When AI replaces "the struggle of writing" where thinking happens

Anthropic recently released what they’re calling the largest qualitative study of AI users ever conducted. While the researchers were likely looking for a laundry list of innovative use cases, the feedback they actually received was much more raw and honest. They talked to over 80,000 Claude users about how they’re actually living with these tools—what they’re hoping for, and what they’re starting to worry about.

A few of those responses really stopped me in my tracks. One heavy user in the US admitted, “I don’t think as much as I used to. I struggle to put the ideas I do have into words.” Then there was a freelance software engineer in France who noted that the promised productivity hasn’t actually given him more rest; he’s just having to run faster and faster to stay in place. These aren’t skeptics; these are power users who benefit from the tech every day, yet they can feel something essential quietly eroding. It’s a realization that didn’t fully hit me until I sat down with the data.

I’ve come to suspect that those of us in communications are more vulnerable to this than almost anyone else in knowledge work. I might be wrong (that’s my standard caveat, obviously!), but this research struck a chord because it suggests that AI might be hollowing us out by degrading the thinking processes that make our work valuable. Those productivity bumps we’re all celebrating might actually be covering up a serious loss.

I’ve been stuck on a question lately: when you ask an AI to draft a sensitive message about a restructure, or prep a CEO for a high-stakes all-hands, and the output is polished and “on-brand,” what did you actually learn about the situation that you didn’t know before? The reality is often that we haven’t actually learned anything, because the machine did the heavy lifting for us. That subtle, empathetic reflection that comes from searching for exactly the right word for a specific audience has been stripped away.

I’ve always felt that the struggle of writing is where the thinking happens, so I went looking for a framework to explain why that matters.

I found some research from 1987 by two cognitive scientists, Marlene Scardamalia and Carl Bereiter, that really should be required reading for anyone using AI professionally.

They identified two totally different ways we write:

• Knowledge telling: Essentially just transcribing what you already know.

• Knowledge transforming: A back-and-forth process where the act of putting an idea into words actually changes your understanding of the idea itself.

You start a paragraph, realize your logic is full of holes, hit delete, and try a different angle—that’s where the actual understanding gets built. I see this in my own work all the time;

I genuinely don’t know what I truly believe until I’ve wrestled with the words for a while. If you skip the journey, you never actually get to the idea.

When I started digging into recent data on how AI affects this, the results were pretty sobering. I found a paper from January 2026 (likely this one or the related Cognitive Atrophy Paradox) that pointed out a paradox: the technology accelerating our productivity is simultaneously threatening the cognitive skills that made us effective in the first place. I also tracked down a CHI 2025 study that surveyed over 300 knowledge workers; nearly 80% of them reported that AI reduced their effort for comprehension, analysis, and synthesis.

What’s even more interesting is that the more confident a user was in the AI, the less critical thinking they applied to its output. We’re making the thinking easier, and as a result, we’re making it shallower. It’s a trend that maps onto the 2014 study on handwriting versus typing, where the slower, more demanding method led to deeper conceptual understanding. The mechanism is identical: cognitive effort during production leads to depth of comprehension.

We also have to be honest about our own tendencies. When was the last time you read an AI draft and thought, “I need to scrap this entirely and start fresh”? I’ve caught myself avoiding that because the output is usually “good enough,” and the moment you start editing it, you’re already anchored to the structure the machine provided.

I looked into this and found that it’s a documented psychological trap called automation bias. Raja Parasuraman and Dietrich Manzey published the gold-standard model for this in 2010, showing that we build trust in automated systems based on experience. The more a system performs acceptably, the less you bother monitoring its work. It’s not laziness—it’s a rational search for efficiency—but it means your personal quality bar starts to sag.

There’s a specific frustration that comes with using AI output: it’s hardly ever “wrong” anymore; it’s just imperfect. And being “almost right” is actually much harder to fix than being flat-out wrong. When something is a disaster, you toss it and start over. When it’s 80% there, you’re stuck trying to remodel a house that’s “adequate” but lacks soul.

I found real-world evidence of this in a 2024 Science Advances study. Researchers looked at writers and found that while AI helped less creative people get better, it also made everyone’s work start to sound the same. The writers were “anchored” to the AI’s output. They individually got better while collectively losing their unique voice.

The time we supposedly save often just gets shifted from creating to correcting. I looked at some adjacent fields to see if this was happening elsewhere and found a 2025 study from Faros AI of 10,000 software developers. AI-assisted teams finished more tasks, but they created 9% more bugs, and the time they spent on code review shot up by nearly 91%. Another RCT by METR found that developers believed they were 20% faster, but they were actually 19% slower. That gap between feeling and fact is something we need to pay attention to in communications, because we haven’t even checked to see if our own gains are real.

Finally, I’ve been trying to articulate why AI speech often sounds so off when read aloud. Having spent years writing for both print and the stage, I know how different those modes feel. I dug into the work of linguist Wallace Chafe, who showed that spoken and written language have totally different registers.

• Spoken language is fragmented, built from shorter ideas, and made for the ear.

• Written language is integrated, dense, and designed for the eye.

Douglas Biber’s 1988 corpus analysis confirms that experienced communicators navigate this multi-dimensional spectrum instinctively. AI, however, tends to spit out text in one single, flat mode. It looks fine on a screen, but the second a leader stands up to deliver those words, the audience can feel they weren’t written for a human breath. The sentences are too long, the vocabulary is too technical for the ear, and there’s no rhythm. You only learn to spot that mismatch by doing the work yourself. If we stop building that muscle through the struggle of translation, we won’t even realize when the machine is getting the tone wrong.

When you stack all of this up—skipping the thinking, trusting the output too much, and getting anchored to “good enough” structures—it creates a foundation that’s thinning right beneath our feet. A benchmark called the Remote Labour Index found that frontier AI systems could only complete about 2.5% of real-world freelance projects to an acceptable level. So while the profession looks intact from the outside, the core expertise is under threat.

Anthropic called this the “light and shade” of AI. The very thing that helps us is slowly diminishing us. The central question for us is whether we’ll still have the ability to do the work that truly matters by the time we finally wake up to what we’ve lost.