I spent three hours last Tuesday trying to make ChatGPT write a scene where two people argue about a burnt piece of toast. It failed. Miserably. It made them sound like two HR managers resolving a conflict in a breakroom. There was no grit, no spit, no actual human resentment. Just a lot of ‘I understand your frustration’ and ‘Let’s find a way forward together.’

It was soul-crushing. I’m not a professional novelist—I work a 9-to-5 in logistics and write this blog because if I don’t, my brain will turn into a spreadsheet—but I know what people sound like. And in 2024, ChatGPT doesn’t sound like a person. It sounds like a person who is terrified of getting fired.

If you’re trying to write anything with a pulse, you’ve probably felt this. The ‘AI smell’ is getting stronger, not weaker. But then I started using Claude 3.5 Sonnet, and for the first time in two years of messing with these models, I actually felt a little bit of a chill. Not the ‘the robots are taking over’ chill, but the ‘oh, this actually sounds like someone I’d grab a beer with’ chill.

The ‘Tapestry’ problem and why GPT is a narcissist

ChatGPT is obsessed with itself. Or rather, it’s obsessed with its own idea of what ‘good writing’ looks like, which apparently involves using the word ‘tapestry’ or ‘testament’ in every single paragraph. I hate those words now. I genuinely despise them. If I see a story start with ‘In the heart of the bustling city,’ I want to throw my laptop out the window.

The problem is the internal rhythm. GPT-4o has this specific, staccato-then-flowery beat that feels like a metronome stuck on 60 BPM while you’re trying to dance to jazz. It’s too balanced. It never lets a sentence sit in its own filth. It’s always trying to wrap things up with a neat little bow of wisdom. Real life isn’t wise. Real life is messy and often ends mid-sentence because someone walked into the room.

I might be wrong about this, but I actually think GPT-3 was better for creative stuff. It was weirder. GPT-4o feels like it’s been lobotomized by a committee of safety auditors who think that ‘creative writing’ means ‘writing a very polite cover letter for a job you don’t want.’

Total garbage.

The 14-day dialogue stress test

I’m a bit of a nerd when it comes to tracking this stuff. Over the last two weeks, I ran a specific test. I gave both ChatGPT and Claude the same 10 prompts every day, all focused on high-tension dialogue. I tracked the ‘Adverb-to-Noun ratio’ and the frequency of sentence fragments.

• ChatGPT (GPT-4o): Average sentence length of 19.8 words. Used ‘delve’ or ‘tapestry’ in 14 out of 20 samples.
• Claude (3.5 Sonnet): Average sentence length of 14.2 words. Used sentence fragments (human-like) 22% more often.

What I mean is—actually, let me put it differently. Claude is okay with being ugly. I asked it to write a scene where a guy realizes his wife is cheating because of the way she folds the laundry. Claude wrote: ‘The socks were too neat. Aggressively neat. Like she was trying to apologize to the cotton.’

ChatGPT wrote: ‘As he looked at the meticulously folded garments, a sense of unease washed over him, a silent testament to the secrets hidden beneath the surface of their domestic bliss.’

See the difference? One is a punch to the gut. The other is a Hallmark card written by a ghostwriter on Xanax.

The part where I change my mind about ‘Settings’

I used to tell everyone that you could fix ChatGPT if you just messed with the System Instructions or the ‘Temperature’ settings. I was completely wrong. You can’t prompt your way out of a boring personality. It’s like trying to teach a golden retriever how to be a cynical noir detective. You can put the trench coat on the dog, but it’s still going to wag its tail when the murderer walks in.

Claude feels like it was trained on actual books—not just Wikipedia entries and Reddit threads. It understands subtext. It knows that when someone says ‘I’m fine,’ they usually mean they want to set the house on fire. GPT thinks ‘I’m fine’ means the person is, in fact, fine. It’s literal to a fault.

Claude is like that one friend who actually listens when you’re crying, while GPT is the friend who hands you a self-help book and asks if you’ve tried meditation.

I honestly think Sam Altman doesn’t read fiction. There, I said it. The way OpenAI has tuned their models suggests they think ‘creativity’ is just ‘adding more adjectives.’ It’s a fundamental misunderstanding of why we read stories. We don’t read for the adjectives; we read for the gaps between them.

Why I’m biased (and I don’t care)

I’ll be honest: I have an irrational hatred for Google Gemini. I refuse to use it even though some people swear by it for research. The UI looks like a Google Doc had a mid-life crisis, and the output feels like it was written by a middle schooler who is trying really hard to get an A-. I know that’s unfair. I know it’s probably improved. I don’t care. I’m staying in my Claude-shaped bubble because it’s the only place that doesn’t make me feel like a machine is patronizing me.

Anyway, back to the point. I had this embarrassing moment a few months ago. I tried to use AI to help me write a toast for my brother’s wedding. I was busy at work—we were dealing with a massive shipping delay at the warehouse—and I was exhausted. I let ChatGPT draft the first version.

It was horrific. It talked about ‘the journey of love’ and ’embarking on a new chapter.’ If I had read that aloud, my brother would have known immediately. It lacked the specific, weird details that make a brotherly bond real—like the time he threw a Nintendo controller at my head in 1998. AI struggles with the ‘Nintendo controller’ moments. Claude gets closer, but even then, you have to push it.

But Claude at least tried to make a joke about the controller. GPT just told me to ‘cherish the memories of youth.’

The verdict for 2024

If you want to write a technical manual or a polite email to your landlord, use ChatGPT. It’s great at being a beige wall. But if you’re trying to write a story, a poem, or even a blog post that doesn’t sound like it was spit out by a corporate bot, Claude 3.5 is the only real choice. It’s not even close anymore.

Claude wins. Hands down.

I do wonder, though, if we’re all just training ourselves to like ‘AI-style’ writing without realizing it. Will my own writing start to include ‘tapestries’ in five years because I’ve read too much generated text? I hope not. I really hope not. But sometimes I catch myself typing a sentence that feels a little too balanced, a little too neat, and I have to go back and break it on purpose just to feel human again.

Does any of this matter if we’re all just shouting into a void of generated text anyway?