What AI tutors still get wrong

We build an AI tutor. So this is not a warning from the outside — it is a list of the things we keep running into, and the ones we have not solved.

If you are choosing an AI tutor, or deciding how far to trust one, these are the limits that matter more than any feature list.

It can be confidently wrong

A language model produces the most plausible next words. Plausible and correct overlap most of the time, which is exactly what makes the gap dangerous — the wrong answer arrives in the same calm, well-structured voice as the right one.

The failure that costs students most is not a wild hallucination. It is a small one: a sign flipped, a constant misremembered, a step that applies to a similar-looking problem but not this one. It reads perfectly.

What to do about it: treat a worked answer as a draft you check, not a result you copy. Ask it to redo a step a different way. Two derivations that disagree tell you something; one that sounds confident tells you nothing.

We push our tutor to recompute before it agrees with you, precisely because a warm "yes, exactly!" on a wrong answer does more damage than a cold correction ever could. But no prompt makes that guarantee absolute.

It does not know what you understood

A human tutor watches your face. They see the exact sentence where you stopped following, and they slow down without being asked.

An AI tutor sees text. It cannot tell the difference between silence-because-obvious and silence-because-lost. It cannot see you frown.

What to do about it: say when you are lost, out loud and early. This is why interruption matters so much to us — the check-in that a human does by watching, you have to do by speaking. A tutor that cannot be interrupted forces you to sit through the rest of an explanation that stopped working three sentences ago.

It has no memory of you

Most AI tutors, ours included, start each session from nothing. It does not know that you struggled with logarithms last week, that you already know Python, or that your exam is in four days.

That is partly a privacy decision and partly a limitation. But the effect is the same: you carry the continuity, not the tool. If context matters, say it in the first sentence.

It is confident about the wrong scale

Ask about a broad topic and a model will happily produce an authoritative-sounding overview of a field that takes years to learn. The overview is not wrong, exactly. It is just flattened — every part sounds equally important, because nothing in the text says which parts are hard.

What to do about it: ask for the part that people usually get wrong. That question pulls out the structure a summary flattens.

It costs real money to run, and that shapes everything

This one is invisible from the outside, so it is worth saying plainly: every answer costs the provider money. Voice costs more. Reading a scanned document costs more again.

That is why free AI tools have caps, why they get slower under load, and why some of them quietly use a smaller model than the one they advertise. It is also why our voice sits behind a sign-in — not to collect you, but because an open endpoint that spends money is an open endpoint that gets drained.

We wrote a separate piece on what it actually costs, with real numbers, because the pricing of AI products makes very little sense until you see them.

What it is genuinely good at

Having listed the limits: none of them make the tool useless. They make it a particular kind of tool.

An AI tutor is unmatched at patience. It will explain the same thing six different ways at two in the morning without a hint of impatience, and that is a real thing that human teaching often cannot offer. It is very good at turning a wall of prose into a structure you can see. It is good at the question you are embarrassed to ask.

Use it for those. Verify anything you will be graded on. And say something the moment it loses you — with a tutor you can interrupt, that costs you one sentence.