The most interesting characters have flaws (and your AI content doesn't)
Language models predict the most likely word. Imperfection is improbable by definition. That's the problem and I'm experimenting with the fix.
I've been playing storytelling RPGs for years. They're not the min-max-your-stats kind where you optimize damage output and armor class, but the kind where the goal is to make your character interesting to play with.
Years ago, I noticed that some players made characters who were perfect in every way. Capable, competent, morally sound. No real weaknesses. I think they were trying to create idealized versions of themselves - compensating for something, filling in gaps, whatever the psychology was.
And these characters were almost always boring to play with.
They were way too predictable. They made the "right" choice every time. There was no tension, no friction, no surprise. You could map out their decisions three sessions in advance.
Then there were players who built in flaws. Real ones. Their characters made bad decisions. Behaved selfishly sometimes. They created chaos and broke things.
These were controversial to play with. Some players hated it. They wanted that predictability, that sense of control over where the story was going. They'd get genuinely frustrated when a flawed character disrupted the neat narrative arc they had in their heads.
But damn, those imperfect characters were interesting. They created the moments everyone remembered and still talk about.
Here's the hard part, though: playing a flawed character is genuinely difficult. You have to make decisions that feel bad. Your character says something cruel, or makes a selfish choice, or behaves like shit - and even though it's just a game and fiction, you're the one making that choice. It creates real discomfort. Most players avoid it because it's easier to be the hero.
This is the main problem with AI-generated content.
It's too perfect.
Not perfect like "high quality." It's perfect like the min-maxed RPG character. Optimized, capable, predictable. Every sentence structured correctly. Every point made clearly. No rough edges, no weird energy, no bad days. It's BORING.
When humans write, we're a mess. That mess shows up on the page.
Research on circadian rhythms and cognitive performance shows that attention, working memory, and executive functions all follow daily rhythms reaching their lowest levels during nighttime and early morning, better levels around noon, and even higher levels during afternoon and evening hours. Your afternoon email is literally a different cognitive product than your early morning Slack message.
But it goes deeper than time of day.
When you're stressed, your writing changes in measurable ways. Studies using linguistic analysis software have found that higher physiological and emotional stress reactivity is linked to lower linguistic cognitive complexity - essentially, the more stressed you are, the simpler your language becomes. Your sentences get shorter. Your vocabulary contracts. The research suggests this happens because acute stress impairs prefrontal cortex functioning, including reduced cognitive flexibility and working memory during stressful tasks. (PubMed Central)
Even your typing patterns change. Research on mental fatigue and typewriting shows that people tend to make more typing errors during prolonged task performance. (PubMed Central) Your accuracy degrades as you get tired. More interesting: younger adults show an increase in typing errors correlated with decreased attention and memory operations as a result of fatigue-related diminished cognitive engagement. (Frontiers)
There's even research from the University of Illinois Chicago using an iPhone app that monitors users' mood and cognition by tracking their typing patterns (University of Illinois at Chicago) - they found a manic episode may be preceded by rising numbers of typos, faster typing, more frequent use of the "delete" key.
Your mental state literally shows up in your keystrokes.
So when a human writes, they're writing with their whole situation embedded in the text. Are you tired? Rushed? Hungry? Did you sleep badly? Are you irritated at something that has nothing to do with what you're writing about? Have you had coffee yet?
All of that creates variation. Imperfection. Unpredictability.
A post you write at 7am before coffee is structurally different from one you write at 2pm after lunch. Not just in quality - in character. The rhythm is different. The word choices shift. The energy moves.
This is the thing AI can't replicate.
Language models work by predicting the most likely next token. They're literally optimized to produce probable sequences. Imperfection is, by definition, improbable. The model will always smooth toward the expected, the normal, the optimized.
It's the min-maxed character every single time.
So I've been experimenting with shitpost senpai.
The question I keep coming back to: how do you maximize humane variation into generated text? Not random noise which is just chaos. But the kind of imperfection that signals a real person wrote this.
I've been adding what I'd call "energy modulation" - ways for the output to shift based on something other than pure optimization. Some of it is tone variation. Some of it is structural choices that aren't the "best" choice but are an interesting choice.
And starting today: typos.
Not a lot. Not random keyboard smashing. But the kind of small errors that real people make when they're typing fast, or tired, or distracted. Weird abbreviations. The occasional letter transposition. The stuff you'd normally autocorrect out.
The stuff that makes text feel like it came from fingers on a keyboard attached to a brain that's also thinking about whether it's too early for a second coffee.
Here's what I've learned from RPGs: perfect characters aren't just boring to play with - they're also boring to watch. The audience checks out. There's no tension. No stakes. No sense that anything real is happening.
Flawed characters make you lean in. You don't know what they're going to do. Their decisions matter because they might be wrong.
I think content works the same way.
Perfect content is background noise. It's the brand-safe, SEO-optimized, thoroughly-edited wall of competent text that your eyes slide right past.
Content with flaws - with weird energy, with rough edges, with the fingerprints of an actual human state embedded in it - that's the stuff that lands.
Let's see if the same is true for AI.