There is a finding in educational research that everyone in this field eventually runs into. Benjamin Bloom published it in 1984 and called it the two-sigma problem. A student who works one-on-one with a skilled tutor performs about two standard deviations better than a student in a thirty-person classroom. Two standard deviations is the gap between an average student and a top one. The catch, of course, is that nobody can afford to put a tutor next to every student. That is the actual constraint, and it is the reason Bloom's paper stayed a problem for forty years instead of becoming a recipe.
AI is the first thing that has changed that math. Not because an AI tutor is as good as a great human one, it is not, but because the second-best version is now affordable enough to put in front of every student who needs it.
What an AI tutor can do well
The honest list is shorter than the marketing copy suggests, but it is longer than skeptics give it credit for. An AI tutor is good at:
- Explaining the same thing six different ways. The first explanation a student hears is the textbook's. If it does not land, the textbook does not have a second one. An AI tutor does, and a third, and a fourth.
- Being available at the moment of confusion. A doubt at 9pm on a Tuesday, the night before the test, is the doubt that matters. A teacher cannot be there. A tutor on a phone can.
- Holding patience the student does not have to earn. Students who would never raise a hand in class will type the same question three times in a row to a chat window. The bar to ask is lower, so the asking goes up.
- Working in the language the student thinks in. For most of the students we serve, that means Hindi as easily as English. A textbook does not switch. Brighta does.
None of this is a replacement for a teacher. None of it is supposed to be. It is a replacement for the silence that used to fill the space between the lesson and the exam.
What an AI tutor should not do
The harder design question is what to refuse. We thought about this for a long time before Brighta shipped, and the rules are simple:
It explains. It does not finish the homework. If a student asks for the answer to question three, Brighta will walk through the method, ask them to try, and then check. The completion is the student's, every time. We could have built it to just answer. We did not, because the point of homework is the trying.
It admits when it does not know. An AI that confidently invents a wrong answer is worse than no AI at all in a learning context. Brighta is tuned to say "I am not sure, here is what I would check" before it guesses.
It does not pretend to be the teacher. A real teacher knows the student. Brighta does not. The tutor and the teacher are two different roles, and conflating them is how trust gets damaged.
Why the name
We could have called it the AI Tutor. It would have been accurate. We did not, for two reasons.
The first is pedagogical. Students treat a thing with a name differently from a thing with a category. A tool gets put away. A study partner gets remembered. Naming changes the relationship from "I used the AI" to "I asked Brighta," and the second sentence is the one that produces a returning student.
The second reason is design discipline. Naming forces you to take a position on what the thing is. Brighta is not a search engine, and she is not a chatbot. She is a tutor, and the rules above follow from that. If we ever decide to add a feature that a tutor would not do, the name pushes back.
What we have learned so far
The thing that surprised us first was the time of day. The peak hour for Brighta is not the hour around the school day. It is the two hours before bedtime, when the student is sitting with the day's lesson and trying to make it stick. That is exactly the hour a human tutor cannot reach.
The thing that surprised us second was who used it most. We expected the strong students, the ones who already asked questions. The pattern is the opposite. The students who use Brighta the longest are the ones who had stopped asking, because the cost of asking, in front of the class, in front of a teacher who already moved on, was too high. A patient interface lowers that cost to almost zero.
The point
Bloom's two-sigma problem is not solved. It will not be solved by AI alone, because the missing piece in most students' learning is not raw access to explanation, it is a teacher who knows them. But the gap between "no help when I need it" and "some help when I need it" is the largest gap in the chain, and that is the gap an AI tutor closes.
Brighta is the version of that we ship today. It will be a better tutor in a year than it is now, and a better one the year after that. The name will stay, because the role is the part we are committed to.
