There is a difference between using a model to pressure-test a paragraph and letting a model become the public speaker who happens to have your byline. The first can be craft. The second is ghostwriting - whether or not you intended the word.
This strategic piece sits on top of the practitioner stack: muse boundary, draft then critique, taste-preserving edit, and disclosure credits. The stage map remains keep your taste. Authorship norms: keep your name on AI-assisted work.
Treat your public voice as a governed asset. Decide which channels may use generative drafting, which may use critique-only assistance, and which are human-only - then enforce that with the same seriousness you would apply to a brand trademark.
Decision framework
Score each public surface (newsletter, launch blog, client letter, personal essay, social thread, talk script, grant statement) on four questions:
- Reliance. Are people relying on this text as your judgment, story, or expertise?
- Substitutability. If a peer could tell a model wrote most of it, does that damage trust?
- Skill. Is writing this kind of piece how you maintain a capability you still need?
- Consequence. What breaks if a hallucinated detail, stolen tone, or misleading fluency ships?
| Score pattern | Policy |
|---|---|
| High reliance or high consequence | Human draft required; AI critique-only at most |
| High skill maintenance | Human draft required on a cadence (e.g., every issue / every talk) |
| Low reliance, low consequence, clear disclosure OK | Draft assist allowed with edit + credit rules |
| Any surface with intimate others’ stories | Consent + privacy rules first; often critique-only |
Illustrative scenario: a ceramicist’s weekly studio notes are where collectors learn to trust their eye. Outsourcing those notes to a cheerful model produces growth metrics and a slow authenticity leak. Product-spec sheets for kiln shelves might tolerate more assist. Same person, different surfaces.
Cost, risk, and capacity
| Approach | Upside | Risk | Capacity need |
|---|---|---|---|
| Human-only on flagship surfaces | Trust, skill, distinctiveness | Slower shipping | Calendar blocks for writing |
| Critique-only AI | Faster revision | Still requires a real draft | Draft discipline |
| Draft-assist + disclosure | Speed on commodity text | Voice sameness; disclosure failures | Edit rigor |
| Silent full ghostwriting | Short-term volume | Reputation shock when noticed; authorship disputes | Crisis management later |
Silent ghostwriting is rarely cheaper once you count the trust repair. Publishing ethics bodies reject AI-as-author because responsibility cannot be assigned to a tool (COPE). Copyright offices likewise center human authorship (Copyrightability Report; registration guidance).
Governance constraints to respect
- Contracts and workplace policy - some clients forbid generative drafting; some require disclosure (workplace AI disclosure).
- Venue rules - journals, festivals, platforms.
- EU transparency trajectory - Article 50 AI Act duties from 2 August 2026 for in-scope content (Regulation (EU) 2024/1689; limited Article 50(2) marking grace to 2 December 2026 for certain systems already on the market).
- Consumer protection - do not automate deceptive claims (FTC announces crackdown on deceptive AI claims and schemes).
Ghostwriting workflows often paste the most sensitive material - unpublished opinions, client conflicts, family stories - into consumer tools. If the surface is sensitive enough to need your real voice, it is sensitive enough for a privacy review before any paste.
Recommended rollout
- Inventory your public surfaces (one afternoon).
- Assign each a policy: human-only / critique-only / draft-assist+disclose.
- Publish the policy where your future self can see it (notes app is fine).
- For draft-assist surfaces, require the taste-preserving edit pass and a credit line when appreciable.
- Quarterly: write one flagship piece with zero generative drafting and compare distinctiveness - same spirit as the unaided check in the creative practice delegation map.
Do not do this yet (or ever, silently)
- Do not let a model invent personal anecdotes for a memoir-like surface.
- Do not ship leadership notes, apology letters, or recommendation letters as un-reviewed AI drafts.
- Do not use AI to imitate another living writer’s voice for public work (visual references ethics applies in prose too).
- Do not treat detector scores as proof of human authorship; detectors are unreliable, and honesty does not depend on them.
Worked surface inventory (illustrative)
| Surface | Reliance | Policy |
|---|---|---|
| Personal essay newsletter | High | Human draft; critique-only AI |
| Product changelog | Low-medium | Draft-assist OK + light disclosure if substantial |
| Client apology after a miss | High consequence | Human-only |
| Conference talk script | High skill + reliance | Human draft; AI for timing critique only (presentation rehearsal patterns) |
| Social promotional blurbs | Lower | Draft-assist OK if voice-checked and non-deceptive |
Your inventory will differ. The artifact matters more than matching this table.
Refusal scripts you can reuse
When a collaborator asks you to “just generate the post in my voice”:
- “I can outline or critique. I will not ship a model draft under my name on this channel.”
- “This surface is on my human-only list. I can trade deadlines, not voice.”
- “If we need speed, cut scope - shorter human text beats long ghostwritten text.”
When a client asks for silent generative drafting against their own audience trust:
- Restate disclosure expectations in writing.
- Offer critique-only or human drafting with a longer timeline.
- Decline work that requires deceptive authorship; see also what not to delegate.
Decide which public surfaces stay yours
Fill the anti-ghostwriting voice checklist for your top five channels this week. Put at least one channel in human-only. Tell one collaborator the rule so it is harder to quietly abandon when a deadline hits.
