Fooling the Reader

8 min read

In visual effects (VFX), it’s a fairly dominant idea that designers go where the “bar is lowest”. This means they cut corners, but only up to the point where the illusion still holds. The story and the VFX has to be believable. However, not so low they use CGI slop (that was even before AI era) and don’t even bother to make it interesting, like they did in the many shots of The Hobbit: Battle of the Five Armies.

The moviegoer has to believe it was a real explosion, actual blood, and a proper martial arts fight. Often it’s the film editor doing that final bit of convincing, and on set there’s usually someone whose whole job is watching continuity: making sure no Starbucks cup gets left in frame, the way it famously did near the end of Game of Thrones when things got lazy!

VFX takes time to get things right, and movies have budgets and deadlines.

Writing has the same problem. Budgets, deadlines. Great writers know how to work around that.

Which brings us to the actual point: AI is speeding up a lot of things, especially coding. It just makes sense to use AI in coding, at least from a very non-coder perspective; I have created websites, applications, a webapp (TBA), python plumbing, databases, paywalls, and other AI pipelines, scrapers, and this blog. Sky is the limit… Eh! Even the sky is not the limit.

Using AI feels like a game: to maximize the amount of tokens I can use per provided window. You gotta prompt those clankers every day, even before you say your morning prayers. I have no coding knowledge, and yet I am able to do all this. I feel empowered and unemployed.

There’s a multitude of bloggers discussing their experiences of AI. So I had my AI work through it, and analyse the sentiment for each. People have mostly mixed feelings.

Against     ██████████████░░   14
Mixed       ████████████████   16
Accepting   ██████░░░░░░░░░░    6

Against: the writing itself should stay human, tools or not.

dergigi · Front Porch Republic · Oxide · alexop.dev · short.is · storiesby.ai · Sean O’Connor · Karen Smiley · gracefulliberty.com · Craig Mod · lambdaland.org · resobscura · dbushell.com · Ted Chiang, why AI isn’t going to make art (paywalled — discussed here)

Mixed: fine for the scaffolding, not for the sentence.

Simon Willison · Ben Congdon · Daniel Wieser · Charlie Guo · botharetrue · Erik Hoel · tombedor.dev · pnppl.cc · Tim O’Reilly · ia.net · amatria.in · Sidharth Kakkar · jampa.dev · mattsayar.com · rmoff.net · uxcontent.com

Accepting: judge the output, not the method.

noenthuda · octopusyarn.com · Linda Caroll · Katie Harbath · manveerbhullar.com · brevity.wordpress.com

In fact, using AI in writing is close to taboo right now. Code is alright because it can help find security issues before the bad guys. Sometimes the bad guys outrun the good guys though…

But writing? Writing is a (hell) no-no. Practically stigmatised if you ask me! People take pride in saying a piece was completely self-made with no AI anywhere near it. Research, brainstorming, back-and-forth Q&A, sure alright.

But the actual sentences have to come from the author. Some writers and news outlets wear the “no clankers” badge. And honestly, my gut agrees: writing carries a soul (whether or not you believe in souls literally) and it carries a style and a proper timing with the em-dash — and AI just, does, not, have that. (You see what I did there? That’s a style AI rarely uses: a thought parenthesized inside another thought, commas stacked word after word to build rhythm.)

So yes, I agree. Especially among writers who can already smell great writing from terrible, clanker booboo. AI writing is often below-average and predictable. That kind of writing erodes trust instantly: one bad, obviously-AI sentence and the reader’s attention shifts entirely to whether the author’s message is even legitimate. As Oxide, put it: “This can leave a reader frustrated: why should they spend more time reading prose than the writer spent writing it?”

Attention, or more precisely presence is precious and we have only so much of it in our lives before we die. Some people forget to live altogether; we are such busy bees, digging and shovelling, digging and shovelling, digging and shovelling, digging and shovelling, digging and shovelling.

I’ve made my share of mistakes: life is trial and error, and we learn, adapt, and get better. We need to make mistakes with AI so that we can get better with it. That includes “slop” as well, I must say!

Trial and error applies everywhere, love and marriage included. So why do we treat AI slop as such a big deal? Because your boss, coworkers and potential customers won’t forgive you the way your husband or wife would. It’s about money, and about wasting time. That’s why. God forgives too, eventually, and only if you’re willing to accept it, and that’s the hardest part. Can you forgive yourself the way God does? I digress. (AI doesn’t.)

With all that said, the models themselves have gotten efficient and pretty intelligent compared to even a year ago. It’s accelerating.

Writers now lean on things like SKILL.md and DESIGN.md files, AI-SDLC pipelines, and other instructive-contextual files an AI reads before writing, so it stays consistent with a particular goal.

What used to be the word “delve,” or the tic of “not X, but Y,” can now be trained out of the text almost entirely. What used to be flat, generic vocabulary can be tuned to be specific and to actually follow an author’s voice. Things change, but taboos take a long time to change.

The people who experiment with these models, learn the weak points, and know how to use AI to write something that actually fools the reader (remember the VFX point). I believe it’s only a matter of time before that becomes normal. But to do that, one has to be a great writer to begin with, in order to use these tools for one’s own benefit. Otherwise they don’t benefit from AI.

That said, not everyone knows how to do this properly. Imagine you’re the CEO of a company that makes boxes. You have employees who know this and that about the business. Your actual job as CEO is being the lead janitor — cleaning up other people’s messes. That means knowing when someone’s straying off course, knowing the answers to the hard questions, and knowing your business at its core. If you didn’t, how could you lead it? If you knew nothing about boxes, how could you run a box company and actually deliver value?

The same logic applies to writing with AI. Today’s AI systems include “agents”. These are AI setups that don’t just answer a single question but carry out a multi-step task somewhat on their own, chaining several actions together toward a goal.

Say, you ask one to solve some very complex math problem. Like an Erdos problem, which apparently AI has solved those, but I have zero idea what they are, and how to make sure I understand their outputs and even assumptions or notation.

When you read the output and have zero idea what it even means, how can you know whether it’s any good? That’s the first thing they teach in kindergarten: if you don’t know what a word means, don’t say it or use it in your story. And if you can’t tell whether the agent did a good job, how can you be sure the writing is actually good?

By default, it’s complete garbage. But as they say, someone’s garbage is another’s treasure. So it can be made better. (Well, the saying doesn’t imply that.) But only if you know how, and if you can see past the initial allure of the AI. Because if you think AI wrote something brilliant, think twice: it’s most likely just writing better than you would have, because your own baseline wasn’t that high to begin with. Most likely the writing misses context-sensitive knowledge, wording and practises, nuance and so on.

Personally, I always start with a vision and write my own draft notes first. I have a taste for what I like to read, what elements I want to see, what sources to use, and which words I reach for. There are plenty of kinds of writing, of course, and I’ve never touched fiction myself, not even poems or songwriting.

Still, a good artist knows exactly when and where to use these tools to fool the reader. That’s always been the game, really. AI slop is just what happens when people without that prior experience assume and shout “this is great!” the first time a model hands them something ostensibly “polished”. If people, especially in school, never learn to write well and just keep leaning on AI, we get more slop, and that could become a (HUGE) problem. You need the actual experience of writing badly first (and in fact, writing bad slop with AI as well), and reading a lot, to know what good writing even looks like.

That’s how you fool the reader. Writers and artists have always manipulated (with love) the people reading and watching their work. Don’t be afraid of using AI in your writing, but make sure you put more effort and thought into it, before you gave it to AI to do the “grammar check” and more.