AI Detection for Ghostwriters: How to Deliver Drafts Clients Can Pass Off as Their Own Voice

Ghostwriting has always been a kind of impersonation. You write the LinkedIn post, the memoir chapter, the CEO’s keynote – and the byline belongs to someone else. None of that is new. What’s new is that a client can now paste your draft into a detector, watch a number climb past some threshold, and ask why the thing they’re about to put their name on reads like it came from a robot instead of from them.

That question deserves a real answer. Here’s how to handle AI detection in ghostwriting without losing the one thing the job actually sells: a client’s own voice, coming out of your keyboard.

Key takeaways

• Ghostwriting has always meant borrowing another person’s voice; AI detection just adds a new failure mode where the draft sounds like a machine instead of like your client.
• Detectors score statistical patterns, not authorship – so a bland draft can flag even when a human wrote every word, and a vivid one can pass even when AI helped.
• Build a reusable voice profile for each client and edit toward *that*, not toward generic “good writing.”
• Use a detector as a smell test for blandness, then fix the flagged passages by hand.
• Be straight with clients about your tools and process. Trust is the real product.

What an AI detector actually measures (and what it doesn’t)

Start here, because the whole panic falls apart once you understand the machine.

AI detectors don’t know who wrote a sentence. They estimate how *predictable* the text is to a language model. Two signals do most of the work: perplexity and burstiness. Low perplexity means each word is the statistically obvious next word. Low burstiness means the sentences march along at a uniform length and rhythm. AI tends to produce both. So does a careful human editor who has sanded every sentence to the same smooth finish.

That’s the trap for ghostwriters. Your professional instinct – clean, tight, on-topic, no wasted words – is exactly the profile detectors lean toward flagging. A first-draft email from your client, full of half-finished thoughts and a weird tangent about their dog, would probably score *more human* than your polished version of the same idea.

A few facts worth keeping straight so you can talk about this honestly:

• Detectors output a probability, not a verdict. A score like “92% AI” is a guess about pattern similarity, not evidence of who held the pen.
• OpenAI pulled its own AI text classifier in July 2023 because its accuracy was too low to be useful. The company that builds the models couldn’t reliably detect them.
• A 2023 Stanford study (Liang et al., published in *Patterns*) found GPT detectors disproportionately flagged writing by non-native English speakers – misclassifying over half the non-native samples as AI while scoring native English text almost perfectly. If you ghostwrite for ESL clients or write in their patterns, expect more false positives.

None of this means detectors are worthless. It means the score is a signal about *texture*, and texture is something you control.

The real job: a voice profile, not a clean draft

Generic competence is the enemy here. The fix isn’t to write “better” in some abstract sense – it’s to write more like one specific person.

Before you draft anything for a new client, build a voice profile. Treat it like a reusable asset you refine over the relationship.

Mine the source material

Pull every scrap of the client’s actual language: old emails, Slack threads, podcast transcripts, recorded calls, social posts, voice memos. You’re not looking for topics. You’re looking for *moves*:

• Cadence – short and punchy, or long and winding? Do they use fragments? One-word sentences for emphasis?
• Diction – “folks” or “stakeholders”? Do they swear? Do they hedge with “I think” or assert flatly?
• Tics – the phrase they overuse, the way they always open with a question, the dash habit, the parenthetical asides.
• Structure – do they argue point-by-point or tell a story and land the lesson at the end?

Write this down. Three or four bullet points beat a vague memory of “they sound casual.”

Edit toward the person, not toward neutral

Once you have a draft, your editing pass shouldn’t be “make this read well.” It should be “make this read like *them*.” That usually means adding back some controlled mess: a sentence that runs long because they get excited, a blunt three-word reaction, an opinion stated without qualification. This is the same variation a detector reads as human – not a coincidence. People are bursty. Brands trained to sound “professional” are not.

A mini-scenario: the flagged founder post

Say you ghostwrite LinkedIn posts for a startup founder. You draft a sharp 200-word post about a hiring mistake. It’s clear, structured, and lands the point. The founder’s assistant runs it through a detector before publishing – new company policy – and it comes back “AI-likely.”

Nobody used AI to write it. But you can see why it flagged: every sentence is roughly the same length, the transitions are textbook, and there’s not a single thing in it that only this founder would say.

So you go back to the voice profile. This founder tells stories. He swears mildly. He starts posts mid-thought, like he’s continuing a conversation. You rewrite the opener as a flat confession – “I hired the wrong person twice for the same role. Same mistake, twice.” You drop in the detail he actually told you on the call: the candidate aced the case study and bombed the first standup. You let one sentence sprawl and cut the next to four words.

You run it through the checker again. The score drops, but more importantly, the founder reads it and says “yeah, that’s how I’d say it.” That second reaction is the one that matters. The detector was just the alarm that told you the first version was wearing a costume.

Where AI tools fit in a ghostwriting workflow

Plenty of ghostwriters use AI to draft, and there’s nothing inherently wrong with it. The danger is shipping the model’s voice instead of the client’s. A reasonable workflow:

1. Capture raw material in the client’s words. Interview them, transcribe it, keep their phrasing.
2. Draft however you draft – by hand, with AI, or both. The drafting method is yours.
3. Rewrite for voice using the profile. This is the non-negotiable human pass.
4. Check for blandness, fix by hand. If a detector lights up a passage, that’s the spot where your draft went generic. Revise that passage toward the person. Don’t run the whole thing through an automated rewriter and call it done – that just swaps one machine texture for another. (If you’re curious why one-click “humanizers” oversell what they do, there’s more reading on detection and writing on the blog.)
5. Fact-check everything. AI invents quotes, dates, and statistics. Your client’s name is on the line, not the model’s.

The detector is a diagnostic tool in step 4, not the goal. The goal is a draft your client recognizes as themselves.

The trust conversation you can’t skip

Here’s the part most guides dodge. The biggest risk in AI-era ghostwriting isn’t a detection score – it’s a client who feels deceived.

Decide your disclosure stance and put it in the open. Some clients don’t care how the sausage gets made as long as the voice is theirs and the facts hold up. Others – especially in regulated fields, academia-adjacent work, or thought leadership where authenticity is the whole pitch – have firm rules. Find out which kind you’re dealing with on the kickoff call, not after a draft blows up.

A simple line in your contract works: “Drafts may be AI-assisted; all content is edited for your voice, fact-checked, and reviewed by you before publication.” That protects you and gives the client a real choice. Surprising someone with an undisclosed tool after they’ve publicly staked their name is how a referral pipeline dries up.

If you’re handling steady client volume and want to bake voice-checking into your process, it’s worth looking at pricing for regular client work so the tooling cost is predictable rather than a per-project scramble.

Frequently asked questions

Does ghostwriting count as AI-generated content? No. A human writing for another human is centuries old. The wrinkle is that many ghostwriters now draft with AI, and if the prose carries machine-like patterns it’ll flag regardless. The fix is the same either way: edit until it reads like your client.

Will an AI detector know my client didn’t write it themselves? No. Detectors estimate AI-likeness, not authorship. They can’t distinguish ghostwriting from solo writing, and they report a probability, not proof.

Should I tell clients I use AI? Yes – ideally before they ask. Many are fine with AI-assisted drafting when the voice and facts are right. Others aren’t. Set the expectation in writing up front.

My human-written draft got flagged. Why? Clean, even, on-topic prose scores low on perplexity and burstiness, which detectors read as AI-like. ESL patterns and tightly outlined writing also trip them. Treat a flag as “this got bland,” not “you cheated.”

How do I capture a voice with few samples? Mine emails, calls, voice memos, old posts. Draft a short passage, send it, ask the client to mark what sounds like them. Two rounds usually lands it.

The bottom line

Ghostwriting was never about hiding that someone else wrote the words. It’s about making sure the words sound like the person whose name is on them. AI detection didn’t change that job – it just gave you a blunt instrument for catching the moments where your draft slipped into nobody’s voice at all.

Use it that way. Write the draft, build the voice profile, and when something flags, fix it by hand until it sounds like a person again. You can run a draft through a checker to find the bland spots before your client ever does.