The honesty test for AI-assisted writing
Fluent sentences from a model don’t make AI-assisted writing dishonest. The real test is where the idea came from, not who typed the words. This piece runs through a drafting pipeline that hands the prose to an AI subagent and keeps the thinking, the judgment, and the final approval with me.
Last updated: 2026-07-23
Why isn’t “a human typed every word” the right test?
A test built entirely around who typed the words misses the actual transaction between writer and reader, the same transaction ghostwriters and PR shops proved mattered decades before any model existed: a polished sentence hiding a hollow idea still wasted the reader’s time.
That’s the standard I use for my own posts now: whether the idea came from me, not whose hands typed the sentence. Only one of those two questions has ever mattered to the person reading.
What’s being handed to the AI in this pipeline?
This piece, written by Antony Evans, went through the exact AI-assisted writing pipeline it’s describing, from hook to publish: a spawned AI subagent handled the drafting work before I ever saw it, running through hook options, a full draft, and a humanizing pass, before the review and sign-off happened in my own hands.
The pipeline runs in a fixed sequence:
- draft hook options
- write the full piece
- run a humanizing pass against a deterministic taxonomy of AI writing tells
- self-check the draft against those same gates
- hand the draft to me for review and final sign-off
Nothing publishes until the last step happens. I’ve found that’s what keeps the standard honest, not any step before it.
Why doesn’t a fluent sentence make an idea any more honest?
Fluency is a property of the sentence, not the claim underneath it, which is why a model can write a clean, well-paced sentence around an idea that is empty, borrowed, or simply untested and still have the sentence read as fully confident to the reader. That gap between fluency and truth is exactly where a reader gets misled.
Ghostwriters figured this out generations ago. PR shops built entire careers on it. The fluency was never the tell of honesty, and it is not the tell now that a model can produce it in seconds.
What happens if I stop doing the thinking myself?
What breaks first is the source of the idea, not the sentence: the moment the thinking behind a post stops being worked out by me before the AI subagent in this pipeline drafts anything, the piece is no longer something I can call mine, no matter how carefully I edit the prose afterward. That’s the one condition attached to the idea-over-keystrokes standard, the practice of judging a post by where the idea came from rather than who typed the sentence, and it holds only for as long as the thinking is mine before any of it reaches the model. The day I start asking the model what to believe instead of how to phrase what I already believe, that condition breaks and the post stops being mine.
This doesn’t work if I outsource the judgment along with the sentence. The trade-off is real: speed on the prose side buys nothing if it costs me the thinking side. Handing over the sentence is delegation. Handing over the belief is something else.
Where does this break down?
The whole idea-over-keystrokes standard only works because I still read every draft myself before it ever goes out, and if I ever stopped doing that one thing, the entire argument in this piece would collapse into exactly the excuse it’s trying to reject. That’s the actual failure point, not the AI.
I don’t have a fix for that beyond doing the read every time. The discipline is the whole system. Remove it and the pipeline is just a faster way to publish something I do not believe.
FAQ
Every answer below is Antony Evans’s own account of running this pipeline, in his own words, not a claim sourced from an outside expert or third party.
Is it dishonest to use AI to help write a LinkedIn post? Not on its own. It becomes dishonest the moment the idea itself, not just the sentence, comes from the model instead of from the person publishing under their name.
How much of this article did AI write? The sentences, mostly. A spawned AI subagent drafted the hooks and the full piece, then ran a humanizing pass against a fixed list of AI writing tells. I read every section, cut what wasn’t mine, and signed off before it went anywhere.
What’s the difference between AI-assisted and AI-generated content? In this pipeline, assisted means the AI carries the sentence and I carry the thinking. Generated, as I use the word, means the AI supplied the idea too, not just the wording.
Does using AI break the trust between a writer and a reader? Only if the writer stops doing the thinking and lets the model supply the position along with the prose. The trust was always about the idea, not the keystrokes.
Where’s the line between getting help with the sentence and outsourcing the thinking? The line is whoever decides what to believe before a word gets typed. If that’s still me, the AI is a tool. If it’s the model, the byline is wrong.
Does this pipeline fact-check itself before anything publishes? It runs a self-check against a fixed set of writing-quality gates before I see the draft, but it can’t verify facts on its own. That gate catches AI writing tells, not factual accuracy, which is why I still read and approve every piece myself before it goes out.
Where would you push back: the sentence, or the thinking behind it?