Narrate First, Polish Later

What Voice Dictation and AI Are Teaching Me About Thinking

By Erika Sanneman

Sometimes I have to stop the AI and say that I haven't actually written anything yet.

I've brought it an article, a connection to something from class, or the beginning of an idea I want to explore. And suddenly there's a finished-looking post in front of me. It might sound reasonable. It might even contain things I agree with. But agreeing with a paragraph isn't the same as having arrived at it myself.

I rarely begin with a perfectly composed prompt. I talk. I interrupt myself, add the thing I forgot, and sometimes realize halfway through that the real point is somewhere else. Then I read what comes out of that narration, and I talk back to it.

I'm finding that trying to write clearly too early can get in the way of figuring out what I think. Speaking gives me room to get more of the thought out before correcting how it sounds. AI helps me make that thought available to examine. The difficult part is making sure the examination still happens when the writing already looks finished.

Letting the thought arrive

When I type, I'm often editing before I've finished thinking. I qualify a sentence or fix a word while the next connection is still trying to arrive. Sometimes I lose that connection because I'm so busy communicating the first one.

Speaking gives me another way in. I can narrate while I'm cooking or walking, while an idea is still there. A few minutes might contain a story, something from class, several questions, and a contradiction. I don't always know which part matters most until I've said more.

I've been told many times that I spiral toward a point. I have worked on being less repetitive, and I want the writing to be clear. But some of that movement is how I find the point. The tangent might explain why an example matters. The sentence I correct halfway through might be the moment I notice an assumption I hadn't questioned.

I've tried journals, notes, and other ways to collect those thoughts. I keep falling away from processes that ask me to stop, type, find the right place, and organize everything. Voice input removes some of that friction.

And that matters in the life I'm actually living. I'm a single mother of two small children. I'm working, pursuing a master's, building a company, and protecting time for my health. Capturing a thought as I move through my day gives me something I might otherwise lose. It also gives me the chance to return, question it, or discover I no longer agree with it.

Making the thinking visible

Once I've narrated, I want help seeing what is there. Which ideas belong together? Where did I start one thought and finish another? What have I assumed a reader already knows?

Usually I read with the microphone open and respond as I go. I might explain why a sentence doesn't sound like me, add an example, or stop because a connection reminds me of something from a book or class. That response becomes more material to work with.

This article itself went through about seven drafts before reaching this version. It began with multiple voice memos, recorded and transcribed at different points over two full days, drawing together ideas and thoughts I'd been developing over the past month and a half and learning from AI courses I took several years ago. Many of those drafts involved me adding more voice notes, corrections, suggestions, and small verbal tweaks alongside the AI's edits. Reading each version gave me another chance to hear what was missing or what I hadn't quite meant.

While developing this article, I asked for the context behind my answers to be woven back into the piece. A draft can include what I've said and still lose why I said it. For example, describing my busy life helps explain why speaking is a useful way to capture ideas. Without that connection, the same details can sound like a defense of using AI.

I don't need the article to reproduce our whole conversation. I need it to carry enough of the meaning for someone who wasn't there to follow it.

The danger is the polishing

AI can turn a messy collection of thoughts into prose that sounds settled. Sometimes that makes the argument easier to see. Sometimes it hides the fact that I haven't settled it.

An awkward sentence may contain a distinction I haven't learned to express yet. An unanswered question may be something I'm still sitting with. A repeated idea may return because each pass adds something. Removing all of that can leave me with a cleaner article and less of what made it worth writing.

I still want editing. Meeting a word limit while keeping the main point and a metaphor I love can take me a very long time. A shorter version gives me something concrete to respond to: this part needs to come back; that can go; this metaphor can become tighter, but the relationship it explains needs to survive.

What I'm trying to protect is the order of the work:

Because AI is extremely good at sanding off the weird little edges, those edges are often where voice, uncertainty, humor, contradiction, and genuine learning live.

So for me the process becomes:

Narrate → notice → organize → question → revise → polish.

Not:

Prompt → generate → publish.

I move back through those steps as needed. Reading a polished sentence can send me back to narration because I realize I haven't explained something. If I had more time, I would do more of the editing manually. I would still want the spoken brainstorming.

Giving the tool boundaries

My approach has been shaped by the Acosta Institute AI Masterclass, a Google AI course, reading, and LinkedIn conversations about responsible use. Alongside those experiences, undergraduate philosophy coursework has given me questions about technology, efficiency, exploitation, the environment, and who bears the cost of what we call progress.

A responsibility checklist from my Google learning gives those questions a practical application: review and personalize outputs, disclose assistance, follow applicable organizational policies, protect private information, and consider accuracy, bias, and effects on others.[4]

I use different configured spaces for educational consulting, restorative-practices coursework, business content, other online work, and parenting. Each involves different goals, voices, responsibilities, and information.

For coursework, I read the assigned material and prompts, then narrate my responses to the reflection questions. I've set instructions asking the tool to question me when information is missing instead of supplying a plausible account of what I must have thought or experienced. A missing answer may represent work I still need to do.

I disclose and cite AI assistance in my coursework, and I've discussed the process in my courses. But the parameters don't always hold. Sometimes the tool drafts before I'm ready or brings in an assumption I've already challenged. Instructions give me something to check against; they don't replace that check.

Being honest about the contribution

Much of this process begins with my thinking, examples, and connections. But describing AI only as a transcription tool would leave out some of the help I've received.

During a change exercise for coursework, I spent roughly an hour and a half to two hours answering questions and working through the framework. Once I laid out my answers, AI suggested connections I hadn't noticed. Some resembled observations my therapist and two close friends had made, even though I hadn't told the tool what they had said.

That felt meaningful to me, and it complicated how I described the process. I supplied the experiences and answers, but I didn't independently generate every interpretation.

Does a suggested connection fit what I know? What might it be missing? Am I accepting it because it feels familiar, because it flatters an explanation I prefer, or because I can account for it with evidence from my experience? My professional knowledge helps me catch some omissions and oversimplifications. It doesn't let me catch every one.

Keeping the learning in the work

That distinction matters when I think about students. I want them to learn to challenge, correct, contextualize, and evaluate what AI produces. Calling it a thinking partner leaves an open question: which parts of the thinking are still theirs, and what can they explain, justify, or do independently afterward?

The purpose of the task matters. If we're trying to understand a student's reasoning, spoken explanation might help us hear more of it. If we're teaching organization, having AI supply the entire structure may remove the very work we're trying to develop. If we're teaching source evaluation, an impressive-looking citation is the beginning of something to investigate.

In a high-school mathematics field experiment, Bastani and colleagues found that AI assistance improved practice performance. Still, students using the less restricted version performed worse than the comparison group once the tool was removed. A version designed with safeguards to support learning largely mitigated that harm. This was a particular mathematics setting, but it gives us a reason to examine assisted performance and learning separately afterward.[1]

For me, that brings the question back to design. What support helps a learner stay engaged with the difficult part? What makes it easy to bypass it? I want the same scrutiny for my own coursework. A finished assignment isn't enough evidence that I've understood the reading or revised my thinking.

What our language asks us to believe

The language I use while narrating deserves attention, too. I say “you,” ask questions, and describe what I want the tool to help me do. That conversational interface is useful. It can also encourage assumptions about what the system understands.

Words such as “thinking,” “intelligence,” and “superintelligence” carry expectations. Does a fluent response establish understanding? What is being claimed when a system is described as more intelligent?

In a position paper about natural-language understanding, Bender and Koller argue that learning from linguistic form alone does not establish meaning. Their argument doesn't settle every question about current AI systems, but it offers a reason to be precise about what convincing language demonstrates.[2]

An interpretation that resonates with me doesn't establish that the tool understands me as a friend or therapist does. I can value the conversational way I work while remaining careful about what I infer from it.

Systems that make room for people

These questions connect to something broader in my work: systems that make people's lives function better and reduce unnecessary pressure. AI helps me capture, organize, and return to work competing for limited time and attention. But what is the time saved being used for? Does it make room to think, learn, rest, or be more present? Or does it make room for more demands?

I've raised a related concern in conversations about a shorter workweek. Compressing the same expectations into fewer days doesn't necessarily create a more livable working life. I think we need to examine the conditions underneath promises that AI will save people time, too.

For educators, I can imagine a teacher capturing observations after a lesson, a coach talking through what they noticed, or a leader recording questions after a meeting. These are possibilities to design and evaluate. In narrative reporting, spoken observations might preserve details about a child that are hard to capture in a checkbox. A tool organizing them would still need to separate observation from interpretation, preserve uncertainty, protect privacy, and avoid inventing what happened.

Implementation also needs opportunities to see responsible use modeled, practice it, receive feedback, and revise. Otherwise, we risk holding individuals accountable while leaving the conditions shaping their work unexamined.

UNESCO's guidance likewise calls for an approach centered on people, with attention to privacy, age-appropriate use, and ethical and pedagogical validation.[3] I read that as a reason to examine the setting in which we're asking people to use a tool alongside what it can produce.

Letting verification change the argument

Source work is one place where these principles get tested. Often I bring the direction: a book, poem, article from class, or connection to something I've read. Sometimes I remember the idea but not the author, and AI helps locate the reference.

While working on this article, I realized I'd been speaking about verification more confidently than my habits justified. A familiar quotation, a plausible author, and a properly formatted citation had sometimes been enough for me to accept it without reopening the original.

That needs to change. I want exact quotations checked against accessible originals, with their location and context recorded. When something can't be checked, I want the uncertainty visible.

And when a passage doesn't support what I remembered, I want to pause before searching for another quotation that does. Did I misunderstand it? Combine it with another idea? Remember the part that suited my argument and lose the complication?

There is learning in that mismatch. I want research to be able to change what I say.

Values have to show up in decisions

My master's coursework in restorative justice also shapes how I approach these questions. It asks me to consider how our decisions affect ourselves, other people, and the environment, both directly and indirectly. To me, being restorative means acknowledging those effects, taking accountability, minimizing the harm we cause, and working to improve the world around us. That responsibility includes consequences we don't immediately see.

That is part of why I think using AI requires attention to its environmental costs, including water use, alongside questions about bias, privacy, labor, and access. Finding a helpful workflow doesn't resolve those concerns. Neither does completing a course or paying for a subscription. I need to keep asking what my use contributes to, what harm I can reduce, and where my choices need to change.

I pay partly so I can configure a process and return to work without rebuilding the context so often. That's a practical decision about my working conditions. My privacy boundaries still need development: I avoid concrete financial details, but I don't yet have an equally clear practice for every kind of information.

Ethical questions also appear in the choices the tool helps me consider. As a white-passing Latina, I participate in some LinkedIn conversations that center Black experiences. I can care about a topic while needing to consider the invitation, the purpose of the space, and what happens to its focus when I enter.

In one exchange, I received appreciation for what I brought and a request to leave room for people with particular identities. I left my comment in place because the person had also engaged positively with it, but stopped extending the conversation. Both the appreciation and the boundary mattered. My explanation doesn't tell me how everyone experienced my response.

AI cannot speak for those people. I've explicitly asked it to question me if I seem to be rationalizing a choice. I would rather be asked to look again than receive a polished explanation that makes my decision easier to defend. That is part of what I want from revision, too: help noticing the question that might complicate what I've already said.

Leaving space for what comes next

I love dialogue. LinkedIn sometimes feels like a delicious dinner party where the conversations are asynchronous. I can return to an exchange, build on a thought, or notice a connection I missed. I want my writing to leave room for that.

I still have questions about my own growth. I don't know how much comes from AI, how much from reading and coursework, or how much becomes visible because I can express and revisit more of my thinking. I wonder about voice, too. After years of editing out sentences beginning with “And,” I'm leaving more of them in work that begins as speech. How much does the tool preserve my patterns, and how much does it amplify them?

What I do know is that narrating first lets me keep more of a thought long enough to work with it. I want that support to reduce the friction around thinking while leaving me engaged in the thinking itself.

When you try to put an idea into words, where does editing help you understand it, and where does it interrupt a thought that hasn't finished arriving?

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122.

Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463.

Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO.

Google LLC. (2024). Responsibility checklist. Course resource from my AI learning; the course title is not printed on the document.

A note on AI assistance

I developed this article through spoken reflection and ongoing revision with ChatGPT. I supplied the personal experiences, questions, and much of the direction; ChatGPT helped organize the material, draft and revise passages, suggest connections, and locate and summarize research. My corrections and follow-up narration shaped the revisions. Its contribution went beyond transcription. I remain responsible for the claims, source use, and wording I choose to publish.