The argument about whether AI can write is the wrong argument.
It has been framed as a binary — AI writes or it doesn’t — when the actual question is more granular: which specific tasks in the content production process does AI handle better than humans, and which does it handle worse?
The answer is not “all of them” and it is not “none of them.” It divides roughly along the research and writing line, and understanding that division is the most practically useful thing you can know about AI in a content operation.
What research requires
Research, as it applies to content production, is a set of specific tasks: understanding what has already been written on a topic, identifying the consensus view and the positions that challenge it, finding the data points and examples that will support an argument, mapping the existing territory so the writer knows where they are entering it.
These tasks require coverage and pattern recognition. They benefit from scale. They do not require the evaluative judgment that determines whether an insight is interesting or whether an argument is true. They require inventory.
AI is good at inventory. It can describe what the field has said about a topic with reasonable accuracy — with important caveats about hallucination and knowledge cutoffs — identify the major positions, surface the common examples and counterarguments, and produce a research scaffold the writer can work from.
This is not a complete research function. The editorial judgment that determines which of those data points is actually useful for this particular piece, which example will land with this specific audience, which argument is worth engaging — that judgment is human.
But building the scaffold is a task where AI’s speed advantage is real and the accuracy requirements are manageable, because the writer will evaluate the scaffold before using it.
What writing requires
Writing, in the editorial sense, requires something different: the ability to make a specific argument, from a specific perspective, for a specific reader, in a specific register.
The word “specific” is doing all the work in that sentence. What AI cannot do is produce the specific from scratch. It can produce the general, the average, the expected version of an argument with high fluency. It cannot produce the specific version that only this writer, with this experience, for this audience, could have written.
This is not a temporary limitation. It is structural. AI produces outputs by predicting what follows from what came before, based on patterns in training data. The most common version of a thing is the most predictable — and therefore the most AI-native — version of a thing. The specific, original, experience-grounded version is, by definition, underrepresented in any training dataset.
The Helpful Content Update targeted content that lacked this specificity. AI-generated content was disproportionately affected not because of how it was produced but because of what it tended to produce: the most common version of topics, competently assembled, without the specificity of expertise that distinguishes content worth reading from content that merely covers a subject.
The workflow implication
The practical implication of this distinction is significant and shapes the entire AI content workflow.
Use AI at the research stage: building the landscape, identifying what has been said, surfacing examples and data, producing the scaffold. The output of this stage is a research document — not a draft, a foundation. Treat it as background briefing material, the same way you would treat research produced by a human researcher.
Use human editorial judgment at the writing stage: making the argument, selecting the specific angle, bringing the direct experience, writing in the register the audience needs. The research document informs. The writing creates.
The failure mode that produces content teams disappointed by AI is almost always the same: they used AI for writing and then wondered why the output felt generic. They used it for the task it does poorly instead of the task it does well.
What this means for team skill development
The content team that becomes good with AI is not the team that learns to prompt better — although prompting well is genuinely valuable. It is the team that becomes more precise about which work they need AI to do and which work they need humans to do.
That precision requires a clear-eyed assessment of what different team members are actually good at. A writer who is exceptional at synthesizing research and producing original arguments is not a researcher. Their time is expensive at the research stage and irreplaceable at the writing stage. A researcher who is excellent at coverage and synthesis but not at original argument is the inverse.
AI changes the economics of the research stage — making comprehensive coverage faster and cheaper — without changing the economics of the writing stage, where the original judgment is the value. The organizations that use it well are the ones that understand this and build their workflows accordingly.
Jacob Clifton is the principal of Clifton Creative, an editorial strategy consultancy based in Austin, Texas. He spent fourteen years as a flagship staff writer at Television Without Pity and has written for Tor.com, Vulture, BuzzFeed News, and the Austin Chronicle.
For inquiries: jacob@cliftoncreative.agency · cal.com/cliftoncreative