Reddit Please wait for verification

Izzo Anyone who's tried to pass off AI writing as their own knows the struggle.

Izzo Welcome back to Exploring Next, I'm Izzo. Today on episode two-sixty-one, Boone and I are tackling something every content creator hits: why does AI-generated text sound so... AI-generated?

Boone And more importantly, what can we actually do about it beyond just crossing our fingers and hitting regenerate.

Izzo Right. I saw this Reddit thread in ChatGPTPro where someone's asking how to make their AI writing sound more human, and honestly? This hits every product team using AI for content.

Boone It's not just a writing problem though — it's fundamentally about how these models are trained and what they're optimizing for.

Izzo Okay, so why does AI writing feel so polished and obvious? What's happening under the hood?

Boone The core issue is that transformers are trained to maximize coherence and minimize perplexity. They're literally optimized to produce the most predictable, grammatically perfect next token.

Izzo Which sounds good in theory but...

Boone But real human writing is messy. We start sentences and abandon them. We use filler words, repeat ourselves, go on tangents. The model sees that as 'bad' writing to avoid.

Izzo Plus they're trained on edited text — published articles, books, formal documents. Not the rough drafts or casual conversations.

Boone Exactly. So when you prompt ChatGPT, it's drawing from this corpus of polished, professional writing. No wonder it sounds like a corporate blog post.

Izzo So what actually works to fix this? The Reddit user mentioned tweaking prompts versus full rewrites.

Boone Prompt engineering can help. Instead of 'write an article about X,' try 'write like you're explaining X to a friend over coffee' or give it a specific persona.

Izzo I've seen people use prompts like 'write in the style of a Reddit comment' or 'casual blog post with some typos.' Does that actually change the output?

Boone It does, but you're still fighting the model's base training. A more robust approach is multi-pass generation — generate, then prompt it to make the tone more conversational, then maybe add some imperfections.

Izzo That's interesting. So instead of trying to get it right in one shot, you're treating it like human editing.

Boone Right. Humans don't write perfect first drafts. We revise, we second-guess ourselves, we add personality in editing. You can simulate that with multiple prompts.

Izzo What about the technical side? Are there parameter tweaks that help?

Boone Temperature is huge. Most people use the default, but bumping it to 0.8 or 0.9 introduces more randomness. Top-k sampling around 40-60 can also help avoid the most predictable word choices.

Izzo Boone, break down temperature for me — I know it affects randomness but how?

Boone Temperature controls how much the model considers lower-probability tokens. At 0, it always picks the most likely next word. Higher temps make it more willing to pick surprising but plausible alternatives.

Izzo So you're trading some coherence for authenticity.

Boone Exactly. And that trade-off is actually what makes human writing feel human — we don't always pick the 'optimal' word choice.

Izzo From a product perspective, I'm curious about the user workflow here. Are people really going to do multi-pass editing?

Boone Probably not manually. But you could build tools that automate it — generate, then run a 'humanize' pass, maybe inject some controlled imperfections.

Izzo That's actually a solid product opportunity. Like Grammarly but in reverse.

Boone Ha! De-Grammarly. I'm adding that to the weekend project list.

Izzo What about training your own models? If you're a company doing lots of content generation? Fine-tuning on your own writing samples can help a lot. Or using retrieval-augmented generation to pull in examples of your actual voice and style. That makes sense. Instead of fighting the base model, you're giving it better examples to work from. And honestly, the best results I've seen combine multiple approaches — better prompts, parameter tuning, AND post-processing. Alright, so wh