Guides
AI-assisted business writing

Why a document drafted by a model still needs an author

A model can produce a full draft from a prompt, but a person still has to decide what the document argues and take responsibility for it. We look at what the author’s job becomes once a model has done the drafting.

Once you’ve seen how much a model can produce from a few lines of instruction, it’s tempting to think that writing is now mostly a matter of asking well, and that a good prompt will always produce a good document. For some documents, such as a routine notice or a first summary of meeting notes, that’s close enough to true. But for the documents that businesses and researchers depend on, such as a proposal or a policy report, that view overlooks much of what the writing has to accomplish.

Producing sentences is only one part of writing. The rest is working out what you think and testing it against the evidence, and then taking responsibility for the result. A model can generate text at any speed you like, but it can’t form a unique perspective or answer for what the document says. Readers rely on you for both.

What’s missing from your prompt

A prompt describes what you want, but it can’t capture everything you know because much of your understanding of a problem isn’t in a form you can easily type into a chat window. That understanding includes the client’s reaction in last month’s meeting, what went wrong on the previous project, the figure you’re fairly sure of but would want to check, and the objection you expect the board to raise. Nor can you prompt for the specifics that make a document yours, since you often don’t know which ones matter until you’re partway through writing it.

A model fills in what’s missing with what has already been written about problems like yours. There’s evidence that writing with a model’s help can make different people’s work more alike: in one experiment, short stories written with ideas from a model were more similar to one another than those written without that help (Doshi & Hauser, 2024). Our guide to why your AI-drafted report sounds generic looks at why this happens and how to fix it.

A prompt also can’t contain what you’d discover by writing. Many people who’ve written a substantial document find that the argument they finished with wasn’t the one they started with because writing it forced them to test it. A model gives you a polished version of your first idea. If you accept that first draft as it stands, you won’t find out what you’d otherwise have concluded by the end.

The author’s job

If a model writes the draft, your job as the author shifts from writing the text to judging it. That judgment has several parts. The first is deciding what’s true. Every claim in the draft is going out under your name, but the model can’t tell you which ones are sound. If you don’t check each one, you’re asserting it on the model’s word alone.

You also decide what matters. Where a model tends to give each point similar weight, you know that the third recommendation is the one the client will act on and that the first is only there for completeness. The document should be shaped around the third. The same knowledge tells you what to leave out: a model has no reason to keep a draft short, so it often includes everything. Only you know what a particular reader can do without.

The way the document sounds is also yours to decide. A model writes in a general-purpose style by default, but your organization has a voice that readers who know you will expect. Read the draft for anything that doesn’t sound like something you’d say. Our guide to keeping your house voice when a model writes the drafts shows you how to describe that voice so that you can edit the model’s first draft against it.

Finally, you stand behind the document, and when it’s questioned, you’re the one who has to answer for it. Readers trust a document partly because a named person or organization is accountable for it; a tool can’t be held to account in the same way.

How to tell whether a document has an author

To test this, ask the person named on the document to explain, without looking at it, what it argues and why each of the main claims is true. Someone who can do that was using the model as a tool and is the document’s real author. Where they can’t, the document still needs work by the named author before anyone else reads it.

Keeping the author’s work with the author

None of this is a reason to stop using models. They can take much of the drudgery out of producing a document, but the jobs that need an author should stay with the author. Our guide to where AI helps in a business writing workflow explains which tasks those are and suggests a process that keeps the important decisions with you.

If you’d like to develop the author’s side of the work yourself, our writing coaching pairs you with one of our editors, who works through a real document of yours alongside you. For a draft that already exists, however it was produced, we can edit it in your organization’s voice, with tracked changes and margin comments. Each decision about what the document asserts stays with you.

References

  • Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), Article eadn5290. https://doi.org/10.1126/sciadv.adn5290

Draft in hand?

Send it over in whatever state it’s in, and we’ll agree the scope with you before quoting.

Get a quote