Human versus AI-generated content

Clients ask this first. The field answers it worst.

The two loud positions — "AI content gets penalized" and "AI content performs exactly the same" — are both wrong, and both usually come from somebody with something to sell. The useful version splits three questions that keep getting mashed into one.

One: is it against the rules?

This part is documented, and the documentation isn't ambiguous.

Google's position, published in February 2023 and unchanged since, is that how you made it isn't the test. It focuses on quality rather than production method, and calls that a useful guide. The same guidance says plainly that appropriate use of AI or automation isn't against its guidelines.

The limit is stated too. Using automation mainly to influence rankings violates the spam policies, which define scaled content abuse as generating lots of pages mainly to manipulate rankings instead of helping people — and name generative AI used to produce many pages without adding value as an example.

Put those together and the actual rule is clear. It isn't about AI at all:

It was always about intent and volume, not about which tool typed the words. Spun articles, doorway pages and content farms were violations long before language models existed. AI didn't invent a new violation. It cut the price of the old one by about everything.

Two: does it actually perform worse?

Here the honest answer gets less satisfying.

No platform has published a study comparing how human-written and AI-written content rank, controlling for quality, topic, site authority and intent. The comparisons that go around are either vendor content marketing or samples too small to separate "written by AI" from "written badly, fast, in bulk, by a site nobody has heard of."

What you can say without making things up:

• The confound is enormous. Most AI content published at scale gets published because it's cheap, by people optimizing for volume. Anything that correlates with that is measuring the strategy, not the tool.
• The failure modes are real and easy to spot. Generated text drifts toward the consensus of its training data, which is often the consensus of marketing copy. It's fluent whether or not it's right. It has never been anywhere. Every one of those is a quality problem, and quality systems are built to catch quality problems.
• Those failure modes are fixable. They come from leaving it unsupervised, not from the tool existing.

So the defensible conclusion is narrow. AI assistance doesn't seem to carry a penalty of its own, and unsupervised AI content production reliably produces exactly the traits that quality systems mark down. Those are two different statements, and the field merges them constantly.

Three: how do the AI engines treat it?

Nobody outside those companies knows, and we're not going to pretend otherwise.

What you can reason about carefully: these systems pick passages that answer a question well, from a source they feel confident about. Nothing in that mentions who typed it. But the things a system can actually observe — whether a findable person or firm stands behind it, whether claims trace to primary sources, whether the page contains specifics somebody had to go and get — correlate with human involvement, because those are exactly what unsupervised generation doesn't produce.

There's a structural argument worth taking seriously too, though it's an argument and not a finding. A model answering from retrieved documents has no use for a page that restates what the model already knows. The content that earns a citation tends to carry something the model couldn't have produced on its own — a specific number with a source behind it, a firsthand account, a primary document, an original piece of analysis.

If that's right, then the lasting advantage of human work isn't that humans write better sentences. It's that humans can go and find things that aren't in the model yet.

What actually separates good from bad, whoever wrote it

This is the practical core, and it's why "human versus AI" is the wrong axis entirely.

Every row on the left is achievable with AI help and achievable without it. So is every row on the right — it's just far cheaper on the right once you automate, which is how volume operations end up there.

Should you disclose it?

Google's guidance is more specific than most summaries of it.

It suggests considering disclosure where someone might reasonably wonder how the content was made, and adding it where a reader would expect it. It also says don't give AI an author byline — naming a model as the author isn't the way to tell people a tool was involved.

The sensible reading: say how it was made, and attribute it to whoever is accountable for it being right. The byline names who's responsible. A note elsewhere explains the process.

What to actually do

If you take one thing from this page:

Use AI for the work that doesn't require having been there. Outlines, first drafts, restructuring, summarizing source material, headings, metadata, spotting what you've left out. It's genuinely good at all of it.

Don't use it for the things that make content worth quoting. The specific numbers, the firsthand account, the position on the contested question, the judgment about what actually matters. Those are what separate your page from what the model already knows, and the model can't give them to you because it doesn't have them.

Put a name on it and mean it. The accountability is the signal. It's also the only part of this that still holds up the next time the technology changes.