The hard part of an AI tutor is the reply students never should have seen

The failure I worried about was never "the tutor could not answer". It was the tutor answering confidently and wrongly, in front of a child who had no reason to doubt it.

I work on an AI tutor used by students, and the asymmetry is uncomfortable. A model that says "I don't know" is a small disappointment. A model that states a wrong fact, drifts into an unsuitable topic, or phrases something in a way that reads badly to a younger student is a different category of problem, and you usually find out after the reply is already on screen. Reviewing a log at the end of the week does not help the student who read it on Tuesday. That is why the review has to sit before the student, not after.

Why a blacklist was not enough

My first instinct was a word filter. It lasted about an afternoon. A list of banned terms misses paraphrases, misses a factually wrong but cleanly worded answer, and produces false positives that make the tool useless. The harder class of problem - a confident factual error, a fairness issue in how a question is handled - is not a string match at all.

The second instinct was to let a model grade every reply. That trades one unreliable judge for another, and it gives you a number nobody can defend when a parent asks why a reply was allowed through. What I actually wanted was a review layer that surfaces red flags for a person, not a verdict that pretends to be final.

So I framed the problem as vetting. Before a tutoring setup goes in front of students, and when replies are reviewed, there are concrete things worth checking: safety red flags, quality and fairness warning signs, and a clear record of what to do when one shows up. That is a checklist job, not an oracle job.

What I built

I built TutorGuard around that framing. It reviews AI tutor arrangements and replies for safety red flags, factual errors, and policy violations, with the goal of catching risky content before students see it. The declared features are: review tutor arrangements for safety red flags, get a vetting and interview checklist, spot quality and fairness warning signs, and receive a parent safety action plan. That is the whole set. It does not certify a tutor, does not monitor a live session on its own, and does not guarantee that every risky reply is caught.

The vetting checklist and the parent safety action plan are the parts I would keep even if the rest went away. A red flag you notice but have no procedure for is just anxiety. A checklist tells you what to look at; an action plan tells a parent what to do about it. "Before students see it" is the design goal, and a procedure is what makes that goal operational rather than a slogan.

I kept "warning signs" phrasing on purpose. The tool surfaces signals for a human to review; it does not issue a pass or fail. A factual error is often domain-specific, and the person closest to the subject is better placed to judge it than any automated read.

The limits, stated plainly

This is decision-support, not a certification, not legal advice, and not a live audit. It does not guarantee that unsafe content, factual errors, or policy violations are all caught, and it does not replace human review or parental supervision. A warning sign is a prompt to look, not a finding. If you need assurance about a specific tutoring setup, that belongs with the people responsible for it.

I also want to be honest about "automatically reviews". The review is a best-effort layer over what it is given; it cannot know a model's training, cannot see a session it was not given, and cannot vouch for content outside that input. The output is a set of signals and a plan, not a clean bill of health.

If the thing keeping you up is a confident wrong answer reaching a student, TutorGuard gives you a review layer, a vetting checklist, and a parent action plan to start from. If there is a red flag you think it misses, that is a gap and I would like to hear about it.

FAQ

What does TutorGuard produce?

A review of tutor arrangements and replies for safety red flags, factual errors, and policy violations, plus a vetting checklist and a parent safety action plan. You start that review in TutorGuard.

Does it guarantee every risky reply is caught?

No. It is a best-effort review layer that surfaces warning signs for a human. It does not replace review or supervision, and it is not a live audit.

Is this a certification or legal advice?

Neither. It is decision-support. It does not certify a tutor, does not issue a pass or fail, and does not give a legal opinion.

Does it monitor a live tutoring session by itself?

No. It reviews what it is given. It cannot see a session it was not given, and it does not vouch for content outside that input.