There are two exhausting ways to talk about AI in learning. One says it changes everything and the tutor is dead. The other says it's a glorified autocomplete that will rot your brain. Both are good for engagement and useless for building anything.
The truth is duller and far more practical: AI is genuinely excellent at some parts of learning and genuinely poor at others, and the entire job is knowing which is which. So here's the honest version, the one we actually use to decide what to build and what to leave alone.
What it's genuinely good at
The thing AI does that nothing before it could is respond. It can hold a real conversation, take a role and stay in it, and react to what you actually said rather than what a script anticipated. For practice and feedback, that one capability is close to the whole game.
- Rehearsal that reacts. A simulated client, interviewer or stakeholder that pushes back, so you practise the real interaction instead of reading about it. This is the part that used to require a willing human and a free afternoon.
- Feedback on demand. Specific, immediate notes on what you just did, in the moment, instead of a generic rubric a week later when the attempt has gone cold.
- Explaining a thing five different ways. Until one lands. A patient, un-embarrassing re-explainer is wildly underrated, because most people stop asking after the second time they don't get it.
- Endless, varied practice. New scenarios, new questions, new awkward follow-ups, without a human having to author each one by hand.
Notice what these have in common. None of them is "tell me the truth about the world." They're all "give me a place to practise and react to how I'm doing." That's the lane where AI is not just useful but genuinely new.
What it's genuinely bad at
Now the part the launch videos skip. AI will state a wrong answer with total confidence, in the same calm tone it uses for correct ones. It has no stake in the outcome, no accountability when it's wrong, and crucially, no reliable sense of when it's out of its depth. The fluency is the danger: it sounds equally sure whether it's right or inventing.
Confident and correct are not the same thing, and AI is much better at the first one.
So treated as an oracle, it's actively dangerous. Treated as a sparring partner, it's brilliant. The failure mode in education is always the same: someone points the "knows everything" framing at a domain where being wrong has consequences, and the calm confident tone does the rest.
The distinction is sharper than it looks. Ask a model to role-play a tough customer and react to your pitch, and a wrong move on its part costs nothing: it's a rep, the whole point is that you're testing yourself against it. Ask the same model to confirm a dosage, a clause, or a tax threshold, and a wrong move costs everything, because you're trusting it. Same tool, same confident voice, opposite risk. The skill isn't in the model. It's in knowing which of those two things you're actually doing.
The third category: things it shouldn't do even if it can
Beyond "good at" and "bad at" there's a category people skip: things AI is perfectly capable of doing that it still shouldn't. It can convincingly pretend to be a person; it shouldn't. It can quietly generate content and let you assume a human wrote it; it shouldn't. It can offer a confident verdict on a medical, legal or financial question; it shouldn't get the final word where a real professional is required.
These aren't capability limits. They're choices. And a product reveals what it actually believes by which of these lines it's willing to cross when no one's checking.
How we draw the line
The rule we build to is short enough to remember: use AI for the practice, keep humans accountable for the truth.
In practice that means the AI runs the rehearsal, scores the attempt, and coaches the next rep, while the substance it's practising against (the qualification, the standard, the real-world consequence) stays anchored to something verifiable rather than to the model's confidence. When our products generate something, they say so. When the AI sounds certain, we don't treat that certainty as evidence. And the model never pretends to be a human, because the entire value of the practice is that you know exactly what you're practising against.
This is one of our stated principles, not a footnote: honest beats flattering. A learning tool that tells you what you want to hear is worse than useless, because you'll believe it and stop improving.
Why the honesty is the product
It would be easier to overclaim. Most of the category does. "AI tutor that knows everything" sells better than "AI sparring partner that's sometimes wrong and that's fine, because you're using it to practise, not to look things up."
But the whole point of practice is that it tells you the truth about where you actually stand. You cannot build a truth-telling product on top of a marketing lie. The moment the tool flatters you to keep you happy, it stops doing the one job that made it worth using.
So we'd rather just say it plainly. AI won't make you ready on its own. Pointed at the right job, as the thing that finally lets you practise the doing instead of only reading about it, it gets you there faster than anything that came before. That's a big enough claim to be worth making. It doesn't need the hype on top.