Table of Contents
Do not index
"If I use AI to draft a client's post, do I have to tell them?"
Ghostwriters have been asking me some version of this for two years, and the answer just stopped being a matter of taste. Anthropic announced model-level watermarking on Claude-generated text in August 2026, arriving alongside most of the EU AI Act's transparency provisions, which became applicable on August 2. The technical detail is the part that changes your workflow. The watermark "is imperceptible to readers, does not alter the meaning or readability of the text, and remains embedded in content when it is copied and pasted." Detection moved from guessing to provenance, and any content a model touched, not only content a model wrote outright, may now carry a signal.
So here is the verdict. Stop working out whether you will be caught, because that question has an expiry date on it. Decide instead which parts of your process you would be comfortable showing a client on a screen share, then build the workflow so the answer is all of them.
I run this on what I call the Authorship Split. Two categories, and every task in a content pipeline goes into exactly one. Mechanics on one side, meaning transcription, restructuring, tightening a paragraph that already says the right thing, generating fifteen headline variants from a claim you made, checking whether a draft contradicts something the client said in March. Judgment on the other side, meaning what the person actually believes, which story gets told, which claim is defensible, what the piece is arguing and why. Mechanics can be delegated to a model without changing what the work is. Judgment cannot be delegated without changing what you are selling.
Who this is for. Ghostwriters charging $5k to $30k per month who publish under someone else's name and carry real reputational exposure. Agency owners between $200k and $2M in revenue with two or three writers on payroll and no written policy about what those writers are allowed to paste into a tool. Founders whose executive content runs through a team and who have never asked where the first draft comes from.
This is not for everyone. Skip this if you are producing SEO volume content where nobody involved has ever pretended a specific human authored it, because the disclosure question does not bite there. If you sell AI-generated content as AI-generated content and your clients bought it that way, you already have alignment and nothing here changes. And if your position is that clients do not need to know how the work gets done as long as the output is good, this article will not change your model, though it is worth noting that watermarking turns that position into a bet on nobody checking rather than a philosophy.
The disclosure conversation is a positioning opportunity
Most people are dreading this conversation. It is actually the easiest premium-rate justification you will get this year. When a client asks whether you use AI, the weak answer is defensive, and the weak answer is what most of the market will give. The strong answer is specific. Here is exactly where a model sits in my process, here is what it is not allowed to touch, and here is why that boundary is what you are paying for.
That answer only works if the boundary is real and documented. Write it down before anyone asks. One page naming the tools, the tasks each one is permitted to do, and the tasks that stay human. Get the client to sign off at kickoff rather than fielding it as an objection in month four. In practice this converts a compliance risk into a differentiator, because almost nobody in the ghostwriting market has this document and every buyer has now read a headline that made them nervous.
What this means for quality control
The operational consequence is that your review process needs a provenance step alongside the quality step. If you already run structured review before anything ships, this is one more checkpoint rather than a new system, which is part of why a content quality control system prevents client churn before the retainer ends rather than just catching typos. The reviewer asks not only whether the piece is good, but whether the argument in it originated with the person whose name is going on it.
For a 3 person agency that is a fifteen minute addition per client per week. For a solo ghostwriter it is a habit rather than a process. Either way it is cheap now and expensive later, because the version of this conversation that happens after a client discovers a watermark on their own is a very different conversation from the one you initiate at kickoff.
The trajectory point is that the value of ghostwriting was never in the typing, and this is the event that makes that undeniable. When the mechanical layer becomes verifiable and close to free, the market splits into people who sell drafting and people who sell thinking, and the first group is facing a hard two years. If your process is voice extraction, argument construction, and knowing which of your client's opinions is worth defending in public, watermarking is not a threat to your business. It is the thing that finally makes that distinction legible to buyers who could not see it before.
