AI Tutors Are Knocking on the '2 Sigma' Door — What Dartmouth's Effect Sizes of 0.71 to 1.30 Actually Mean

Education research has one famous homework problem it hasn’t solved in over 40 years. It’s called Bloom’s 2 sigma problem. And in a recent real-world course at Dartmouth, an AI tutor turned in a report card that knocks on that old wall: effect sizes of 0.71 to 1.30 SD. The numbers alone are eye-catching. Let’s unpack what they actually mean.

A quick bit of honesty first: this isn’t a hot topic that’s been blowing up on Hacker News or X over the past month. It’s more of an evergreen debate that education researchers have circled for decades. So instead of chasing live reactions, I want to focus on why these numbers matter — and where the context sits.

What the ‘2 Sigma Problem’ Even Is

In 1984, educational psychologist Benjamin Bloom published a study that still haunts the field. He split students into three groups. The first got ordinary classroom lectures. The second used “mastery learning,” where you can’t move on until you clear a bar. The third got one-on-one tutoring.

The result was stunning. Students who received one-on-one tutoring scored a full 2 sigma — two standard deviations — higher than students in the ordinary classroom. In plain terms, an average student getting private tutoring leapt to roughly the top 2% of a regular class. A middle-of-the-pack kid suddenly performing like the sharpest student in the room. That’s the magic.

And that’s where Bloom’s question lands. One-on-one tutoring clearly works. But no society on Earth has the money to hand every student a dedicated private tutor. So: is there a way to get tutoring-level results at scale? That question has been the holy grail of education for four decades.

How Big Is an Effect Size of 0.71 to 1.30?

Back to Dartmouth’s 0.71 to 1.30 SD. Here, SD means standard deviation — the unit we use to measure effect size.

Education research has a rough rulebook. An effect size of 0.2 is “small,” 0.5 is “medium,” and 0.8 is “large.” By that scale, 0.71 is already brushing up against “large,” and the upper bound of 1.30 is unusually powerful for any classroom intervention.

If that still feels abstract, try this: an effect size of 1.0 lifts an average student to about the top 16%. Given that most education policies and study methods fight it out somewhere between 0.2 and 0.4, the range this AI tutor posted genuinely stands out.

But remember, Bloom’s magic number was 2.0. The AI tutor’s report card lands at roughly half to two-thirds of the way there. So the honest framing isn’t “it broke through the wall.” It’s “it started knocking on the wall for real.”

Why AI Tutors, and Why Now

Computer-based learning isn’t new. For decades, e-learning platforms and adaptive-learning tools have taken swings at the 2 sigma target and struck out. Most stalled around an effect size of 0.3 to 0.4.

Large language models changed the equation. The key difference is conversation. Older systems served up a fixed problem and simply checked whether you got it right. An LLM tutor answers your weird tangential questions, asks you why you got something wrong, and reshapes its explanations to fit your level. It’s far closer to what a real tutor does.

The elements Bloom prized in one-on-one tutoring were exactly this: immediate feedback and personalized response. Forty years ago, only humans could deliver them. Now software is starting to approximate them — and Dartmouth’s numbers are the receipt.

Before You Take the Numbers at Face Value

None of this is a reason to go starry-eyed. A few things deserve a cold look.

First, the range from 0.71 to 1.30 is wide. That likely means the effect swung hard depending on the student, the subject, or how it was measured. It wasn’t uniformly great for everyone — it probably worked especially well under certain conditions.

Second, this was a university course. College students, who tend to be self-motivated, aren’t the same as K-12 students who need help managing focus. A lab win doesn’t automatically scale to every classroom.

Third, there’s the novelty effect. The thrill of using a shiny new tool can juice scores, and that buzz fades. The real test is whether the gains hold up several semesters later.

The Bottom Line

Bloom’s 2 sigma problem is, at heart, a story about equity. It asks whether the benefit of one-on-one tutoring — long reserved for kids from wealthy homes — can be handed cheaply to everyone. Dartmouth’s 0.71 to 1.30 isn’t a complete answer, but it’s a meaningful marker pointing in that direction.

So where do you land? Do you think AI tutors will eventually replace human ones, or stay parked in the “helpful assistant” seat? The day these numbers reach 2.0, the classroom will look very different from the one we know today.