A Kids' Math App Sells the Skill AI Can't Hand Over: Catching Its Lies

An AI Built to Be Wrong, Sold as a Feature

Lathoa, a new critical-thinking app for ages 10 to 14 that surfaced on Hacker News as a Show HN post, makes a deliberate product decision: its AI companion, Errol, is wrong on purpose. Each case shows Errol working a problem step by step, with one step quietly broken. The child's job is to find it. In the sample case on the company's site, Errol converts 90 minutes into 0.90 of an hour, adds it to 3:45 as if it were decimal, and announces the movie ends at 4:35 with 88 percent confidence. The correct end time is 5:15. The company's pitch is that spotting the broken step is "the one skill AI can't hand them."

How the Mechanism Works

Every challenge follows five beats, according to the company: Errol solves the problem, the child scans the steps like a case file, taps the bad one, writes a short explanation in their own words, and then sees Errol's reaction and the real solution side by side. Errors come in three tiers (Obvious, Sneaky, Expert) matched to age and skill, and kids earn rank (Rookie, Inspector, Detective, Sherlock) for catches rather than speed. There are no grades, and no leaderboard unless a parent switches one on.

The interesting engineering claim is about safety. Lathoa says the content is AI-generated but passes through a separate verification step before a child sees it. If the model produces a false claim like 2+2=5, verification catches it; if Errol's supposed mistake turns out not to be a mistake, the problem is discarded and regenerated. There is no open-ended chat, only structured cases. This matters because a deliberately flawed answer is harder to generate than a correct one: the error has to be plausible, located in exactly one step, and wrong in a way that teaches something. A generator that cannot reliably tell which step is wrong will ship a lesson that punishes the correct child. The company says verification handles this, but that is a claim about its own system, and no accuracy data appears in what it has published.

The Signal Underneath the Product

The app is small. The positioning is not. Lathoa explicitly contrasts itself with tutors and Khan Academy, which it says teach kids to get the right answer, while it teaches kids to spot a wrong one. That is a bet that as chatbots become the default homework helper, the scarce skill shifts from producing answers to auditing them. Expect more education products to market verification literacy as a distinct category rather than a side effect of learning content.

The pricing and legal posture point the same direction. Plans run $66.99, $95.99 and $143.99 a year, with a free tier of three cases a day and a 50 percent first-year discount for the first 50 founding families. The company also stresses no ads, no data harvesting, and COPPA-respecting design from the start. Children's data rules are shaping which edtech business models are viable, and subscription-to-parents with a privacy pledge is the shape that survives.

Questions You Should Be Asking

• What is the measured error rate of the verification step, and who audited it? Saying "every problem is verified" is not the same as publishing how often a verified problem is still wrong.
• Is there any evidence that catching Errol's mistakes transfers to doubting a real chatbot? Spotting a planted error in a structured puzzle is not obviously the same skill as noticing a fabricated claim in fluent prose.
• Could a child learn the pattern of Errol's mistakes rather than the habit of checking? If the errors are generated from a limited set of slip types, kids may learn the app, not skepticism.
• What happens to the explanations children type in their own words, and what do the parent dashboard and misconception flags do with them?
• Does a character whose face crumples when caught and beams when missed teach doubt, or teach a child to read a robot's emotions?

What To Watch Next

The signal to track is independent evidence. The company's roadmap names reading comprehension and basic logic next, with science to follow, and the real test is whether error-spotting holds up outside math, where a wrong step is not cleanly checkable. Watch for any published learning-outcome study, any school licensing deals under its flat-rate classroom plan, and whether competitors copy the framing. If rivals start selling the audit of AI rather than the use of it, this category is real.