Aubrey Codes
EssayPublished · 7 min read

Writing

There will be slop.

But humanity's best writing will use AI

I have been using AI in my writing for years now. It started with editing and then grew to planning, structuring, outlining, and refining ideas. I tried for a while to get it to write the prose for me as well. I spent countless hours struggling to get it to recreate my thought process and voice. Eventually, I gave up. AI isn’t good at writing; it’s good at writing support. I write all of my own prose now.

Substack recently announced its Pangram AI detector integration, and the platform has been abuzz with everyone’s opinions. A sentiment I have seen repeated often is “If it is AI-generated, I’m not interested.”

This is a sane response to AI-slop, but slop is not all that is being produced by AI.

Since the early days of the internet, there has been no shortage of shallow, stupid, shitty writing. Online forums have been, and probably always will be, clogged by it. Tumblr encouraged it. Amazon has long allowed anyone who is willing to put in a modicum of effort to self-publish their drivel and make it available to the mass market.

AI mimics the bland body of the bell curve of its training data. The impact of outliers on the logic, style, and even the consideration of what is worth thinking or writing about is insufficient to overcome the insipid majority. The prose that AI produces by default is banal. Unremarkable. Unworthy of our attention… but the same is true of the default prose of humans.

AI didn’t invent stupidity or bad taste; it has just made articulating it nearly frictionless.

While AI has opened the floodgates for slop, it’s also reduced the barriers to excellence. The floor is lower, but the ceiling is higher. More slop and more beauty are emerging together.

AI-generated and AI-assisted writing are not the same thing. Not the same practice, not the same product. You shouldn’t need much convincing of this.

AI makes research, opposition, synthesis, experimentation, and iteration faster, easier, and cheaper than ever before. Writers striving for excellence can make broader and deeper attempts.

For some writers, the effort to construct, scrutinize, and refine an argument is enough that by the time they are sure of their position, they’ve got nothing left in the tank to try and communicate it.

In reality, completing a project sooner is often the difference that means completing it at all. And even for projects that would have been completed either way, earlier completion has a cumulative effect at the level of collective endeavor.

The causal argument is this:

  1. AI reduces the cost of writing support.
  2. Some ambitious writers reinvest time, effort, money, attention, or capacity.
  3. Reinvestment permits more numerous, deeper, or broader attempts.
  4. Those attempts must maintain sufficient variation.
  5. Competent human judgement must select, redirect, or reject them.
  6. Faster achievement permits faster cycles and cumulative progress.

The most common opinions I have seen, while hostile to fully AI-generated writing, are much more accepting of writing that is only AI-assisted. But there are of course some opponents of even this more moderate and effortful approach… some objections that deserve to be answered.

More attempts to write something useful don’t help if they are effectively the same attempt. AI homogenizes premises, structures, evidence, and language.

We now have access to a level of support and depth of knowledge that was previously accessible to comparatively few. AI is interfacing with and operationalizing largely the same core knowledge base that writers were using before, but the quantitative change in mediation, availability, and shared framing is so immense that it produces qualitative cultural effects. AI is a human product, a grand synthesis, but it creates a bottleneck. Meaningful homogenization is happening, but sufficient variation remains. Each writer’s context is unique. Different situations, purposes, audiences, lived encounters, and frames of reference. We still have to practice selective adoption, judgement, discerning composition, and revision.

Delegation weakens our judgement.

We’ll start with the easy layer. Delegation doesn’t invariably weaken the product of a collaborative endeavor; it often strengthens it. Used well, AI can strengthen a writer’s judgement. It can increase our ability to compare alternatives, answer objections, identify conflicts, detect failures, reject attractive errors, and redirect exploration. It doesn’t matter if AI is the source of these corrections. Judgement doesn’t need to precede AI output to regulate it.

AI does facilitate passivity, avoidance, and self-confirmation. To answer this, we don’t have to totally prevent these failings. We just need enough competence among enough writers to reduce their frequency. We have to make increased contact with reality and corrective capacity the criteria for competent AI-assisted writing.

The harder layer is abstraction. When we abstract away the components of a process, we become less practiced at doing them manually. This is effectively unavoidable. It’s a gamble that we have taken countless times and will have to take countless more. We don’t know what will be required of us in the future, as individuals, as a society, as a species. We take our best guess, we aim at it to the best of our ability, and we take a swing. If it works, we start streamlining the process. Once it’s streamlined, we slowly become less capable of doing it manually. Rinse and repeat ad infinitum. You can object to this process, and insist correctly that there is a chance that society will collapse, and only the doomsday preppers and bushcrafters will have been right about what to spend their time and resources on. Fine. I’m placing my bets elsewhere. Let’s not be friends.

Slop cheapens the whole game. We can’t communicate if we can’t hear each other over the noise.

Maybe. We do have to mitigate slop. If we let it destroy our culture’s ability to find or share excellence, we’re lost. The good news is that slop doesn’t have to disappear; it only has to be managed. As long as we have editorial communities dedicated to excellence, provenance and disclosure practices, discerning readers, and incentives for distinction, we’ll find our way through. We have those things. The internet is already burdened with and adapted to low-quality material. We just have to stay sharp and skeptical.

Some objections aren’t about whether AI assistance can be good or useful. Some people want to feel personhood in the writing. They want an unadulterated representation of someone else’s thought. They imagine that thought somehow wasn’t already intractably responsive. They are mistaken. Our thoughts can’t escape their inheritance. What matters is that we now have a new synthesis to respond to. AI can’t reliably mimic human point of view. Some who are trying primarily to produce goods built on that point of view probably know or suspect this. Some of those who don’t will quickly realize it. Some won’t. There will be slop. But we balance it against the potential for explosive growth in other kinds of writing. Refusing important, true, beautiful, powerful, or useful work solely because it is AI-assisted means you are judging process purity over the quality of the work itself. Were this to become the consensus view, it would strangle the life out of writing just as readily as unmanaged slop and process homogenization.

I’ll summarize: For things to go well, we need smart, discerning people, dedicated to excellence, and a sufficiently healthy cultural ecology. We need those people to use AI assistance well, and we need to limit how much people misuse it. If we can make those things happen, they will accumulate into better writing. We are headed that way right now. Whether or not things continue in that direction is partly up to chance, or fate, or God, whatever your preference… but it is also up to us.

Rejecting or screening AI-generated writing is a reasonable heuristic. Figuring out if something is worth your attention is effortful. So filter the slop, but don’t let that judgement extend to all AI-assisted writing.

AI, used competently, is a force multiplier. We are building something together. What we build is made more sophisticated by the participation of more competent contributors. It is strengthened by more robust support structures. It’s streamlined by accessibility. These differences will eventually accumulate into the most effective collaborations, the most useful models, and the most enduring truths. If you refuse to read anything AI-assisted, you are inevitably going to filter out indispensable insight. Humanity’s best writing will use AI.

More to read

Keep going

Where the software comes in

Some of the same questions show up in the software. Oathcraft matters here because it keeps the writing honest. Ideas that sound persuasive on the page still have to survive contact with real constraints.

See Oathcraft