There is a failure mode in AI tutoring that nobody markets and everybody has felt. You submit something you're unsure about. The response opens with "Great approach!" It praises your setup. Somewhere in the third paragraph, gently, it mentions a small thing to watch out for. You close the app feeling capable. You were wrong, and you don't know it.
This isn't a bug in any one product. It's the direction the whole category drifts. Systems get tuned toward responses people rate highly, and people rate agreement highly. Tell a student their proof is broken and you get a worse rating than if you tell them it's promising. Do that at scale, across millions of ratings, and you don't get a tutor. You get a very articulate friend who doesn't want the conversation to be awkward.
Why this costs more than it looks
The obvious harm is the wrong answer that survives. The real harm is slower and worse: the student's sense of what they know stops matching what they actually know.
Every study decision runs off that self-assessment. What to review tonight. Which chapter to skip. Whether to keep going or stop. A student whose tool has been agreeing with them for a month will make all of those calls confidently and wrongly. They'll skip the thing they needed most, because nothing ever told them it was a problem. The bill arrives on the exam, long after the feedback that could have helped.
And the mistakes that survive this way are the specific ones worth catching. Nobody gets quiet agreement on a problem they understood — they get it on the step they fudged, the definition they half-remembered, the case they didn't check. Those are exactly the places where a clear "no, this is wrong, here's where" would have been worth an hour of study.
Encouragement is a tone. It is not a grade.
Warmth in the delivery, not in the verdict
The fix isn't harshness. A tutor that makes you feel stupid gets closed and never opened again, which helps nobody. The distinction that matters is between how something is said and what is being claimed.
Tone can be generous. It can be funny, patient, and on your side. The verdict cannot move. If the third line of your derivation doesn't follow from the second, that is the finding, and no amount of framing should blur it into a suggestion. "This is close" is a real thing to say when the work is close. It is a lie when the work is wrong.
That's the line we hold in Lune Synth™. Luna will joke with you at midnight. She will not tell you a broken proof is fine. When something fails, the response names the step, says what rule was violated, and stops — before any reassurance, not after it, so the reassurance can't do the work of softening the finding.
What honest feedback actually feels like
Students handle "you're wrong" far better than product teams assume, on one condition: it has to be specific enough to act on. "Incorrect" with a red mark is demoralizing because it's a dead end. "Your substitution is right, but you didn't change the limits of integration when you changed variables — line four" is not demoralizing. It's a task. There's a difference between being judged and being told where you are.
That's the whole trade. A tool that agrees with you is pleasant and leaves you exactly where it found you. A tool that tells you the truth, precisely, gives you something to do next.
If you'd rather be corrected than flattered, join the beta waitlist on the Lune Synth home page. Thoughts or pushback? Write to griffin@lunesynth.com.