Talking, Listening, Reading, Writing How the Way We Share Ideas Shapes Learning

Open in new tab ↗

Talking, Listening, Reading, Writing: How the Way We Share Ideas
Shapes Learning
Oct 5, 2026 · @David Rakestraw
Why the format matters
How information reaches you, and how you send it back out, changes what your brain does with it.
Reading, writing, listening, and talking are not four versions of the same thing. Each one is good at some
jobs and weak at others.
Think about the last time you really understood something hard. Maybe you read about it. Maybe you
wrote it up for a class. But there is a good chance that somewhere along the way, you talked it through
with someone who asked you questions until it clicked.
Until recently, that kind of conversation was scarce. Now an AI can hold a spoken, back-and-forth
conversation with anyone, any time. That raises a question this lesson asks you to investigate: what are
listening and speaking actually good for in learning, and how should we design new tools around them?
This reading takes about 15 minutes. It summarizes research from education and cognitive science. As
you read, watch for two things: what each mode does well, and where the evidence is still thin.
Afterward, you will talk through a short set of questions with an AI in voice mode.
A quick note on numbers. Some studies below report an effect size, a way of saying how big a difference
was. An effect size of 0.5 means the average student in the helped group scored higher than about 69%
of students in the comparison group. An effect size of 0.8 means higher than about 79%. Anything
around 0.2 is small, 0.5 is medium, and 0.8 is large.
Why school runs on text
Schools are built around reading and writing largely because text scales. One teacher can assign a
chapter to 30 students and collect 30 essays. That same teacher cannot hold 30 one-on-one
conversations at the same time.
People have worried about this trade-off for a very long time. In Plato's dialogue Phaedrus, Socrates
complains that a written text cannot respond to you: ask it a question and it just repeats the same
words. Centuries later, the scholar Walter Ong argued that writing changed how people think, making
ideas more organized, analytical, and easy to revise. Both were right. Writing gave us something
powerful, and it also took away the live exchange of a conversation.
Voice AI changes the cost of that exchange. You can now explain an idea out loud, be questioned about
it, and get an immediate reply, all without a human tutor. The real question is not whether voice will
replace reading and writing. It is which kinds of learning speech supports well, which kinds text supports
better, and how to combine them.

Conversation as a thinking tool
Conversation helps learning when you do the reasoning: explaining, defending, and building on ideas
while a partner pushes back. Talk that only checks whether you remember the answer does much less.
In classrooms. In England, a program trained teachers in "dialogic teaching," in which students explain
their reasoning and respond to one another's ideas. It was tested with nearly 5,000 students aged 9 to
10. After 20 weeks, they were about two months ahead of comparison students in English, math, and
science (Alexander, 2018). Another team recorded lessons in 72 classrooms. Classes where people
elaborated on and questioned each other's ideas did better on tests, but only when most students were
taking part (Howe et al., 2019).
One-on-one tutoring. A large review of controlled studies found that human tutoring raised scores with
an effect size of about 0.79 compared with no tutoring (VanLehn, 2011). That is large, although smaller
than the famous claim of 2.0 from the 1980s. Much of the benefit came from tutors keeping students
reasoning step by step, rather than from the tutor's own explanations.
A ladder of engagement. Learning scientists Michelene Chi and Ruth Wylie sort learning activities into
four levels, each predicted to produce more learning than the one before (Chi & Wylie, 2014):
1. Passive: listening to a lecture or reading without doing anything else.
2. Active: doing something with the material, such as highlighting or repeating it.
3. Constructive: generating something new, such as an explanation or a prediction.
4. Interactive: building ideas together with a partner, each responding to the other.
A good conversation sits at the top of that ladder. Hearing an explanation read aloud sits at the bottom,
even though both involve voice.
Turns matter more than words. In a study of 36 children aged 4 to 6, the number of back-and-forth
conversational turns with adults predicted language skills better than the total number of words the
children heard (Romeo et al., 2018). Being talked at is not the same as being talked with.
Try this: Next time you study with a friend, notice who is doing the explaining. Whoever is talking and
being questioned is usually doing most of the learning.
Explaining out loud versus writing it down
Explaining an idea is one of the most reliable ways to learn it, and the format changes what the
explanation does for you. Speaking tends to produce richer, more detailed explanations. Writing tends
to produce better-organized ones.
Explaining to yourself works. A review of 64 studies with about 5,900 learners found that prompting
students to explain material to themselves while studying improved learning, with an effect size of 0.55
(Bisra et al., 2018). Most of those studies used written explanations, so the researchers flagged speaking
versus writing as an open question.
A head-to-head test. In one experiment, 48 students read about how car engines work. Then they
explained it to an imaginary classmate, either out loud or in writing (Lachner, Ly & Nückles, 2018).

• Students who wrote produced better-organized explanations and did better on questions about
the core concepts.
• Students who spoke produced more elaborate explanations, adding examples and connections,
and did better at applying the ideas to new problems.
Why speaking helps: an audience. Follow-up research suggests that talking makes the listener feel more
real, even an imaginary one. Students who explained aloud used more words like "you" and "we" and
covered more ideas, and that predicted learning (PLOS ONE, 2021). A voice AI could go a step further: a
listener who actually answers back.
Teaching makes you try harder. In studies with 5th and 8th graders, students who believed they were
teaching a computer character named Betty spent more time on the learning activities than students
learning for themselves. Researchers call this the protégé effect (Chase et al., 2009).
The takeaway: speaking and writing are partners. Talking an idea through can help you discover what
you think; writing it down afterward can help you organize it.
Does saying it make it stick?
Saying something out loud gives your memory a small boost, but it does not by itself make you
understand more. For practicing recall, what you retrieve matters far more than whether you say it, type
it, or write it.
The production effect. Words you read aloud are remembered better than words you read silently. In
one set of experiments, reading aloud also beat hearing, mouthing, and writing the words (MacLeod et
al., 2010). The boost is biggest when you say some items aloud and read others silently, because the
spoken ones stand out. If you say everything aloud, the benefit shrinks.
Memory is not comprehension. A 2023 study found that reading whole passages aloud helped people
remember the text but did not help them understand it any better (Memory & Cognition, 2023).
Retrieval practice works in any mode. Testing yourself is one of the best-supported study strategies.
Researchers found only small differences, if any, between recalling answers by speaking and by typing
(Putnam, 2011). Even silently recalling an answer in your head often helped about as much.
The takeaway: voice is a convenient way to quiz yourself, especially with a partner who keeps asking the
next question. It is not a shortcut to understanding.
Listening versus reading
For most people who read well, listening and reading lead to about the same understanding. But reading
wins when you control the pace, and long stretches of speech can overload your memory. Spoken words
disappear; written words stay on the page.
About equal overall. A review of 46 studies with 4,687 participants, from 1955 to 2020, found no
reliable overall difference in comprehension between reading and listening (Clinton-Lisell, 2022,
summary). So listening to an audiobook is not "cheating."

Pacing makes the difference. Reading did better than listening when readers could set their own pace
and listeners could not. Reading also tended to win for informational text, like a textbook, though not
for stories (Clinton-Lisell, 2023). When you read, you can slow down, skip back, and reread a confusing
sentence. A recording just keeps going.
Disappearing words. Psychologists call this the transient information effect. Short spoken comments
that go with a diagram usually help, because your eyes stay on the picture while your ears take in the
words. But when the spoken text gets long, learning drops below what people get from reading the
same words, because they cannot hold it all in working memory (Leahy & Sweller, 2011).
Where conversation is different. A recording has the pacing problem; a conversation partner doesn't
have to. In a conversation, you can interrupt, ask for something to be repeated, or say "slow down."
Whether that is enough to match the advantage of self-paced reading has not been tested yet. That is
one of the questions this lesson leaves open.
Try this: Have someone read you a phone number, a street address, and a three-step set of directions all
in one breath. How much do you remember? Now imagine the same thing with equations.
Timing: wait time and responsive feedback
In a spoken conversation, timing shapes thinking. Fast replies help when they respond to your
reasoning; they hurt when they cut your thinking short.
Wait time. In the 1970s and 1980s, science educator Mary Budd Rowe spent years recording real
classroom lessons, mostly in science. She found that teachers usually waited less than one second after
asking a question before speaking again. When teachers waited about three seconds, both after the
question and after a student's answer, students gave longer, better-reasoned answers, and quieter
students joined in more (Rowe, 1986, summary).
Voice AI often does the opposite. Many systems reply the instant you pause. If you are pausing to think,
that can cut you off mid-idea. A well-designed voice tutor should wait, and you should feel free to say
"let me think."
Feedback that carries information. A review of 435 studies found that feedback improves learning, with
an average effect size of about 0.48. Detailed, informative feedback helped much more than simple
praise or "right"/"wrong" (Wisniewski, Zierer & Hattie, 2020). The review could not settle whether
immediate or delayed feedback is better in general.
The takeaway: the special thing about conversation isn't speed. It is that each reply can depend on what
you just said: probing a half-finished idea, catching a mistake before it grows, or asking the next
question while the idea is fresh.
What reading and writing still do best
Text is not an old format waiting to be replaced. Writing about what you are learning reliably helps, and
text offers things speech cannot: permanence, self-pacing, visible structure, and precise symbols.

Writing to learn. Across 56 school experiments in science, social studies, and math, students who wrote
about the content learned more, with an effect size of 0.30. The benefit was similar across subjects and
grade levels (Graham, Kiuhara & MacKay, 2020).
Writing organizes. In the out-loud-versus-writing experiment earlier, written explanations were better
structured. Writing lets you see your whole argument at once and rearrange it.
Some things have to be seen. Equations, graphs, data tables, code, and diagrams don't translate well
into speech. Try saying the quadratic formula out loud and have a friend write it down from your words
alone.
Here is how the research divides the work:
Learning job Talking and listening Reading and writing
Generating ideas and explaining
them in detail
Stronger Weaker
Organizing an argument Weaker: hard to restructure as
you go
Stronger: visible, easy to revise
Being questioned and pushed
back on
Natural and immediate Slower; needs someone to
respond
Long or dense information Overloads memory Self-paced and rereadable
Equations, data, and diagrams Poor Essential
Quizzing yourself About equal About equal
Keeping a record to revisit Only with a transcript Built in
The best learning tools probably use both: talk an idea through, then write a short summary in your own
words.
Voice AI so far
Talking computers have a decent track record for young children and language learners, and well-
designed AI tutors have shown big gains in college physics. But the research we reviewed for this lesson
found no study directly showing that a voice AI tutor teaches concepts better than the same tutor in
text. That question is still open.
Young children. In the PBS KIDS show Elinor Wonders Why, researchers added an AI version of the main
character who asked children questions and responded to their spoken answers. Children who watched
the talking version scored higher on a science test than children who watched the regular episode (Xu et
al., 2022).
Language learners. Several reviews find medium-sized benefits from practicing a new language by
talking with a chatbot, with effect sizes around 0.5 to 0.6 (Wang et al., 2024; Bibauw et al., 2022). Many
students also report feeling less nervous. But in at least one study, students' anxiety went down and
confidence went up while their measured speaking skill did not improve (System, 2024).

College physics. At Harvard, 194 introductory physics students learned one topic from a carefully
designed AI tutor and another from an active-learning class. They learned more, in less time, from the AI
tutor (Kestin et al., 2025). That tutor was text-based, and it was built around research on how people
learn.
A warning about design. Nearly 1,000 high school math students in Turkey used one of two AI helpers. A
plain ChatGPT-style helper raised their practice scores by 48%. But on a later exam taken without AI,
those students scored 17% lower than students who never had AI at all. A second helper, designed to
guide students instead of giving answers, largely avoided that drop (Bastani et al., 2025).
The takeaway: whether a tool helps depends less on the technology than on its design, especially on
who is doing the thinking. Feeling like you learned is not the same as being able to do it on your own.
Who voice helps, and who it can leave out
Voice can open learning up for some people and shut others out, so a good voice tool always needs a
text option that works just as well.
Who might benefit most. In the dialogic teaching trial, students from low-income families gained more
than the group as a whole (EEF report, 2017). In the children's AI research, English language learners
showed some of the biggest gains (EdSurge). Listening can also help people who struggle with reading,
such as some students with dyslexia.
Speech recognition isn't equally accurate for everyone. A 2020 test of five major speech recognition
systems found about 35 errors per 100 words for Black speakers, compared with 19 for white speakers
(Koenecke et al., 2020). The systems have improved since, but accents, dialects, and technical
vocabulary still cause mistakes. A tutor that mishears you may misjudge what you know.
Voice doesn't work for everyone or everywhere. People who are deaf or hard of hearing, people who
stutter or have other speech differences, and anyone without a quiet, private space may need a
different way in.
Privacy. Your voice is personal data. Before using a voice tool, check whether it saves recordings or
transcripts. Don't share private information or record other people without their permission.
Thinking like a designer
In class, your team will try a voice AI tool that exists today and design one that doesn't exist yet. The
research above turns into a set of questions you can ask about any idea:
1. Who is doing the thinking? Does the learner explain, predict, decide, and get questioned, or
mostly listen?
2. Why voice? What does talking add that typing or reading wouldn't: turn-taking, interruption,
speaking practice, hands-free use, a sense of audience?
3. What belongs in text? Which parts are long, dense, numerical, or visual, and should be on a
screen or page? Is there a written step at the end?
4. Is there time to think? Does the design leave pauses, let the learner hold the floor, and avoid
answering its own questions?

5. Who gets left out? What happens if the tool mishears someone, or if a learner can't use voice?
6. How would you know it worked? What could the learner do afterward, without the AI, that
would show they learned something? Enjoyment and confidence are worth noticing, but they
aren't proof.
As you read this, you may already have an idea for a voice-based tool. Hold onto it. You'll be asked about
it next.
Sources
• Alexander, R. (2018). Developing dialogic teaching: genesis, process, trial. PDF
• Bastani, H., et al. (2025). Generative AI without guardrails can harm learning. PNAS. Record
• Bibauw, S., et al. (2022). Dialogue systems for language learning: A meta-analysis. Record
• Bisra, K., et al. (2018). Inducing self-explanation: A meta-analysis. Summary
• Chase, C., et al. (2009). Teachable agents and the protégé effect. Record
• Chi, M., & Wylie, R. (2014). The ICAP framework. Record
• Clinton-Lisell, V. (2022). Listening ears or reading eyes. Summary; (2023) ERIC PDF
• Education Endowment Foundation (2017). Dialogic Teaching evaluation report. PDF
• Graham, S., Kiuhara, S., & MacKay, M. (2020). The effects of writing on learning. Summary
• Howe, C., et al. (2019). Teacher–student dialogue during classroom teaching. Record
• Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning. Scientific Reports.
Article
• Koenecke, A., et al. (2020). Racial disparities in automated speech recognition. PNAS. Article
• Lachner, A., Ly, K., & Nückles, M. (2018). Providing written or oral explanations? ERIC
• Leahy, W., & Sweller, J. (2011). Transient information and the modality effect. Record
• MacLeod, C., et al. (2010). The production effect. PDF
• Putnam, A. (2011). The effects of response modality on retrieval. Thesis
• Reading aloud and comprehension (2023). Memory & Cognition. Article
• Romeo, R., et al. (2018). Beyond the 30-million-word gap. Record
• Rowe, M. B. (1986). Wait time. Summary
• VanLehn, K. (2011). The relative effectiveness of human tutoring. Record
• Wang, F., et al. (2024). Does chatting with chatbots improve language learning? Record
• Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited. Record
• Xu, Y., et al. (2022). "Elinor's talking to me!" ACM
• Plus: PLOS ONE (2021) on social presence in explaining; System (2024) on AI speaking practice;
EdSurge (2021) on conversational agents for children.