Every learning product answers two questions. Does it teach? Do students come back? We think the one that wins answers a third: how little does it ask of the student? Our answer is voice. The simplest interface is no interface, and that is what we are building with Sophie.

This essay covers that third circle: why it has stayed empty, what it takes to build, and where our effort goes next. It follows Talk to discuss, write to think, which covers why talking helps students learn.

9:40 on a Tuesday night

A junior is on number 4 of her AP Calculus homework, and she is stuck. The worksheet is on the desk. Her phone is next to it.

Here is what help costs her in a typical learning app. Unlock the phone. Find the app. Pick the course. Find the unit, then an exercise that looks like hers. Type ∫ x² eˣ dx on a phone keyboard, hunting for the superscript. Read a long answer on a small screen. Put the phone down and find her place on the paper again.

Every one of those steps is a place to quit. The math was already hard. The steps on top of it are what make her close the book and hope it's not on the test.

Two circles everyone builds in

For forty years, learning technology has worked on two questions.

The first is pedagogy. Does it teach? This is the method: explain, ask, pick the right next problem, check the work. The bar was set in 1984, when Benjamin Bloom found that students tutored one to one scored above 98% of students taught in a regular class.1 We wrote about it in The 2 sigma problem. Software has been closing the gap ever since. Tutoring systems that work through a problem step by step come close to a human tutor.2 In 2025, a Harvard team built a physics tutor that guided students one step at a time, and its students learned about twice as much as in an active-learning class, in less time.3 The method is known, and it fits in software. More in Why Sophie is not a chatbot.

The second is motivation. Do they come back? This is the reason to open the app tomorrow: streaks, points, leagues, badges, push notifications. A large share of the industry's product work lives here, and we use some of it too. It helps a little. A 2020 review of gamification studies found small gains on learning and on motivation, and the smallest, least reliable effect on behavior, the part that brings a student back.4

Draw every learning product on one page and almost all of them land in the overlap of these two circles. They teach well enough. They work hard to bring you back. And they are all screens you tap and type into.

Figure 1

Where learning products are built, and where we invest.

Three circles: pedagogy, motivation and simplicityPedagogy and motivation overlap at the top, where most learning products sit. Simplicity, which means voice, is the third circle at the bottom. Sophie sits where all three overlap.WHERE EVERYONE BUILDSWHERE WE INVESTPedagogyMotivationSimplicityDoes it teach?Do they come back?You just speak.Sophie
Each dot is a kind of learning product: adaptive practice, video courses, gamified apps, homework chatbots. Most sit between pedagogy and motivation. Sophie sits where all three circles meet, and the third circle is where our next effort goes.

The third circle

The third question is simplicity. How little does it ask of the student?

For us, simplicity means voice. Think of the best tutor you ever had. There was no login, no menu, no course picker. You sat down, said "I don't get this one," and they looked at your page. The designer Golden Krishna made the general case in a 2015 book whose title we keep coming back to: The Best Interface Is No Interface.5 For tutoring, the idea is literal. A tutor you can talk to needs nothing between you and help.

So with Sophie, the student talks. She says what's wrong in her own words, and Sophie answers out loud. The pencil stays on the paper, where the thinking happens. We call that split talk to discuss, write to think.

Why the circle is empty

If voice is so much simpler, why has almost nobody built a learning product around it? We see four reasons.

It didn't work until recently. A spoken tutor has to follow a teenager who mumbles, trails off and corrects herself halfway through a sentence, sometimes with a TV on in the next room. Then it has to answer without an awkward pause. Speech technology only got good enough for that in the last couple of years. Before, the keyboard was the only reliable way in.

Voice was bolted on as a button. When learning apps use voice at all, it's usually a microphone icon on the same screens. Press, speak, wait, read the answer. That's dictation. Building the whole product around a conversation is a different job, and most of it is about timing: when to speak, when to wait, what a silence means.

Math is the hardest subject to type. Fractions stack, exponents float, an integral needs its bounds. A phone keyboard flattens all of it into one line. Every math app asks the student to translate the problem into the app's format before help can start, and that translation is exactly the step a tired student skips. Out loud, she says "number 4."

Simplicity is built by removing things, and removing things doesn't demo. A roadmap full of new features looks like progress. A release that deletes three screens looks like nothing happened. Most product teams get rewarded for the first.

Speaking makes the other two circles work

Voice is how the other two circles reach the student.

It feeds motivation. BJ Fogg, who founded the Behavior Design Lab at Stanford, describes a behavior as three things meeting at the same moment: motivation, ability and a prompt.6 When motivation is low, the dependable lever is ability, which means making the action easier. We can't make a tired sixteen-year-old want calculus at 9:40 pm. We can make asking for help cost one sentence. Streaks try to raise motivation. Voice lowers the bar.

It feeds pedagogy. Working memory is small. John Sweller's cognitive load theory shows that effort spent on the form of a task is taken from learning it.7 Attention spent on menus, buttons and the superscript key is attention not spent on the math. Talking also makes students explain. Students who explain each step out loud understand it better,8 and learning through back-and-forth dialogue sits at the top of Chi and Wylie's ICAP framework.9 A tutor you type to gets short questions. A tutor you talk to hears the reasoning, like "I thought you had to multiply first." That one sentence tells the tutor exactly where to start.

It's faster. People speak about three times faster than they type on a phone.10 For math, with all its symbols, the gap is wider.

Figure 2

Getting help with number 4.

In a typical learning app

  1. Unlock the phone and find the app
  2. Pick the course
  3. Find the unit, then an exercise that looks like hers
  4. Type the problem: ∫ x² eˣ dx
  5. Read a long answer on a small screen
  6. Find her place on the paper again

With Sophie

“Sophie, I'm stuck on number 4.”

She looks at the page, asks what you tried, and gives one hint. The work stays on the paper.

The app steps illustrate a typical flow. They are not a measurement.

Voice is a design problem

A screen waits for a click. A conversation has timing, silence, interruptions and tone, and a spoken tutor gets each of them wrong in its first version. Most of our work on Sophie's voice has gone into a handful of moments that last a few seconds each.

Knowing when you're done. In one of our early test sessions, four of six long spoken answers got no reply until a "did you hear me?" The cause was politeness. Sophie had said "take your time, tell me when you're finished," then waited for a signal that never came. We changed how she listens, and her own "take your time" is no longer a reason to wait. In our tests, she now catches the end of the answer 7 times out of 7, up from 4.

A "yeah" is not a "stop." Students say "mhm" and "yeah" while they listen. In one session, a quick "yeah" halfway through a question made Sophie drop the question and jump ahead. Now she hears it as a sign the student is following, and finishes the question.

Silence is thinking time. When a student goes quiet after a question, Sophie waits. After about nine seconds, she asks once more, shorter or simpler, then waits again. She never moves on by herself. It is our second principle, 10% speak, 90% listen.

Noticing a drift. After three quiet minutes, Sophie checks in once, with a light joke, then leaves the student alone.

A calm voice. Math is stressful. When a student sounds frustrated or hard on themselves, Sophie says a few calm words about that first, then carries on. She never hurries them.

Listening first. When a student mentions that a test went badly, Sophie asks about it before she puts up a problem. A bad test is the best map of what to work on.

None of these moments shows up in a screenshot. Together, they decide whether talking to Sophie feels like talking to a tutor or like shouting at a phone.

The voice also has something to point at. Sophie looks at the student's page through the camera, so "number 4" means something, and the work stays on paper.

Easy to ask also means easy to ask for the answer. Sophie never gives it. She asks what the student tried, gives one hint, and waits.

The pedagogy stays. Only the interface goes.

Where our effort goes next

We keep working on teaching and on motivation. Our extra effort goes into voice, in four places.

Thinking versus done. The hardest problem in a spoken tutor is telling "I'm thinking" from "I'm finished." When we searched our sessions for "let me think" and "you keep interrupting," most came right after Sophie filled a pause with a "hmm" or corrected a student halfway through a sentence. Every early reply breaks someone's focus.

Math out loud. "One over x plus one" can mean two different things. Sophie has to say formulas so they're easy to follow by ear, and when a student's spoken math could mean two things, ask which one they meant instead of guessing.

Sophie talks less. The student should do most of the talking, because their reasoning is what we want to hear. Shorter turns from Sophie leave room for it.

The keyboard as a fallback. Voice first, every time. Typing is for the moments when speaking isn't possible.

Three numbers tell us whether it's working: how long Sophie takes to answer after a student stops talking, the share of a session in which the student is the one speaking, and how often a student has to say "wait" or "let me think."

What we don't know yet

Voice isn't always welcome. A library, a shared bedroom, a shy student. A quiet way in has to stay.

Mishearing costs more than mistyping. A tutor that hears "fifteen" as "fifty" is worse than a keyboard. Accents, cheap microphones and noisy kitchens are part of the job.

Talk can crowd out thinking. Silence is part of doing math. A spoken tutor has to be comfortable with long pauses and never fill one just because it can.

The evidence is early. The strongest study of tutoring software built this way is small and short. Our own sessions need to show the same thing over a full school year.

Back to number 4

The junior at her desk doesn't open an app. She says, "Sophie, I'm stuck on number 4." Sophie looks at the page, sees the integral, and asks what she tried first. Her pencil stays on the page.