We have now completed our tour of the major tutor moves: (which is not very effective, except as the coda to a learning episode), scaffolding (effective for surface learning but difficult for deep), feed back (ditto), and checks for understanding (easy for surface understanding but harder for deep diagnoses).

The clear pattern is that surface learning is quite easy to obtain—indeed, it is difficult to avoid. That perhaps accounts for the observation we started with: the general effectiveness of tutoring however and by whomever it is delivered. But deep learning is much harder to achieve, requiring a great and targeted effort by tutors. Even then, some of a student’s misconceptions—such as a “folk” understanding of the physics of motion that objects slow down and stop unless a force is pushing them—may have been held for so long that a single session of tutoring is simply not enough to dislodge them.

Or, to restate the previous paragraph for the optimistic reader, there is a colossal opportunity to improve what tutors and tutoring can achieve. We can imagine two kinds of tutors: those who produce perfectly respectable surface learning easily and declare victory; and those who make a determined effort to engender deep learning—succeeding, even so, only part of the time. What follows is intended for the second group.

Let us first repeat that surface learning is not without value. Some content isn’t sufficiently meaningful to allow anything else: Knowing your multiplication facts does not require a great conceptual leap. Neither does decoding written text in the early grades, nor punctuating writing in later elementary. Most content, though, can be treated at both a surface and deep level. Indeed, this might be the definition of good art.

“The world breaks everyone and afterward many are strong at the broken places.” “At the still point, there the dance is.” “The pieces I am, she gather them and gave them back to me in all the right order.”

In math and science, students may find themselves being given a procedural, surface-level understanding— tips and tricks such as dividing fractions by flipping and multiplying or reversing the sign when you move a term from one side of an equation to the other. Instruction designed to impart a deeper, conceptual understanding is harder to serve up, and so, it is often simply left out for several grades. This is why, when you arrive at calculus, the wheels come off.

Whatever the topic, traveling beyond beginner levels and into the realm of experts requires depth. Perhaps this is why schools have evolved to turn out beginners in a wide range of topics and experts in none.

So deep learning, sooner or later, is a cognitive necessity. What remains is for us to identify the tutor moves that lead reliably to it. Unfortunately, detailed studies of tutoring sessions have not been able to find them. Micki Chi analyzed the rare occasions when deep learning was glimpsed in tutoring sessions in order to answer the question: which tutor moves predicted it? She couldn’t find any. Nothing tutors did in the sessions she studied reliably led to deep learning.

This points to the strong possibility that, like Bigfoot and ufos, deep-learning tutor moves are difficult to find for the simple reason that they don’t exist. Kurt VanLehn, a computer scientist and education researcher at Arizona State University, writes, rather shockingly, of “the intriguing possibility that the content of the tutor’s comments may not matter much.” He doesn’t mean a tutor is unnecessary. Rather, he means that tutors’ key contribution to learning may not be the content they share but their ability to orchestrate the session so that the student engages with the topic in a sustained way. Indeed, it is a striking fact of tutoring that tutees are so engaged. They almost never fail to respond when the tutor asks a question. (If that doesn’t seem odd, contrast it with the classroom, where most students don’t respond to most questions, or to online learning, where student attention wanders all over the place.) Perhaps the secret sauce of tutoring is simply making it difficult for a tired or bored student to hide.

There is something in that, but Chi the indefatigable— who, just to make things interesting, is VanLehn’s spouse—wasn’t content with it. Spooling compulsively through the rows of numbers in her statistics app, she finally noticed one thing that actually did predict deep learning. It was the moment of a student reflecting on her own progress—for instance, saying something like “Hmm, I understand most of this, but not all of it.” Students who did that demonstrated more deep learning later on. Perhaps, thought Chi, we have been looking in the wrong place all along: it isn’t what the tutor does that matters; it’s what the student does.

Once we switch focus from tutor-moves to studentmoves, a whole raft of potential signs of deep learning suddenly comes into view: forming hypotheses (“Maybe, germs can get in through your skin, through a cut”), extrapolating to other situations (“There’s fighting going on between the good and the bad germs … a whole new world inside the body”), coming up with predictions (“What if there was a super fighter germ?”), forming analogies (“So, the septum is like a wall in your heart”), generating justifications, generating critiques, and revising existing knowledge to deal with conflicting information.

What everything in that list has in common is, once again, connecting knowledge, fitting newly learned information together with something that was already there, like rain falling on a pond. For instance, to reflect is to ask whether the new information fits. Hypotheses, extrapolations, predictions, and analogies require you to begin with the new information and take it somewhere. Generating justifications and critiques and resolving conflicts require you to bring existing knowledge to bear on the new information.

If we can get a student to make any of these connecting moves with new knowledge they just encountered—or even better, more than one connecting move—deep learning is more likely to follow. Are there tutor moves that prompt students to make such a connection? And if so, why didn’t those moves show up in the research? Didn’t we just conclude that, maybe, they don’t exist?

We did. Though perhaps they do exist, just not where we are looking. The corpus of tutoring sessions that researchers sift through may not be rich enough to allow them to be found. Sessions in that corpus are typically led by inexperienced tutors—frequently college students, a ubiquitous and affordable resource right outside every research lab’s door—and focus on surface topics, such as solving simple physics problems. It’s like trying to discover new species of whales by studying the lakes of England. They aren’t there. You won’t find deep souls in shallow water.