When AI fills a gap with something convincing.

Why the error can look so polished

A language model produces likely sequences of text from patterns and context. That ability can generate clear explanations, but clarity is not a built-in truth test. When the model lacks reliable context—or when a prompt asks for a detail that may not exist—it can still produce an answer in the expected format.

The result may include realistic author names, journal titles, URLs, quotations, legal sections, product features, or historical details. Specificity can make an error feel more trustworthy even though every detail still requires evidence.

Common forms of hallucination

• Invented citation: a plausible article, author, journal, or link that cannot be found.
• Fabricated quotation: words attributed to a real person who did not say them.
• Wrong number: a percentage, price, measurement, or date generated without support.
• Blended identity: details from two people, places, products, or events combined into one description.
• Outdated answer: a once-correct rule or product detail presented as current.
• False causal story: a neat explanation for why something happened when the evidence only shows correlation or uncertainty.

What makes an error more dangerous?

Risk increases when the answer affects health, law, money, school assessment, personal safety, reputation, another person’s rights, or an irreversible decision. It also increases when a learner lacks background knowledge, the source is hard to access, or the interface encourages a fast response.

A child-friendly explanation

Try: “The AI is very good at making a sentence that fits. Sometimes it fills a blank with something that looks right instead of something we have proved is right. Our job is to find the parts that need evidence.”

Avoid saying that AI “knows it is lying.” That gives the system human intentions and can obscure the real lesson: output must be evaluated regardless of tone.

Four habits that reduce the risk

1. Ask what is checkable. Mark every date, number, quotation, citation, current rule, and claim about a real person.
2. Move outside the conversation. Use an independent source instead of asking the same model to validate itself.
3. Match the source to the claim. Use official documentation, the original research, a government authority, or another appropriate primary source.
4. Remove what cannot be verified. Do not keep a citation or quotation because it is convenient or plausible.

Family demonstration: the confidence test

1. Write one accurate statement and one invented statement in the same confident tone.
2. Ask the learner which one “sounds” more believable.
3. Reveal that tone did not determine truth.
4. List the evidence that could verify each statement.
5. Repeat with a number or citation.

This demonstrates the underlying lesson without requiring a child to enter information into a connected AI tool.

What a provider safeguard can—and cannot—do

Tools may use retrieval, citations, warning labels, content filters, or additional models to reduce errors. These controls can help, but none turns every output into verified evidence. A citation feature is only useful when the cited source exists and actually supports the claim.

When to stop checking and ask a person

Verification is not a do-it-yourself substitute for professional judgment. For medical, legal, financial, safeguarding, crisis, abuse, self-harm, or emergency topics, involve a trusted and qualified human. If someone may be in immediate danger, contact the appropriate local emergency or safeguarding service.