You put real thinking into a piece. You researched the claim, built the argument from your own experience, wrote it in your own voice, and attached your name to it. Then a platform flagged it. An editor got nervous, or a reader left a comment implying you fed a prompt to a machine and walked away, and the frustration is legitimate, because what happened to your work and what they're reacting to are not the same thing. That conflation is doing real damage to writers who use AI the way they've always used better tools, to work faster without working worse, the same way search engines and spellcheck were quietly absorbed into the job a generation ago.
The distinction that keeps getting skipped
"AI writing" is doing too much work as a phrase. It covers two completely different acts, and treating them as one is what's causing the mess. A writer who uses AI to find sources, stress-test structure, and tighten prose is doing something fundamentally different from an operator who runs a topic list through a model and publishes whatever comes out. The first act has a human author. The second one doesn't, whatever the byline says. That distinction is practical, not semantic, because the corrections platforms and readers are applying catch both categories, and only one of them deserves it.
Five tests that separate assisted work from slop
The distinction is concrete enough to test: assisted writing passes all five of the following tests, while volume filler fails most of them.
Did you choose the topic for a specific reader, based on something you actually know? Can you walk someone through the argument in conversation, without the text in front of you? The voice should sound like you, trained on your writing or shaped by your editing, and you should be able to defend every factual claim with a source you read. Would you put your name on it without flinching if someone quoted it back to you in public?
Those tests have nothing to do with which tools you used during drafting. A piece that passes all five is authored work. Something nobody chose, nobody sourced, and nobody would own, published at scale because volume is the strategy, is slop. The backlash is aimed at that work, however often the anger lands on the wrong desk.
What assistance is genuinely good at
AI is good at research and tightening. The opinion still comes from you, along with the sense of what a specific reader needs, which means the limiting factor in assisted writing is still the writer, and no tool release has changed that yet.
Go read what currently outranks you. Most of it cites no authority for a single claim, and almost none of it names an author a reader could look up, which is the actual quality bar most writers are competing against. Bring a real argument, source every claim, and sign your name, and you are already ahead of nearly everything ranking above you. AI assistance changes how much time it takes to clear that bar, which stays where it always was: a real argument, sourced claims, and a name a reader can look up.
The gap assistance fills is mostly mechanical. It finds the three studies that support your point, flags the paragraph where the logic goes soft, and cuts the sentence that says the same thing twice. Hand-done, that is the same work. Assistance compresses the hours, while the idea, the experience behind it, and the judgment about what to publish travel with you from the first note to the published page.
What detection scores actually tell you
Detection tools are a signal, not a verdict. They flag genuine human writing often enough that treating a score as proof is a mistake, and the underlying methodology is contested enough that no reputable researcher claims otherwise. A score might matter to a specific editor or platform, which is a practical constraint worth knowing about before you submit. Whether the work is honest is a different question, and the five tests above are a better place to look for that answer than any percentage a scoring tool prints.
How Ghosts is built for the assisted case
Ghosts was built because the tools I was watching were optimized for volume, not for writers who cared about their byline. Our workflow is built for the honest case. I'll say directly that we have a commercial stake in that framing.
Every factual claim in a draft gets a source attached to it, something a writer can read and verify before it ships, and voice training pulls from the writer's own samples, so the output sounds like the writer rather than a generic model doing an impression of one. A revision pass targets machine patterns: the hedges, the structure, the phrasing that reads as generated rather than thought through. Our Humanity score rates writing quality, because that's the measure that actually predicts whether a reader trusts the piece. Nothing ships on its own. The writer decides what publishes, and a writer can bring an existing draft to improve rather than starting from scratch.
Every part of the system rests on the assumption that the writer is the author and the work has to be defensible, because that's the only version of AI assistance that holds up when someone asks hard questions about it.
The byline stays with you, and so does the accountability
That's the point of assisted writing done honestly: you keep both. Tools change what's feasible in a day, but who is responsible for what gets published stays the same. A writer who can answer every question about a piece, defend every claim in it, and recognize their own thinking on every page has authored it, whatever tools were in the room. That's been true with every tool that came before this one.