Fluent nothing

Say a piece was written with AI assistance and watch what happens. People stop reading. Not because they found a flaw in it - they never got that far. The label did the work.

It is the wrong axis, and it is worth being precise about why, because the instinct underneath it is sound.

Slop was never about the tool

Ask a model for five tips on productivity and you get slop. Every time. Not because it writes badly - it writes better than most people - but because there was nothing in the request. Nothing was known, nothing was risked, nothing was learned the hard way. You cannot get an idea out of a brief that contained none.

Now give the same model something you learned by doing the thing. A constraint that forced a decision. The failure that was not the one you expected. The thing that turned out to be subtler than the documentation admits. It will express that better than you would have. It cannot have it for you.

Same tool, same session, completely different output - and the difference had nothing to do with the machine.

Here is the part people skip: plenty of writing with no machine anywhere near it is also empty. Most of it, in fact. We had slop long before we had models. We just didn't have so much of it, so quickly.

The other way it goes wrong

There is a second route to slop that has nothing to do with the brief being empty, and it is the one I see more often.

You bring a real idea. The draft comes back fluent. Nobody reads it properly. It ships.

It is still slop, and for a reason worth naming precisely: a model weights every sentence equally. It does not know which line is the point. So it will bury your best sentence in the fourth paragraph, bolt generic framing onto something specific, hedge a position you actually hold firmly, and pad the whole thing out to the shape of an article because that is what articles look like.

None of that is visible to it. All of it is obvious to you, in about ninety seconds.

Review here does not mean proofreading. Spelling is the one thing that will already be correct. It means judgement: is that the point, is that emphasis right, is that claim actually true, did it quietly invent a specific that sounds plausible and is not.

That last one is the dangerous one, because invented specifics read exactly like real ones. They are the right shape. The only person who can catch them is the one who knows what actually happened.

This piece is an example. An earlier draft of the page describing how this site is made claimed I write the posts and a model tidies them. Not true - they are all AI-assisted, and have been from the start. It read well and it was wrong, and the only thing that caught it was somebody who knew the facts reading it and saying so.

Untouched output is the tell. Not that a model was involved - that nobody came back afterwards.

The right question

Not was a model involved. It was, in most things you read now, whether the author says so or not.

The question is: could this have been written by someone who had not done the thing?

If yes, it does not matter who typed it. If no - if there is a detail in there that only comes from having been in the situation - then it is worth your time, and how the sentences got assembled is nobody's business.

That test is annoying because it takes a minute of actual reading. The label takes none. That is the whole reason people prefer the label.

Why the suspicion is reasonable anyway

Here is the honest part, and it explains the instinct rather than dismissing it.

Empty writing used to announce itself. It was badly organised, it repeated, it wandered. You could tell in a paragraph and move on. The style was the filter.

That filter is gone. Emptiness now arrives well-structured, correctly punctuated and confident. It reads exactly like something worth reading, right up until you finish and notice you learned nothing. It costs you the whole piece to find out.

So people fell back on the only cheap signal left - was AI involved - and started discarding on that. It is a bad rule. It is also the only fast one available, which is why it caught on.

I have written before about models being confidently wrong in a register indistinguishable from being right. This is the same thing pointed at prose instead of answers. The tells we spent years learning to read are simply not there any more, and we have not built new ones yet.

Original ideas are still the hard part

This is the thing that has not moved, and I do not think it is close to moving.

A model is superb inside what already exists. It interpolates, connects, rephrases, and finds the shape in a mess of half-formed thinking. What it does not do is arrive at a position nobody has arrived at before - because that is, by construction, outside what it was built from.

Which makes the whole thing a magnifier rather than a substitute. Have something to say and it will help you enormously. Have nothing and it will help you produce a great deal of nothing, quickly, in a form that looks like the real thing.

The scarce input was never the writing. It was having been somewhere and paid attention.

So say it

Every piece here says it was written with AI assistance, including this one. Not as a confession - it is close to the default now, and claiming otherwise would be the less believable statement.

The disclosure was never the risk. The emptiness is. And if what you brought was a real idea, disclosing the tool costs you nothing, because the tool was never the part that was scarce.

← All writing