern-matcher. Fix those and a "worse" model will beat a better one you're using badly.

"It's neutral and objective"

Because it's software, it feels impartial. It isn't. A model learns from human writing, and human writing is full of our assumptions, blind spots, and skew. The model absorbs all of it. Ask it something where opinions and biases live (hiring, culture, history, anything contested) and you're getting a blurred average of what people have written, not some clean view from nowhere.

That doesn't make it useless there. It makes it a source to weigh, like any other, rather than a referee.

"There's a secret prompt trick that unlocks everything"

You'll see threads promising the "one prompt that 10x's your output." Mostly hype. The genuinely useful part of prompting isn't a magic phrase, it's just communicating clearly: say who it's for, what "good" looks like, show an example, ask for the format you want. That's it. It's the same skill as briefing a colleague well, and it's completely learnable in an afternoon. The practical version is in writing better AI prompts.

"Whatever I paste in just… disappears"

People drop contracts, customer lists, and company code into a chatbot without a second thought, as if the text evaporates after it answers. With a cloud tool, pasting is uploading, your words go to someone else's servers and live under their rules, not yours. Worth a genuine pause before the sensitive stuff. There's a short, non-paranoid guide to this in AI and your privacy, and if you're ever about to share code, a quick sweep with our secret scanner will catch keys and passwords before they leave your machine.

The myth I meet as a builder

The myth I hit most often is "AI-powered means better." I run a tools platform, and our most reliable tools contain no model at all: deterministic code that does the same thing every time is exactly what you want for converting a file or generating a regex. We wrote up how our regex generator works without AI because the transparency is the feature. Meanwhile the places we do use models (in-browser Whisper for speech, one disclosed analysis tool) are the places where pattern recognition genuinely beats hand-written rules. The skill is matching the tool to the job, not maximising the AI content of your software.

The through-line

Notice what all of these have in common. Every myth is really the same mistake wearing a different hat: treating AI as more human, more certain, and more trustworthy than it actually is. It's a spectacular tool. It is not a colleague, an oracle, or a neutral judge, no matter how much it sounds like one.

Keep that straight and you get almost everything good about AI with almost none of the ways it quietly embarrasses you in front of a client.

Sources

• Large language model (Wikipedia)