How do I know when AI is telling my child something false?
You often cannot tell by reading it, and neither can they. That is precisely the problem. AI writes the most likely-sounding answer, so a wrong answer sounds exactly as sure as a right one. The fix is a habit, not a detector: flag the specific claim, find a real source you can actually open, and compare. And you cannot use AI to fact-check AI.
Free to read. No account, nothing to sign up for.
What we teach: how sure an answer sounds tells you nothing about whether it is true.
Why you cannot spot it by reading
Every cue people normally use to judge whether writing is reliable (fluency, confidence, specificity, a citation at the end) is a cue these systems produce effortlessly and produce identically whether the content is true or invented. The usual instincts are not weakened here, they are actively misleading.
The reason sits in the mechanism. The system is producing text that fits the pattern of a good answer. Where it has solid material, the most plausible-sounding answer is usually the right one. Where it does not, the most plausible-sounding answer is a fabrication in the same confident register. Nothing in the output marks the boundary, because the system is not tracking one.
This is why "it sounded really sure" is worth teaching a child to treat as no evidence at all. How certain an answer sounds tells you about the writing, not the world.
The habit that works
Checking a whole answer is too vague to actually do, so nobody does it. Checking one claim is a concrete task that takes a minute.
Pull out the specific checkable thing: a date, a number, a quote, a name, a claim about what some study found. Then find a source you can actually open and read, and compare. Not a search result summary: the source.
If it does not check out, the useful next step is to notice what kind of claim it was, because the failures cluster. A child who has been burned twice on invented statistics starts treating statistics differently, which is a far more durable skill than any single correction.
You cannot use AI to check AI
The most common mistake we see is asking the same system, or another one, whether the answer was right. It will answer confidently again, often agreeing with itself, occasionally reversing under nothing more than the social pressure of being questioned.
Asking "are you sure?" is not verification. It frequently produces an apology and a different answer, which feels like a correction and is simply a second guess.
The three kinds of claim most likely to be wrong
Specific numbers and statistics. Precision is exactly what these systems are good at imitating, so "roughly 34 percent of students" arrives with no more hesitation than a figure it actually has.
Citations, quotes and sources. Plausible-looking references to papers, books and articles that do not exist are a well-documented failure, and they are unusually damaging because a citation is the thing a reader checks least.
Anything recent. A model has a cutoff, and questions about the last few months are where it is most likely to fill a gap confidently rather than say it does not know.
IF YOU WANT TO GO FURTHER
This answer comes from the program we teach: sixteen video lessons that take a student from “what even is this” to directing AI on purpose. Parents get their own access alongside their child, because it is hard to guide something you have not seen.
The first lesson is free to watch, no account needed. Start there before you decide anything.