The Three Questions Framework: The Entire Verification Skill in One Tool
Three questions — How do they know? Can I verify? What if it's wrong? — are the entire AI verification skill in portable form. Here's how to teach them and why they work at any age.
The Three Questions Framework: The Entire Verification Skill in One Tool
A parent in a homeschool forum described her breakthrough moment like this: her daughter had been using AI for research, and they'd tried several approaches to build better habits around it. None of them stuck until the day the parent stopped explaining and started asking. Her daughter was reading an AI-generated summary about the American Revolution, and instead of reviewing the content with her, the parent asked one question: "How do you think the AI knows that?"
Her daughter paused, looked at the screen again, and said, "I don't know. It just said it."
That pause is the whole game. Everything else is building the habit of creating that pause reliably, across every interaction, without a parent in the room to prompt it.
This post introduces the Three Questions Framework, the simplest tool we've found for building that pause into a child's natural response to AI output. It isn't a checklist or a worksheet. It's three questions that, practiced consistently, collapse into a single automatic reflex: the habit of treating AI outputs as claims that need to be evaluated rather than facts that can be accepted.
Why a Framework
The research we've covered in this series points to a consistent gap. Students know, in the abstract, that AI can be wrong. They've been told this. Most of them could tell you that AI "sometimes makes mistakes." But abstract awareness doesn't activate verification behavior. The student who knows AI can be wrong and still submits its output unverified isn't being careless — they just haven't connected the abstract knowledge to a concrete action they take every time.
A framework solves this by giving the abstract knowledge a procedural shape. Instead of relying on general skepticism to activate at the right moment, you have a specific set of questions that run automatically whenever AI output appears. The questions prompt the behavior directly, without requiring the student to first remember that AI can fail and then decide to do something about it.
The Three Questions Framework is also designed to be genuinely portable. It works for a seven-year-old asking Siri about dinosaurs and for a fifteen-year-old writing a research paper. It works across subjects, across AI tools, and across the range of stakes from low-consequence curiosity to high-stakes academic work. That portability is the point — a tool that requires age-specific versions or subject-specific adaptations is a tool that gets used inconsistently.
The Three Questions
Question 1: How do they know?
Before accepting any AI claim, the first question is: what is this AI's actual basis for knowing this? Not "is the AI usually reliable" in general, but specifically: for this claim, about this topic, what access does the AI have to the underlying truth?
This question matters because AI language models don't retrieve information the way a search engine does. They generate text based on patterns learned during training. When an AI states that a historical event happened in a particular year, or that a scientific study found a particular result, or that a public figure said a particular thing, it isn't drawing on a verified database of facts. It's generating text that fits the pattern of what a correct answer to that question would look like.
For stable, well-documented topics, this process often produces accurate output. But the model has no mechanism to distinguish between topics where its training data was reliable and topics where it wasn't. Its tone is equally confident either way. The question "how do they know?" is the tool that introduces that distinction.
With young children, this question is most naturally introduced as curiosity: where did the robot get that from? What did it read to learn that? With older students, it becomes more analytical: what kind of information would the AI have needed to produce this claim accurately, and is that the kind of information AI is trained on? The surface question is the same at both ages. The sophistication of the answer scales with development.
Question 2: Can I verify this?
The second question moves from skepticism to action. Having noticed that the AI is making a claim based on pattern rather than ground truth, the next step is: where would I go to check this independently?
This question does several things at once. It forces the student to identify what kind of claim they're evaluating, which is itself a useful thinking move. A claim about a verifiable fact (a date, a measurement, a quote) has a different verification path than a claim about interpretation or causation. Asking "can I verify this?" prompts the student to figure out what kind of claim it is before deciding how to evaluate it.
It also creates a practical decision point. Not every claim in an AI output needs deep verification for every purpose. A student writing a casual summary for their own notes operates at a different standard than a student preparing something to be submitted or shared. The question "can I verify this?" makes that decision explicit rather than leaving it to chance — and the habit of making the decision consciously, even when the answer is "yes but I'm going to accept it anyway given the stakes," is itself meaningful.
For younger children, the verification question is most naturally paired with a physical action: let's look it up. The verification habit forms when looking things up becomes the automatic response to uncertainty, rather than something that happens only when a parent asks. For older students, it becomes a more sophisticated judgment about sources — where is the most reliable place to check this particular kind of claim, and how would I know if the sources I'm consulting are themselves trustworthy?
Question 3: What if it's wrong?
The third question introduces consequence reasoning. Before accepting an AI claim and moving forward, the student asks: if this were wrong, what would happen?
This question is what makes the framework genuinely calibrated rather than uniformly skeptical. Not every AI error has the same cost. A student who applies maximum verification effort to every AI claim regardless of stakes will burn out quickly and eventually abandon the habit entirely. A student who has learned to assess consequences applies verification effort proportional to what's actually at stake, which is both more practical and more sustainable.
For a low-stakes claim — a fun fact about an animal, a general description of a historical period, an overview of how something works — the answer to "what if it's wrong?" might be: not much. The context is informal, the information isn't being cited or acted on, and the cost of a minor error is low. The student notes the uncertainty and moves on.
For a higher-stakes claim — a citation that will appear in a paper, a statistic that will be shared with others, information that will be used to make a real decision — the answer to "what if it's wrong?" is consequential. The student applies more verification effort, consults primary sources, and doesn't consider the claim usable until it's been independently confirmed.
This is the same judgment process that working adults apply to information evaluation in professional contexts. It's learnable at any age when the question is made explicit, and it's the question that separates calibrated oversight from blanket distrust on one side and uncritical acceptance on the other.
How the Three Questions Map to Everything Else
These three questions are the portable form of the metacognitive cycle we described in our post on metacognitive oversight. Planning corresponds to asking "what if it's wrong?" before you start — understanding the stakes so you know how much verification effort is warranted. Monitoring corresponds to asking "how do they know?" while you're reading — staying actively skeptical rather than passively absorbing. Evaluating corresponds to asking "can I verify this?" after — following through on the flagged claims before accepting the output.
They're also the direct application of the distinction between critical thinking toward the AI and critical thinking for the assignment that we explored in our post on the two types of critical thinking. Running the three questions is what critical thinking toward the AI looks like in practice, claim by claim.
The "How do you know?" habit that forms the spine of this series is the seed version of Question 1. The Citation Hunt activity for middle schoolers is a structured exercise in Question 2. The Confidence Calibration Challenge for teenagers is Question 3 in game form, practiced across a range of topics until calibrated skepticism becomes automatic.
The framework isn't new material. It's the common structure that connects everything that's come before it in a form compact enough to fit in a child's working memory.
Introducing the Framework at Home
The best way to introduce the three questions isn't to hand a child a list and ask them to apply it. It's to model the questions yourself, out loud, in a real interaction.
Find a moment when your child is using AI for something low-stakes — a fun question, casual curiosity, something with no homework pressure attached. Read the output together. Then ask the questions conversationally, as if you're genuinely curious: I wonder how the AI knows that. Do you think we could check it somewhere? And what would happen if that turned out to be off?
Done naturally, in the right moment, the questions land as curiosity rather than protocol. The child picks up not just the three questions but the posture behind them — that these are interesting things to wonder about, not tasks to complete before moving on.
From there, the transition to independence is gradual. You ask the questions alongside them. Then you ask them as prompts: "what are we going to ask first?" Then you ask whether they ran through the questions at all after they've finished. Then you don't ask, and they report back on what they found. The scaffolding fades as the habit forms.
Most families find that the questions become genuinely internalized faster than expected, particularly when they're applied to something that fails. The first time a child runs Question 2, looks up a citation, and finds that it doesn't exist — that's the moment the framework becomes real. Everything before it is setup. That moment is the habit taking root.
This framework is the foundation of Orbrya's AI literacy curriculum, and it's the tool we'll reference throughout every post that follows in this series. If you want to be the first to access the structured curriculum we're building around it — 52 lessons across 12 weeks, launching August 2026 — join the waitlist.
Next week we take a step back and look at the bigger picture: what the difference between teaching kids to use AI and teaching them to audit it actually means for the families making decisions right now, and why the distinction matters more than anything else being debated in AI education.
Orbrya builds AI literacy curriculum for K-12 families — homeschool and supplemental — focused on the verification and evaluation skills that separate effective AI use from dependent AI use. Curriculum launches August 2026.