AI-generated misinformation

Misinformation is false or misleading content. When AI generates it, the falsehood can arrive as a paragraph, a quote that nobody said, or a picture of a crowd that never met. NIST discusses the honest failure mode as confabulation: the system states something erroneous as if it knew. A person can also ask the tool to fabricate. The reader's problem is the same until they check.

A citation does not end the problem. Models attach real links to claims the page does not make. Synthetic media makes the lie look like a recording. Provenance tools and watermarks are attempts to label the file. They are not on every image a child will see this year.

Homework is where this lands first. A made-up statistic with a calm tone gets submitted. A fake image of a classmate gets forwarded. Schools need a rule for both, and a way to see the prompt so the conversation is about what the student asked.

Panic is a bad teacher. The useful habit is short: who said this, where else does it appear, and what would we do if it is false. Kids who expect some AI answers to be wrong share fewer of them.

When Otto helps with schoolwork, it teaches the step and asks the student to try it, so a made-up answer is harder to paste unseen. Kid-safe web search returns cited sources inside a parent-visible chat. A citation is still not proof until someone opens the page. Schools can see the session and can turn search or image tools off.

• Parents can read the prompt, the reply, and whether a search happened.
• Image generation is age-filtered and can be disabled per child.
• HeyOtto does not claim a watermark on every file a child might meet elsewhere.