Does Your Sight Reading Tool Actually Need AI?

Last week MakeMusic released correlated sight reading templates. A teacher picks the lesson they’re on in Habits of a Successful Musician, Sound Innovations, or Sing at First Sight, clicks generate, and gets a fresh exercise built around that lesson’s skills. Click again for another. MakeMusic’s announcement never uses the word “AI.” I think that deserves attention, because most people will assume it does.

Two Ways to Generate Music

Picture two kitchens.

The first kitchen runs on recipe cards. A head chef wrote every card in advance, and the line cook follows them exactly. The card for Tuesday’s special says: one protein, two vegetables, this sauce, this plating, 12 minutes. The cook can vary the details. Tonight it’s chicken and green beans, tomorrow it’s pork and carrots. But nothing leaves the kitchen that the card didn’t allow. If a diner has a shellfish allergy, you can trust the plate, because shellfish was never on the card. The food may not surprise anyone, and that is the whole point. Consistency is the product.

The second kitchen is run by a chef who has eaten at ten thousand restaurants. Nobody handed this chef a recipe card. Over years of tasting, the chef absorbed what flavors tend to go together and how a plate usually comes together, and now cooks from that instinct. Ask for “something like a Tuesday special” and you’ll get something plausible, often with more flair and variety than any card would produce. But the chef is working from patterns, not from a rulebook. Occasionally an ingredient slips onto the plate that nobody approved, and the chef can’t always tell you why it was there. For a dinner party, that’s delightful. For the shellfish-allergy diner, it’s a problem.

Now translate the kitchens into music.

The first kitchen is rule-based (algorithmic) generation. A person who understands pedagogy writes the rules ahead of time: which note values are allowed, which time signatures, which keys, how far the melody may leap, what range it must stay inside, and how phrases should end so the exercise sounds finished. For a lesson on dotted quarter notes, the rules permit dotted quarters and everything students already know, and nothing else. The software then assembles a new melody inside those boundaries each time you click generate. Every exercise is different, yet every exercise is guaranteed to fit the lesson, because the rules define what can happen. No machine learning is involved. A good version of this approach can still sound like real music, since the rules can encode phrasing, cadences, and melodic shape, not only note values.

The second kitchen is generative AI. A model is trained on a very large collection of existing music and learns which notes and rhythms tend to follow which. It doesn’t follow a rule that says “no sixteenth notes in lesson 14.” It predicts what is likely to come next, based on patterns it has absorbed. That gives it a wide range and a natural feel, and it can produce styles and textures that a rule set would take years to write by hand. The price is that nobody wrote the boundaries. Unless the developer adds extra checks, such as filters that reject any exercise containing a rhythm outside the lesson, the model can hand a beginning class a syncopation they’ve never seen. The software did not break its own rules, because it never had any. It followed its patterns.

There is also a middle path, and it is where I expect the most interesting work to happen: a hybrid kitchen. The chef cooks from instinct, but a second person at the pass checks every plate against the allergy list before it goes out. In music terms, a generative model proposes the exercise, and a rule-based checker confirms it fits the lesson, the range, and the level before a student sees it. You keep some of the variety and gain a guarantee.

None of these kitchens is the right answer in every case. The right one depends on what the diner needs. For a sight reading exercise tied to a specific lesson, what the diner needs is a plate with nothing surprising on it.

Why that difference matters in a rehearsal hall

A sight reading exercise is a precision instrument. If it contains a rhythm the class hasn’t learned, a student can stumble for reasons that have nothing to do with their reading skill. The exercise was simply unfair. For this job, a guarantee beats novelty.

As far as public materials show, tools in this category describe their output in engine and algorithm language. Sight Reading Factory calls its approach a unique algorithm for composing music on demand. MakeMusic describes a musicality engine. I can’t see anyone’s code, so I won’t claim to know what’s under the hood. What I can say is that nothing in their public materials claims generative AI, and for this task that may be a sound choice.

The measure that matters

A student who reads fresh, appropriately leveled music for five minutes a day gets better. That is the goal. A fourth grader cannot tell whether a rule engine or a neural network wrote their exercise, and they shouldn’t have to. If a tool doesn’t need AI to deliver that growth, it shouldn’t use AI. If it does use AI, the vendor should be able to explain why in plain language.

So when a product shows up with the “AI-powered” badge, ask three things:

1. What specific job is the AI doing?
2. Would a simpler method do that job as reliably?
3. How do you know the output is appropriate for my students?

Good answers matter more than the badge.