I barely type anymore. Not to my AI, and honestly not to much else either. I talk. To the Claude desktop app with its built-in voice integration, and to a handful of dictation apps that work across my whole computer, in every text field there is. I spoke most of this article too, walking around the room. My keyboard is still there. I just reach for it less and less.
This is not a party trick. It comes from a conviction I have carried for a while: a human being is made to speak, not to type. We learn to talk without a single lesson, years before we get one letter onto paper. Typing is a learned detour, an interface we accepted because the machine wanted it that way. The moment the machine can listen too, that detour falls away.
My hunch gets a name
For a long time this was mostly a feeling. This week Allen Pike of Forestwalk Labs gave it language, in a talk titled “Voice In, Visuals Out”. His argument, building on an idea from Andrej Karpathy: speaking is the richest and fastest way a human gets something into a machine, but the answer usually should not come back as speech. It should come back visually. A screen, a list, a mockup, a chart. Something you take in at a glance and can act on straight away.
That matched exactly what I notice in practice. Talking to an AI that speaks everything back is tiring, you are stuck inside the slowness of sound. But talking to an AI that shows you what it is doing while I talk, that feels like working together. Pike names the trap well: do not build a talking text box. The voice is for getting your intent out. The screen is for making it precise.
Talking to it, and talking with it
There are two things in here that I confused for a long time. There is talking to my AI, and there is talking with my AI.
To it is dictation. I dump a thought, a paragraph, half an idea, and it lands as text. One direction, fast, no waiting. That is how I write. With it is a conversation, back and forth, where the model comes back with a question or a pushback. That is slower and more demanding, but sometimes that is exactly what you want.
The difference sounds small but decides everything about the experience. Most voice features treat both the same, as if I always want a conversation. I do not. Most of the time I just want my voice turned into text with nothing talking back.
Where it breaks
I am not going to turn this into a success story, because voice is nowhere near finished.
The biggest problem is speed. Pike puts hard numbers on it: under seven hundred milliseconds it feels fluid, above that it turns clunky. I feel that every day. Half a second too much and I have already lost my sentence. My head disengages and I reach for the keyboard after all.
There is also the fact that talking to your computer is still socially awkward. In a quiet office or a train I will not do it. And the moment something has to be precise, moving a word or pointing at an exact spot, the mouse still beats my voice without effort. Voice is strong for the broad stroke and weak for the detail. Then there is the recognition itself, which still misses on names and jargon, precisely the words that matter most.
Voice-first as a view of people, not a gadget
Even so, for me this is more than a handy way of working. It has become a design principle. In my work on public innovation I now think about the voice first for everything I come up with: can this also work without a keyboard, without someone having to learn to type or navigate first?
That is not a technical question, it is a human one. Voice-first means the barrier sits lower for someone who is not handy with a screen, or who speaks another language, or who needs their hands for something else. Human-centered is not a coat of paint you brush on at the end. It starts with the question of how a person naturally communicates, and that is with their voice. The technology can bend toward that, not the other way around.
The bill that comes with it
There is a price I do not want to paper over. Almost everything I say goes, with most of these apps, to a cloud service that turns my speech into text. That is a constant stream of my voice and my ideas, sometimes half a work conversation along with it, to a server I do not own. When you type you choose your words. When you talk more leaks out, the hesitation and the aside included.
There is local dictation that runs on your own machine, and it is getting better fast, but it has not caught up with the cloud yet. So I make the call moment by moment: sensitive things I still type, or I speak them into a model that runs locally. And I now lean heavily on one voice stack, which means I am exposed if it gets more expensive or disappears. That is the flip side of something that works so smoothly you no longer want to be without it.
At the same time there is a gain I do not want to underrate. For someone who struggles to type, this is not convenience but access. That side weighs heavily for me, and it is exactly why voice-first is not a fad to me. It is a direction.
Try it yourself
If you take one thing from this: pick a day where you do not type to your AI, but talk. Turn on the dictation feature you already have, and just speak your next message. Notice where it flows and where your hand crawls back to the keyboard anyway. That line tells you more about how you actually work than any blog post can.
I would love to hear how it goes. Find me on LinkedIn.
This piece also appeared in Dutch: Ik typ bijna niet meer tegen mijn AI.