AI Content Production: A Workflow That Does Not Produce Slop

AI made producing content almost free and made verifying it more expensive per published piece.

That is the whole strategic picture, and most content operations have absorbed only the first half. Output went up. The work of checking whether the output is true, specific, and worth reading did not go down — it went up, because there is more of it and because the errors are harder to spot.

A fabricated statistic looks exactly like a real one. A generic paragraph looks exactly like a considered one until you ask what it actually said.

So the useful question is not how to generate more. It is where in a content workflow a model belongs, and where a person has to stay.

---

The structural problem with generated content

A model produces something close to the average of what has already been written on a subject. That is what it is for, and on most tasks it is exactly what you want.

It is the opposite of what content marketing needs.

Average content does not get linked to. Nobody cites the fourth article saying the same thing. Average content does not get chosen by a reader comparing three tabs. And average content does not get repeated by a language model summarising a topic, because a model summarising has no reason to prefer your restatement of the consensus over anyone else's. What does get cited.

Which produces a rule with real teeth: the parts of a piece that make it worth publishing are precisely the parts a model cannot supply. What you tested. What a client said. The number you measured. The thing you got wrong and corrected. Everything else is scaffolding, and scaffolding is what AI is genuinely good at.

Two consequences follow:

• Use AI for the scaffolding, which is most of the words and none of the value
• Do not publish anything whose distinctive content is zero. If you removed every generated sentence and nothing remained, there was nothing to publish

---

Where AI belongs in the workflow

Nine stages. AI helps at four, hurts at two, and is neutral at the rest.

Stage	AI role
Deciding what to write	None. This comes from customers, sales calls, support tickets, search data
Research and gathering	Assist, with sources attached. Never as the only source
Outlining	Strong. Fast, cheap, easy to correct
Drafting scaffolding	Strong. Definitions, background, transitions, structure
Drafting the distinctive parts	None. This is the reason the piece exists
Fact-checking	Assist at most. A model checking a model's claims is not verification
Editing for structure	Useful. Reordering, tightening, spotting repetition
Editing for voice	Weak. It converges on a register that reads as generated
Final read	None. A person, out loud, before publishing

The line to hold: a model can produce the parts of a piece that are true of everyone, and cannot produce the parts that are true only of you.

---

Does Google penalise it?

Short answer: not for being AI-generated. Google's stated position is that content is judged on quality regardless of how it was produced.

The rule that actually applies is scaled content abuse — producing many pages primarily to manipulate rankings rather than to help people. Google's spam policy explicitly states this applies regardless of how the content is created, which means a hundred thin human-written pages and a hundred thin generated ones are treated the same way.

So the risk is not the tool. It is the pattern: volume, thinness, and no reason for any individual page to exist. The full answer, including what people get wrong about it.

---

The verification problem

This is the part that gets skipped, and it is the part that ends careers.

Models fabricate confidently. A statistic with a plausible number, attributed to a plausible organisation, in a plausible year. It arrives looking identical to a real one, which is why skim-reading does not catch it and why "I checked it over" is not checking.

Three shapes it takes:

• A real figure attached to the wrong source
• A real source credited with a figure it does not contain
• A number with no origin at all, phrased as though widely known

The third is the most dangerous because it is the hardest to disprove. Searching finds other articles repeating it, which reads as corroboration and is actually circulation.

The discipline that works is boring and it works: list every factual claim in the piece, find each one at its primary source, and delete the ones you cannot. Deleting is a legitimate outcome and it happens more often than people expect. How to run the check.

One practice worth adopting from this site: keep a record of what you refused to claim. Every article here ends with a "deliberately not claimed" note listing figures that did not survive verification. It costs nothing, it stops the same unverifiable statistic being re-added in six months, and it makes the piece more trustworthy than a citation would.

---

The specificity pass

The single edit that most improves generated text is replacing general statements with specific ones.

Generated prose defaults to the safe general form because the general form is the average of every source. The general form is also unfalsifiable, unmemorable and uncitable.

Generic	Specific
"Many businesses struggle with deliverability"	"A domain sending under 300 messages a day to Gmail cannot see reliable data in Postmaster Tools"
"Regular updates are important for security"	"The window between a published vulnerability and automated exploitation is short, which is what makes 'I'll update next month' the actual risk"
"Personalisation improves engagement"	"A message that works with no personalisation tokens cannot break; one that says 'Thanks for your order, there' has"

Three questions that force specificity onto any paragraph:

• Compared to what?
• How would somebody check this?
• What happens if it is wrong?

A paragraph that survives all three is worth keeping. Most generated paragraphs survive none. Writing that says something.

---

Voice

Editing generated text into your voice is harder than writing in it, and this is the stage where the time saving disappears.

Four markers that identify generated prose to a reader who reads a lot:

• Balanced-clause rhythm. Every sentence weighted the same, no short ones
• The three-item list, applied to everything, including things that have two parts or five
• Hedged completeness — "it's important to note", "while X, it's also true that Y" — the tone of avoiding being wrong rather than trying to be right
• Summary paragraphs that restate what was just said without adding to it

The practical fix is not rewriting. It is cutting. Most generated drafts improve by 30% deletion, and the deleted parts are the connective and summary material rather than the substance.

Then read it out loud. It catches what the eye does not — and it is the one quality step that cannot be delegated to anything.

---

Scale, and where it breaks

The constraint is not how much you can produce. It is how much you can verify, differentiate and distribute.

Verification does not scale with generation. Checking twenty claims takes twenty times as long as checking one, and the checking is human work.

Differentiation does not scale at all. The distinctive material — what you tested, what a customer said — is produced by doing things, not by writing. You cannot generate your way to having more experience.

Distribution got harder as supply went up. More content chasing the same finite attention, in search results increasingly summarised before the click. What that did to click-through.

Which gives a realistic ceiling: you can publish as much as you can verify and differentiate, and that number is not much larger than it was before. What changed is that the same number of pieces now takes less time to draft — which is a real saving, and a smaller one than the pitch implies. Where scale actually breaks.

---

A workflow that holds up

Eight steps. The AI-assisted ones are marked.

1. Choose the topic from evidence — customer questions, support tickets, search data. Not from a model's suggestion list
2. Decide the distinctive claim. What does this piece say that the existing results do not? If there is no answer, stop here. This is the step that prevents most bad publishing
3. Gather sources yourself, primary where they exist. AI-assisted, with every source opened
4. Outline. AI-assisted
5. Draft the scaffolding. AI-assisted
6. Write the distinctive parts by hand. The claim, the example, the number, the correction
7. Run three passes: facts, specificity, cut. Facts by hand. The cut can be AI-assisted; the judgement cannot
8. Read it aloud, then publish. Record what you refused to claim

The two steps that cannot move: deciding what is distinctive, and verifying what is true. Everything else is negotiable.

---

Frequently asked questions

Does Google penalise AI-generated content?
Not for being AI-generated. Google judges content on quality regardless of production method, and the policy that applies is scaled content abuse — producing many pages primarily to manipulate rankings, which the policy states applies regardless of how the content is created. The risk is thinness and volume, not the tool.

What should AI not be used for in content production?
Deciding what to write, producing the distinctive claim, and verifying facts. A model can write what is true of everyone; it cannot write what is true only of you, and a model checking a model's claims is not verification.

How do I stop AI content sounding generated?
Cut roughly 30% — the connective and summary material rather than the substance — replace general statements with specific ones, and read it out loud before publishing. The markers readers notice are uniform sentence rhythm, three-item lists applied to everything, and summary paragraphs that restate without adding.

How much content can I actually produce with AI?
As much as you can verify and differentiate, which is not much more than before. Generation got cheap; checking claims is still human work that scales linearly, and the distinctive material comes from doing things rather than writing about them.

Are AI content detectors reliable?
Not reliably enough to act on. OpenAI withdrew its own classifier in 2023 citing low accuracy, and detectors are known to misclassify writing by non-native English speakers. Treat a detector result as an opinion, not evidence.

How do I fact-check AI-generated claims?
List every factual claim, find each at its primary source, and delete what you cannot verify. Watch particularly for a real figure attached to the wrong source, a real source credited with a figure it does not contain, and a number with no origin that other articles repeat — repetition is circulation, not corroboration.

Is it worth publishing AI-assisted content at all?
Yes, where the assistance is on the scaffolding and the distinctive material is yours. The test is simple: remove every generated sentence and see what remains. If nothing does, there was nothing worth publishing.