Most of what gets repeated about AI is half-right. That's what makes it dangerous: the believable half carries the wrong half along with it. The result is a population of confident users who trust these tools in precisely the situations where they should be skeptical — and distrust them where they're actually useful.
So here are seven claims you've almost certainly heard, each one steelmanned before it's taken apart. The goal isn't to be cynical about AI. It's to be accurate about it, because accuracy is the only thing that tells you when to lean on it and when to double-check.
1. "AI understands what it's saying."
The case for it is strong. You ask a nuanced question, you get a nuanced, on-topic answer. It tracks context, catches your tone, fixes your logic. If that isn't understanding, what is?
But fluency and understanding aren't the same thing. A large language model — the technology behind chatbots — is trained to predict the next word in a sequence, over and over, across an enormous amount of text. What it has captured is the statistical structure of language: which ideas tend to follow which. That structure is real and genuinely useful, which is why the output is so good. But there's no grounded model of the world underneath it — no checking its sentences against reality, only against patterns. It can describe a chemistry experiment it would have no way of actually predicting. Concede the point: the patterns it learned encode a lot of real knowledge. Just don't mistake a convincing description for comprehension.
2. "It looks things up before it answers."
Reasonable assumption. The answers are specific, factual, and detailed — surely it's retrieving them from somewhere.
Usually it isn't. A standard model isn't querying a database when it replies; it's generating text from the patterns baked into it during training. That's the single most important thing to internalize, because it explains the most notorious failure mode: when the most statistically plausible continuation isn't the true one, the model produces a confident, fluent, completely fabricated answer — a "hallucination." It isn't lying or malfunctioning. It's doing exactly what it always does: predicting likely words. (Tools that do look things up — search-connected assistants, or systems that read your own documents — exist, and they're far more trustworthy on facts. But that grounding is an add-on, not the default.)
3. "If it sounds confident, it's probably right."
We use this heuristic on humans for good reason — confidence often tracks competence, because people who know a subject tend to sound surer about it.
That correlation breaks completely with AI. A language model's fluency is constant; it sounds exactly as smooth and assured when it's right as when it's inventing a fake citation. There's no tremor in its voice when it's guessing. So the one signal you've relied on your whole life to gauge reliability — does this person sound like they know? — carries zero information here. Tone tells you nothing. Verification tells you everything.
4. "A machine can't be biased — it's just math."
The intuition feels almost definitional. Bias is a human flaw; an algorithm has no feelings, no agenda. How could math be prejudiced?
But the model didn't learn from math. It learned from text — books, websites, forums, comment sections — written by people, carrying every pattern those people carry, including the ugly ones. If a stereotype is statistically common in the training data, the model absorbs it as just another pattern to reproduce. The math is neutral; the data is not, and the model faithfully mirrors whatever it was fed. "It's just math" doesn't launder the bias out — it bakes it in and gives it a veneer of objectivity, which is arguably worse.
5. "A bigger model is always smarter."
There's truth here, and it drove the last several years of progress: scaling up the size of models and the data they train on really did produce large, sometimes startling jumps in capability. Bet on "bigger" and you were right for a long time.
But "always" is doing too much work. Returns diminish, and size isn't the only lever — often not even the main one. A smaller model fine-tuned on high-quality, relevant examples can outperform a giant general one on a specific task. Connecting a modest model to the right tools or your own documents beats raw scale for most real work. Treating parameter count as a wisdom score leads you to overpay for capability you don't need and to ignore the cheaper levers — data quality, retrieval, fine-tuning — that actually move the needle.
6. "It's coming for your whole job."
Don't wave this away — the disruption is real, and pretending otherwise is its own kind of denial. These tools genuinely automate work that used to require a person, and that has consequences for employment worth taking seriously.
But a job is a bundle of tasks, and AI is good at some of them, useless at others, and actively risky at a few. The pattern that keeps showing up isn't wholesale replacement; it's reshuffling — the tool absorbs the repetitive middle, and the human work shifts toward judgment, taste, verification, and deciding what's worth doing at all. "Whole job" is the wrong unit. "Which tasks" is the right one, and it's a far more useful question to ask about your own.
7. "You need to be a programmer or a 'prompt engineer' to use it well."
You can see why people believe it. There's a whole genre of secret-syntax tips, magic phrases, and elaborate prompt templates promising dramatically better results.
Most of that is theater. The single biggest driver of a good result is mundane: clear, specific, plain-language instructions — saying what you want, who it's for, what to include, what to avoid, and showing one example. That's not engineering; it's just good communication, the same skill that makes you good at briefing a colleague. The mystique mostly sells courses. The actual skill is one you already have and can sharpen for free.
Where to start
If you keep only one of these, keep #2: the model generates, it doesn't retrieve. Once that's lodged in your head, the rest follow — why it hallucinates, why confidence means nothing, why grounding it in real sources is the upgrade that matters. The fix in practice is to stop asking a model to recall facts from thin air and start pointing it at material you actually trust. In JustJot.ai, that's the whole idea behind asking questions against your own notes — the AI answers from what you've genuinely saved and can check, not from the statistical fog. Same tool, far better question.