When AI sounds sure — three questions that teach checking

When AI answers, it often sounds finished.

The sentences are smooth. The tone is calm. There is no hesitation in the punctuation. For a child — and for a tired adult — that polish can feel like certainty.

It is not always certainty. Sometimes it is a confident guess.

Parents do not need a lecture about “never trust robots.” They need a few questions that teach checking without turning the kitchen into a courtroom.

Why “sure” is not the same as “checked”

OpenAI’s research on why language models hallucinate argues that standard training and evaluation often reward guessing over saying “I don’t know.” The result can be a fluent answer that is wrong — delivered with the same polish as a careful one. Read their overview here: Why language models hallucinate.

NIST’s generative AI profile uses the word confabulation for the same problem: confidently stated but erroneous or false content that can mislead people. The profile is public: NIST AI 600-1.

You do not need the papers memorised. The kitchen version is enough: sounding sure is a style, not a proof.

That matters at homework time because kids are trained — by school and by screens — to prefer the answer that looks done. A confident wrong date and a careful right date can wear the same font.

Three questions that teach checking

When the screen looks finished, try one of these — calmly, once, then listen.

1. “Where did that come from?” You are asking for a source, a page, a lesson, or “the chatbot guessed.” Naming the origin is already a check.

2. “What would prove this wrong?” This slows the “finished” feeling. A good answer can survive a challenge. A guess often cannot.

3. “Can we find a second place that says the same thing?” Textbook, teacher notes, a reputable site, or another trusted adult. Two sources beat one smooth paragraph.

Rotate the questions. You do not need all three every night. One good question, asked like a teammate, beats a speech.

If they shrug, stay light. Offer a model: “Let’s check the chapter heading together.” Then stop. The habit grows from small reps, not from a nightly audit.

How to use them without fearmongering

Keep the tone ordinary. You are teaching a habit, not recruiting for a panic.

• Start with your own example if it helps: “I almost believed a tidy answer until I checked the date.”
• Praise the check, not the perfect score. “Nice — you looked again” matters more than “gotcha.”
• If the topic is health, money, or safety, checking is non-negotiable. For low-stakes homework, one second look still builds muscle.

Grandparents, fathers, and older siblings can ask the same three lines. No specialist vocabulary required. The car ride home works as well as the kitchen table.

What you are modelling is not distrust of every tool. It is respect for evidence. That is a life skill with or without chatbots.

What this builds over a month

After a few weeks, children learn the family stance: tools are welcome, and judgement is expected. That is closer to school and work than a blanket ban or a silent wink.

You are not trying to make them afraid of AI. You are trying to make “sounds finished” feel incomplete until someone has checked.

Better questions today. The certainty on the screen does less of the parenting — and your three questions do more of it.