You know the feeling. You spend 30 minutes crafting the perfect prompt, feed it into your AI, and get back something that reads like a corporate memo written by a committee of robots. You sigh, then spend another 30 minutes editing out the pomposity, the empty phrases, the ‘load-bearing’ and ‘genuinely’ that infest every paragraph. And somewhere in the middle, you realize: you could have just written it yourself.
This isn’t a prompt problem. It’s not a fine-tuning problem. It’s a fundamental design flaw in how language models are built.
Your AI isn’t bad at writing. It’s trained to be average — and it’s terrifyingly good at it.
Here’s the uncomfortable truth that most AI boosters won’t tell you: the training objective that makes models fluent also makes them mediocre. These systems are optimized to predict the most statistically likely next token from trillions of examples. The most likely word is the safest, blandest, most generic choice. Creativity is the opposite — it’s the unexpected, the deliberate violation of probability. So the model inherently resists voice, style, and personality.
You’ve probably noticed the symptoms. The pompous tone. The endless qualifiers. The eerie absence of any real point of view. That’s not a bug you can fix with a better prompt — it’s the model doing exactly what it was trained to do.
The architecture that makes AI fluent is the same architecture that makes it a mediocrity-generating machine.
Think about it. Every time you ask for prose, the model averages over every author it’s ever seen. It becomes a blend of everyone — and therefore, no one. It’s like asking a committee of a million writers to produce a single sentence. The result is technically correct, grammatically flawless, and utterly lifeless.
This is why the HN user who posted ‘Ask HN: How to make AI better at writing prose?’ is right to be frustrated. He tried feeding examples, style guides, iterative editing — and still ended up writing it himself. Because the problem isn’t on your side of the screen. It’s in the training data and the objective function.
Now, here’s the twist: This isn’t a bug. It’s a feature. The models are designed to be safe, not surprising. They are optimized to never offend, never take a controversial stance, never break a rule. That’s exactly what you don’t want for compelling prose. Prose requires risk. It requires opinion. It requires the willingness to be wrong.
So what do you do? Stop trying to force a general-purpose LLM to be a prose writer. Use it for what it’s good at: generating outlines, summarizing research, suggesting alternatives, breaking through writer’s block. But write the final draft yourself. Or, if you must, use a small model fine-tuned on a curated collection of high-quality literature — something that has a defined voice, not a trillion-voice average.
The best AI writing tool isn’t a better model. It’s knowing when to turn it off.
That’s the real lesson from the state of AI prose. The technology is powerful, but it’s power lies in amplification, not creation. It can mimic structure, but it cannot invent soul. If you want writing that feels like a person, you need a person. And that’s not a limitation — it’s a reminder of what makes us human.
FAQ
Q: But aren't models getting better? Will they eventually write well?
A: Not unless the training objective changes. Current models are rewarded for being safe and predictable, not for taking risks or developing a distinct voice. As long as they optimize for average, they will remain mediocre at prose.
Q: So what should I do if I need AI to write?
A: Use AI for structure, research, outlines, and brainstorming. Write the prose yourself. For specialized tasks, consider fine-tuning a small model on a curated dataset of high-quality writing that matches your desired style.
Q: Maybe the problem is that we're expecting AI to be creative when it's just a tool for efficiency?
A: Exactly. The hype has sold AI as a creative partner, but it's really a pattern-matching machine. That's useful for many tasks, but not for writing that requires intentionality, voice, and risk. Accept the limitation, and use it where it shines.