I've been running AI detection on the AI-assisted things I post. The pattern is consistent - it comes back 50% +/- 5% every time. I've started to think that this range is the target.
99% AI reads as outsourced. No stakes, no voice, no judgment. Any prompt could have produced it. That's the slop readers are learning to spot on sight, and rightly so.
0% AI is worse than people realize. You're leaving capability on the table. Your thoughts are only as clear as your first pass of typing. You lose the editorial distance a second party provides. You lose the structural scaffolding that makes complex arguments legible. For most people trying to write publicly, 0% reads as muddled because humans under time pressure tend to be muddled. High-AI is at least organized. 0% is often just rough.
50% is the handshake. AI does what AI does well: structure, breadth, holding many threads, proposing angles the human didn't think of. The human does what humans do well: voice, stakes, specific examples, judgment about what to keep and cut, and the last pass. Neither dominates. The seams are visible if you scan for them, but the voice reads as one person because the human holds authorship.
The prompt isn't where the work happens. The prompt is mostly done in the GPT or Project design upstream. That's where you upload your corpus, your writing samples, your personality profile, your style rules, your domain expertise. By the time you're typing a message in a session, the heavy lift is already done. The AI isn't generating text in a void, it's reflecting back an organized version of what you've already fed it.
Which is why "show me the prompt" is such a good challenge for those who comment "AI-slop" simply because a piece is polished. They assume a single magic prompt produced the output. It didn't. The prompt that produced it was the person who spent months building the GPT, Gem, or Project in the first place, then edited the output to feel right.
This isn't amplification. Amplification suggests volume, and that's not what good AI assistance does. It's more like extension. You take what a person actually knows, thinks, and has lived through, and you extend it into forms that first-pass typing can't reach. Long-form arguments. Structural consistency across many pieces of writing. The ability to hold fifteen threads visible at once instead of one. Your voice stays your voice. What changes is what you can do with it.
Dead internet theory says most of what's online is AI-generated content talking to AI-generated content with humans at the margins. That future is coming whether we like it or not. The humans who'll still be legible through the noise will be the ones whose AI assistance is visibly downstream of something real. A corpus of actual thought. Years of specific domain expertise. A distinctive voice the AI was trained to reflect rather than replace. 50% output is what that looks like in practice.
To build an AI voice replicator well, three things have to be in place:
Content matters. You have to actually know what you're talking about. The AI can organize your thinking. It can't replace it. If you try to generate opinions you don't hold, you'll get generic writing that sounds plausible and means nothing.
Structure matters. AI is exceptional at structure. This is where it earns its keep. Outlines, arguments that build, transitions, callbacks, the scaffolding that holds a long piece together.
Voice matters. Voice is still the human's job. Specific word choices, cadence, tics, the small register shifts that make writing feel like someone. Every system's default voice is smooth and anonymous. If you don't put your voice back in, whatever comes out will read as the platform, not you.
Get all three right and you land in the 50% range without trying. Miss any of them and the scanner will tell you which direction you missed in.
AI-assistance matters. It's a real thing. Pretending otherwise is the same mistake as pretending spellcheck doesn't matter, or pretending Google doesn't matter. The tools shape the writing. What's new is that the tool can now hold structure at the scale of a whole essay, not just a sentence.
When the internet dies properly and every post is suspect, the people who still read as real will be the ones whose method was legible and whose substance was their own. Build the project well, do the actual thinking, edit, fine-tune, and post at 50%.
Humanize button? Nah.. Collaborate button.
.
(btw, this post gets 54% AI on undetectable)
[link] [comments]
