Prose has no compiler

In the distributed, fast-paced world we occupy, writing has never been a more important tool to share ideas, align goals, and record events. In my piece On the importance of writing, I propose that writing is the sieve through which you pass the brain slop of fuzzy thinking to clarify and pressure-test it. I have never believed more strongly in that idea.

Meanwhile, an ever-growing array of AI tools is perpetuating the same flawed and dangerous assumption that I saw when I wrote that piece: that writing is a mechanical process that turns existing ideas into an artifact, and that it can be delegated to someone (or something) else.

This is making me angry, and when I get angry, I write.

Writing is the connective tissue of distributed teams, the shared brain of the collective organism that is your engineering organization, and the rudder of your product development juggernaut. The question isn’t whether writing matters (it does), but what’s lost when you stop doing it yourself.

AI writing tools smuggle in a false assumption: that writing prose is mechanical, a machine can do it, and you need only feed in an outline of your ideas. With one click of the mouse, a thorough and fluent document emerges from the other side.

The fatal flaw in that line of thinking is that the value of writing isn’t entirely, or even mostly, within the artifact. The value of writing is created in the act of writing.

Writing is thinking

Bad writing requires little effort. Taking some half-formed ideas and stringing together words that express them is pretty easy, and I regret to inform you that this way of writing is the default approach used by most of the people you interact with every day.

As notable computer scientist Leslie Lamport put it, “Writing is nature’s way of telling you how fuzzy your thinking is.” Decades of cognitive science support the idea that writing is not simply a recording medium, but a tool for thought construction. Amazon famously weaponized this by replacing all presentations with “six-pagers,” which I covered in my previous article.

It’s good that it’s hard. The hard part is the actual work. If you’ve exerted little effort, you’ve produced little value.

When you delegate the part of the process that feels hard and requires effort, where is the value created? Spoiler alert: it isn’t.

Bad prose doesn’t throw exceptions

The value of code is in what it does. If the code does what it is supposed to do, it has fulfilled its purpose, whether it’s legible to a human or not. We have piles of tools to help us write code that both runs and performs the task that we need it to, and the goal of the code is met if all of those tools agree with each other.

Unlike code, writing has no “validation layer” between your fuzzy thinking and your reader’s experience of it. There is no runtime environment, no virtual machine, no compiler. The only error state is that the reader stops reading, or doesn’t understand, and that can happen without you ever knowing.

Most of us have made the mistake of touching the “Cthulu code.” That one neglected area of the system that is critically important and totally works but is written in such a way that it seems almost like magic that it runs at all and nobody wants to go near it because they know if they’re even mildly successful they’ll end up responsible for it until the heat death of the universe.

There is no “Cthulu code” of writing. Writing that is unintelligible to the people who need to understand it, or that is written in such a manner that those people contract sudden, acute cases of narcolepsy when exposed to it, is simply writing that doesn’t work.

Induced demand

In economics, there is a concept called induced demand, which describes the phenomenon whereby an increase in supply causes a decrease in price and an increase in demand. The availability of more supply induces the demand.

This phenomenon can be seen in highway projects. In an attempt to ease congestion, we add lanes to a highway. That increased supply suddenly reduces congestion and mean travel time, which causes people who were using other modes of transport to choose to drive instead. Suddenly, the highway is congested again. The increase in supply of highway capacity induced greater demand to drive on it.

The irony is that while induced demand is well-known, it’s chronically ignored. As a result, our efforts to resolve a problem like highway congestion quite often make the problem worse.

The same thing seems to be happening with AI-generated prose. AI increases the supply of prose and decreases its effort cost; as a result we’re just generating more and more of it. Instead of reducing communication overhead and driving clarity, we’re adding more noise around the signal.

What we’re losing

The best engineering leaders are technical enough to ask really good questions. They don’t know all the answers, but they know what a good answer looks like, and they constantly pressure-test their reasoning. Great leaders can pull the signal up out of the noise.

When engineers and engineering leaders stop writing—themselves, by hand, using their actual brains—they stop pressure-testing, stop challenging their assumptions, stop noticing the cracks in their ideas. Their teams inherit confident-sounding words built on unexamined thinking, published at scale.

AI can write in a tone of your choosing, but it can’t think for you.

What is AI good for?

I use AI every day, and it helps me write these articles. But I write all of these words, I think all of these thoughts, I drag myself through all the editing and deleting and rewriting.

AI is useful as a “collaborator.” Reacting to it, arguing with it, and using it to help organize and summarize. What it does well is the routine stuff I would normally do myself; collecting a bunch of my ideas, listing them out, helping me select which ones belong in the piece and in what order, and suggesting different ways of framing each section.

It’s the equivalent of a freshman-level research assistant brown-nosing their way through an undergrad program. Don’t give it too much responsibility, and try not to listen too intently to all of its praise.

Think for yourself

AI is great at a lot of things: web research, suggesting a pleasing ordering of ideas, being a conversational partner in exploring different ways of approaching your storytelling. But if you use AI to do all the writing, you lose virtually all of its benefits.

It’s tantalizing to ask AI to do the stuff that feels hard, but all of the value created by writing comes out of the hard parts.

Always, always do the hard parts.

Comments