Why AI Rewrites Sometimes Sound Worse Than the Original, and What to Do

What I wrote

The tiler showed up on a Tuesday, looked at the subfloor for about ten seconds, and left.

What came back

Upon arrival, the tiling contractor conducted a preliminary assessment of the subfloor and subsequently determined that conditions were unsuitable for installation to proceed.

Cause 01 / Sampling

It picks likely words, not good ones

A language model generates text by repeatedly estimating which token plausibly follows the ones before it. Settings such as temperature and top-p decide how far it strays from the front runner. Near their defaults, it strays very little.

"Showed up" is a common phrase. "Upon arrival" is commoner still inside the formal register the model associates with words like polish and improve. Every single choice tugs toward the middle of the distribution. Do that four hundred times in a row and the middle is where you land, which reads as competent and anonymous at once.

Cause 02 / The brief

Your instruction was vaguer than it felt

"Make this better" is not a brief. It carries no information about audience, register, length target, or what you personally count as better. The model has to guess, and its guess is the average of everything ever labelled polished, improved or professional.

Those documents skew corporate. The word professional is a particularly reliable way to summon passive voice, nominalisations and hedging. Swapping it for a concrete constraint changed my output more than switching models did.

Cause 03 / Compression

Detail is costly, abstraction is cheap

Specifics are statistically rare. "4.2 seconds", "the tiler", "on a Tuesday" are low frequency and have no obvious tie to the words around them. Generic stand-ins such as performance, the contractor and at the outset are high frequency and fit nearly anywhere.

So when the model rephrases, the safe move is to climb one rung up the ladder of abstraction, because a vague claim is harder to get factually wrong. Sensible for a system penalised for errors. Ruinous for a sentence whose only value was the figure inside it.

Cause 04 / Preference tuning

Hedging and balance were rewarded

After pretraining, these models are tuned against human preference data. Raters, on average, approve of answers that sound balanced, thorough and careful not to overclaim. That preference gets baked in and then shows up in your paragraph as qualifiers you never asked for.

"This does not work" comes back as "this may not be the most effective approach in every context." Nothing factual was added. Certainty was removed, and your certainty was doing real work.

Cause 05 / Missing context

It cannot hear the rest of your document

Paste one paragraph and the model has no idea what surrounds it. It cannot tell whether you are three thousand words into a dry personal essay or opening a compliance memo.

With no register to match, it defaults to a neutral mid-formal voice that sits badly inside anything with a strong tone. This is why rewrites so often read fine in isolation and wrong in place. It is also the cheapest problem here to fix and the one most people skip.

Cause 06 / Iteration

Each pass compounds the last

Rewriting a rewrite is where it gets properly bad. Pass one sands the edges. Pass two treats the sanded version as the source of truth and sands that. By pass three, vocabulary has narrowed, sentence lengths have converged, and any surviving oddity has been read as an error and corrected.

Researchers looking at what happens when models are trained on their own output describe a similar narrowing at the training level, where rare cases vanish first. You can watch a small version of it inside one afternoon of editing.

Accept or reject, in six questions

Tick what survived

□ Every number, name and date is still there. Not paraphrased, not rounded, not turned into "significant".

□ The word count did not grow. A rewrite that adds 30% has almost certainly added padding, not clarity.

□ Your position is still the same strength. Search for may, could, often, generally, in some cases. Count them before and after.

□ Sentence lengths still vary. If every sentence now runs 18 to 24 words, the rhythm has been flattened.

□ You would say this out loud. Read it to the wall. Anything you would not say to a colleague comes out.

□ It still fits the paragraph before and after it. Paste it back in place and read all three together.

The one that catches most people

Word count. It is the fastest proxy for the whole problem, because nearly every failure mode on this page shows up as extra words. If the rewrite is longer and you did not ask for longer, that is your signal to look harder.

Worth knowing

These tools tend to go along with the framing you hand them. Ask "can you make this stronger?" and you will get something longer, which is a different thing. Ask "is this already clear enough?" and you will often be told yes. Phrase the question so that leaving your draft alone is a permitted answer.

01   Finish the ugly draft first

Never draft into the tool. If the first version came from the model, there is no original voice left to protect and everything downstream is a negotiation with an average.

02   Mark your own untouchables

Before pasting anything, note the sentences you would fight for. Usually five or six in a two thousand word piece. Those never go in, no matter how the prompt is phrased.

03   Ask for diagnosis, not surgery

First pass is always "point at what is unclear and explain why, without rewriting." This is where these tools earn their keep, and the output costs you nothing in voice.

04   Fix the top three yourself

Take the diagnosis and do the work by hand. It takes longer than accepting a rewrite and it is the entire reason the piece still sounds like you at the end.

05   Send only what is still broken

If two or three passages resist, send them individually with the surrounding paragraphs attached for register, plus an explicit list of what must not change.

06   Read the whole thing aloud, once

Every seam shows up in the voice. A sentence that came from the machine will make you stumble, because it is not phrased the way you speak. Stumbles mark the edits to undo.

Final verdict

Useful editor, unreliable writer

After a few hundred paragraphs and a lot of undoing, the split in my head is clean. As a reader, these tools are excellent. They will tell you which paragraph does not earn its place, spot the sentence where you lost the thread, find the point where you assumed knowledge the reader does not have, and do it in seconds without getting bored on page nine. I use them for that most days and would miss it.

As a writer, they are a downgrade on anything that had a voice to begin with. Not because the output is bad. Because it is average, which is worse in a way that takes longer to notice. The first rewrite I accepted felt like an improvement. The tenth one I accepted was the point where a piece stopped sounding like me, and I only caught it reading a printout in a coffee shop, wondering who had written it.

The practical answer is not to use less AI, it is to change what you hand over. Give it the diagnosis job and keep the prose job. Tell it what must not change before you tell it what should. Protect your openings, your closings and your stories, because those carry the most voice per word and lose the most in the exchange. And if you take one thing from all of this: check the word count. If the rewrite is longer than what you wrote and you did not ask for longer, you almost certainly just traded meaning for polish.

My kitchen paragraph still says the tiler showed up on a Tuesday. It is the better sentence, and it took me four minutes of owning a worse one to work that out.