Developer builds math app where AI tutor makes deliberate errors

By Splitfeed EditorialLast updated Sep 30, 2026, 22:30 UTC

On September 30, 2026, a developer posted Lathoa to Hacker News under the Show HN format. The app targets children aged 10 to 14 and centers on a robot character named Errol who walks through math problems step by step, with one step deliberately containing an error. Students earn XP by identifying which step is wrong and writing an explanation of the mistake, with speed factoring into the points awarded. To prevent students from gaming the system by flagging an error on every problem, the app includes problems where no mistake is present at all. The developer noted that there is no direct large language model chat interface and that the underlying system uses a stable harness with multiple evaluation steps to handle reliability.

The design choice here is pedagogically specific: it inverts the usual role of an AI system by making the machine visibly fallible on purpose, asking the student to act as the critic rather than the recipient of correct information. This draws on a well-established learning technique, sometimes called error analysis, in which spotting and articulating a mistake requires deeper understanding than simply solving a problem from scratch. The tension worth watching is whether an AI-generated error harness can produce mistakes that are plausible enough to be instructive but not so subtle or so obvious that they lose educational value. The no-error problems add a layer of critical thinking, since students must decide whether to intervene rather than simply hunt for something wrong.

Because no reactions have been posted from the Pro-AI, Anti-AI, or Middle Ground camps yet, only typical positions can be sketched. Pro-AI commentators would likely praise this as an example of AI being deployed in a structured, bounded educational context with clear safeguards rather than open-ended chat. Anti-AI voices would probably raise questions about whether algorithmically generated errors are reliable and age-appropriate, and whether a child internalizing a wrong explanation before correcting it carries any risk. The Middle Ground camp would typically focus on the implementation details, asking how well the evaluation harness actually judges a child's written explanation and whether teachers or parents have any visibility into the results.

The most revealing next step would be any user testing data the developer shares, particularly evidence of how accurately the system scores student explanations and whether the deliberately wrong steps hold up across a range of problem types and difficulty levels.