Why AI Writing Keeps Sounding Like AI Writing

The DSL Project, Part 1 of 4

AI-generated writing has a strange problem.

The grammar is usually fine, and the sentences make sense. The information is often organized better than what many people would write themselves.

After enough exposure, though, you start recognizing a familiar pattern. It is rarely a single phrase or punctuation habit that gives it away. You recognize the shape.

The Obvious Fixes Did Not Fix It

We started where most people would. Tell the AI to sound more natural. Make it less formal. Vary the sentence length, stop the repetition, avoid certain phrases. Cut back on the em dashes.

Some of that helped. None of it solved the problem. The wording changed, but something underneath it stayed the same. A casual answer could still sound generated. So could a more direct one. Even after removing the obvious phrases people associate with AI writing, the same strange familiarity remained.

That was when we stopped looking primarily at the words.

The Problem Was Hiding in the Structure

Once we started paying attention to sentence and paragraph structure, the patterns became difficult to miss. AI likes balance. An idea is introduced, explained, clarified, reinforced, and often closed by restating it. Sentence structures repeat, while contrasts tend to resolve with suspicious neatness.

The problem was not X. It was Y.

Lists tend to give every item similar treatment. Transitions announce what is coming next. Conclusions often summarize material the reader just finished reading. None of those habits are automatically bad. Frequency is the problem. Use a construction once and it may be effective. Repeat the same architecture across unrelated topics and it starts becoming a signature.

Human Writing Is Messier Than That

People do not normally write with perfect structural consistency. We spend more time on some ideas than others. We assume the reader can follow an implication. A dense paragraph may end abruptly.

Sometimes with three words. We abandon patterns before completing them or one bullet may be much longer than the others because the idea needs more room. The obvious part gets passed over while the important part receives attention.

AI tends to resist that kind of unevenness. Once it begins a pattern, the next likely continuation is often the one that completes it. A comparison invites a second side. A list invites similar treatment for each item. An explanation invites a conclusion. The model keeps extending the structure it has already started.

Start a comparison and it wants both sides. Introduce several ideas and it tends to give them similar treatment. Make an important point and the model may explain it again. The result is orderly, sometimes more orderly than the subject requires.

We Had Been Solving the Wrong Problem

Once we started looking at structure instead of surface style, we realized we had been solving the wrong problem. We tried better tone instructions, lists of phrases to avoid, and examples of the voice we wanted. Those things still have value, but they do not reach deep enough.

A casual tone can sit on top of repetitive structure. Humor can too. Technical language, profanity, fragments, deliberate informality—none of them necessarily change the machinery underneath.

We needed a way to control that machinery.

“Write Naturally” Is Not a Useful Instruction

One of the biggest problems with AI writing prompts is that the instructions sound clear until you try to define them. Words such as natural, smooth, and less robotic describe reactions. They do not identify a behavior. We needed something more precise.

Did the response repeat a point that was already clear?

Did several paragraphs follow the same progression?

Did the model explain an implication the reader could reasonably infer?

Did the ending summarize the article because the article needed a summary, or because the model tends to produce one?

Those questions point to observable behavior.

The First Rule Was Simple

We began comparing outputs and pulling the repeated structures apart. When a pattern kept appearing, we isolated it from the surrounding writing and asked what, exactly, was happening. Was the model restating a point that was already clear? Completing a pattern the reader no longer needed? Repeating the same paragraph rhythm? Resolving contrasts the same way?

Each recurring behavior became a separate problem we could describe, observe, and test.

For example:

• Do not restate a point after it has already been made. Allow some conclusions to remain implied.

• Break a long, dense passage with a short sentence, or interrupt a paragraph rhythm that has started repeating.

• Suppress mirrored contrasts when they begin becoming a pattern.

Each behavior was narrow enough to observe and test independently. That is where the DSL began. We were trying to solve a smaller problem than “make AI writing sound human.”

The question was more specific: Why did the writing still sound generated after we had already fixed the words?

Next: Part 2 — We Stopped Editing Words and Started Studying Structure