Should You Let AI Write Your First Draft?

Should you let AI write your first draft?

For some kinds of writing, yes. For anything that depends on your judgement, experience or verified facts, no. The distinction is not about quality of prose. It is about whether the hard part of the piece is arrangement, which a model handles well, or knowing, which it cannot do for you.

I publish a lot. More than 350 articles on answer engines, schema and E-E-A-T, which means I have had every version of this argument with myself. My position is not that drafting with a model is wrong. It is that most people are pointing it at the wrong step.

Here is where I actually use it, where I refuse to, and how I check the result.

What is a first draft actually for?

It is where you find out what you think. That is the uncomfortable truth about drafting, and it is why handing the step away feels efficient and often is not. The struggle in a first draft is usually the thinking arriving, not the sentences being slow.

When a draft appears fully formed without that struggle, you skip the part where you discover your own argument. You then spend the editing pass reacting to somebody else's structure, which is a different and generally weaker activity than building your own.

This is why the advice splits by task. If you already know exactly what you think and the job is to arrange it, generation helps enormously. If you do not yet know, generation produces something plausible that quietly replaces the thinking you had not done.

Where does an AI draft genuinely help?

Restructuring, summarising, expanding notes you already wrote, producing variations to react against, and handling the mechanical middle of a piece where the content is settled and only the phrasing is open. In all of those the substance already exists and the model is arranging it.

The clearest case for me is turning rough notes into connected prose. I have the argument, I have the evidence, and what remains is joinery. A model does joinery quickly and I edit heavily afterwards, which is a genuinely good trade.

The second is producing an option I disagree with. Reading a version that is wrong is often the fastest way to work out what right looks like, and generating that on purpose is much cheaper than writing it yourself.

What goes wrong when the model drafts from nothing?

You get fluent averageness that reads as an article and contains no position. It will be structurally correct, adequately phrased and completely forgettable, because it was assembled from what is generally said about the topic rather than from what you specifically know about it.

The worse failure is invented specificity. A draft produced from a thin prompt will supply details to make itself sound substantiated, and those details are exactly the ones that get a piece taken down. A confident sentence with a fabricated statistic in it is the most dangerous thing you can publish.

That is why I treat every factual claim in any draft as unverified until I have checked it against a primary source myself, which is the whole reason I verify every fact before publishing rather than trusting fluency.

How do you keep your own voice in the output?

Write the parts only you could write, yourself. The opinion, the lesson, the thing you got wrong, the specific judgement about when the usual advice fails. Let generation handle the connective material and never the position, because the position is the reason anyone is reading.

In practice this means I write the spine first as bare statements, then generate around it, then rewrite anything that sounds like it could have appeared on any site. The test I apply is simple: if this paragraph could sit unchanged on a competitor's blog, it is not earning its place.

The other thing that preserves voice is keeping your own sentence rhythm. Models tend toward a smooth, even cadence. Real writing is lumpy. Short sentence. Then a longer one that carries a qualification the short one could not. That unevenness is most of what people recognise as a voice.

What must never come from a model?

Any first person claim about your own experience, and any number. Not the client result, not the time it took, not what you learned from a project. Those are the assets that make the piece worth reading, and inventing them is both dishonest and easy to do accidentally.

This sounds obvious and it is violated constantly, usually not on purpose. A model asked to write in your voice will produce plausible first person specifics because that is what the genre contains, and a tired writer skimming a draft will leave them in.

My rule is that personal facts and figures only enter a draft by my hand, never by editing something generated. That way there is no version of the file in which a fabricated personal claim ever existed and could survive a careless pass.

How do you check a draft you did not write?

Go through it claim by claim and label each one. Is this something I know, something I verified, something I am asserting as opinion, or something that just appeared? Anything in the fourth category is either verified now or deleted. There is no middle option.

Reading for quality is not enough here, because the failure mode is a sentence that reads beautifully and is false. You have to read adversarially, sentence by sentence, asking what evidence supports each one. It is slower than writing the paragraph would have been, sometimes.

That last point is the honest trade nobody mentions. For claim heavy work, generating and verifying can take longer than writing carefully in the first place, which is something I had to learn the hard way about publishing at volume.

Does publishing AI assisted writing hurt you in search?

Not by itself, according to Google. In guidance published on 8 February 2023, Google stated that its focus is on the quality of content rather than how content is produced, and that its ranking systems aim to reward original, high quality content demonstrating E-E-A-T.

Google was equally clear about the limit. That same guidance states that using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results is a violation of its spam policies. It also says that not all use of automation is spam, pointing out that automation has long produced helpful content such as sports scores, weather forecasts and transcripts.

The test Google offers creators is to think in terms of Who, How and Why. I find Why the most useful of the three. If the honest answer to why this page exists is that you wanted a ranking rather than a reader, no amount of process fixes it, and the rest of your trust signals will not compensate.

What should you do next?

Take your next piece and decide, before you open anything, whether the hard part is knowing or arranging. If it is knowing, write the spine by hand first. If it is arranging, generate freely and edit hard. That one decision resolves most of this debate.

Then write down your own list of what may never be generated. Mine is personal claims and numbers. Yours may be longer. Having it written down means you are not making the judgement while tired at the end of a draft, which is when it goes wrong.

If you are trying to work out where a model fits in your own writing process without losing what makes your writing yours, I am happy to talk it through. Let's chat.