Why AI shouldn’t replace rough drafts

I consider writing quite a bit simple. For me, it’s a transfer and organizational process, a way to put ideas out of my head and on the page.

However, fascinating research suggests that the way thinking externalizes can change the nature of thought itself and how well we remember them. It made me think and I could write about it.

This research, Published in Consciousness and cognitionwe compared two simple tasks, including writing words with the hand and drawing pictures of those words. On the surface, both methods may seem equally effective for learning. However, the researchers found something very interesting: drawing has been consistently better Memory Performance rather than writing.

As you can imagine, I thought about LLMS and entered it into its context window and entered it as the central tool for “thinking” these days.

Cognitive Benefits of Drawing

This study used a method called multidimensional experience sampling. This interrupted participants from asking about the “quality and focus” of their ongoing thoughts during the task. This allowed them to go beyond the final written words and sketches, and to see the mental space where work was taking place.

When people painted, their ideas tended to be more visual, more elaborate and more connected. They not only transcribed meaning, but also constructed it in multiple modes at once with a triad of visual images, motor movements, and spatial relationships. These rich thinking patterns supported a strong recall later. In comparison, writing tended to stay more linear and verbal, and to activate narrower slices of mental abilities.

From pens and paper to AI keyboards

This study did not mention it artificial intelligenceHowever, its meaning is difficult to ignore in an age where our “writing” is increasingly large. Co-authored with large-scale language models. A leap from pen to keyboard could already remove learning from durable sensations and some of the motor loops, and a leap from keyboard to AI output risks further elimination. At least that’s my hypothesis based on the findings of the research.

With LLMS, it is often possible to completely bypass the slow and messy generation stages of writing. Instead of assembling thoughts through trials, errors, and revisions, you can feed the system prompts and receive almost complete blocks of text in seconds. Our roles transition from creator to selector. The selector scans, tweaks and approves drive content rather than building from scratch.

Danger of skipping the “create” phase

Drawing studies suggest that this shift is important. Elaborate multimodal thinking patterns that enhance memory come from creation, not only evaluating them.

When AI takes over the creation, mental traces and paths fade. Words exist, but they live outside of us, and are stored in the output of the machine rather than embedded in our own cognitive architecture.

In my view, this is not a discussion of AI use. It’s a discussion to use it with more intent.

If the benefits of Drawing come from engaging the mind across multiple channels, AI workflows may need to create room for such engagement. That means sketching the outline before prompting, schematizing the relationships of ideas, or freewriting a rough draft before asking the model to improve it.

Designed for cognitive abundance

The key is to preserve or further amplify the friction that AI produces very efficiently. That friction is where deeper connections form. It’s where the idea is pushed, pulled and sometimes broken before it is rebuilt into its final form. Without it, we risk producing beautiful, but rather vague pieces on the page.

Of course, we have known for a long time that the tools we use can change how we think. Quills, typewriters and word processors all leave marks on our mental habits. And today, LLM is just the latest tool. However, they arrive at speed and flow ency that may short-circumvent the mental process of clinging to knowledge.

My sense is that this research provides an important reminder that the process is just as important as the product. If AI helps think rather than replacing it, then interactions should be designed so that they remain active participants, not just passive recipients of the results. This means keeping our hands and minds in work, even if the machine can do better (in some cases) faster.

And the key point here is that memory and understanding are not built into the output. They are made.