bhushan agrawal

there's a point where you read enough ai-generated writing that you stop noticing individual sentences.

you start noticing the pattern.

the grammar is clean. the paragraphs are neatly arranged. every point connects to the next one. nothing is obviously wrong.

and yet, somewhere around paragraph four, you want to close the tab.

that's the strange thing about ai slop. it often isn't bad in any obvious way. it just has a certain texture.

there isn't one phrase that proves something was written by an llm. these are signals, not a test. humans write this way too, and ai-generated text can be heavily edited until it doesn't look like this at all.

but once you know what to look for, some patterns get hard to unsee.

1. "not x, but y"

it's not about working harder. it's about working smarter.

this isn't just a tool. it's a new way of thinking.

you're not building a product. you're building an experience.

nothing wrong with this construction on its own. humans use it constantly.

the tell is repetition. when an article keeps turning ordinary statements into neat rhetorical contrasts, the writing starts feeling templated. one example means nothing. five in a short article are worth noticing.

2. everything comes in threes

three benefits. three examples. three reasons. three adjectives describing the same thing.

fast, simple, and reliable.

the rule of three is an old rhetorical trick, not an ai invention. what gives it away is frequency when every section has another trio, the piece starts feeling assembled rather than written.

3. synonym cycling

one idea, repeated with different vocabulary.

the system is fast, efficient, and performant.

then:

it saves time, reduces effort, and improves productivity.

then:

developers can build faster, work smarter, and get more done.

the wording changes. the idea doesn't.

llms are good at producing linguistic variation, which can make repetition look like new information. ask yourself: did this paragraph add an idea, or just rename the last one?

4. generic specificity

some sentences sound informative without telling you much.

businesses can leverage modern tools to streamline workflows and improve operational efficiency.

which businesses? which tools? what actually changed?

same with an impressive number that has no source attached:

companies can reduce costs by 20–40%.

a range isn't proof of ai. it's just an unsupported claim, no matter who wrote it. specific details are harder to fake:

postgres with read replicas cut our average query time from x to y.

that's something you can actually go check.

5. suspiciously perfect transitions

now that we've covered x, let's take a look at y.

with that in mind, let's explore...

this brings us to another important consideration.

on their own, none of these are bad. when every section needs one, the writing starts sounding assembled from a template. human writers don't always announce where they're headed next sometimes they just move on.

6. sounds like a person, isn't one

the writing is friendly, confident, polished. but you learn almost nothing about the person behind it. no unusual observation, no strong preference, no detail that makes you think someone actually spent time with the subject.

instead:

developers often struggle with...

businesses are increasingly looking for...

users want a simple, intuitive experience...

might all be true. also incredibly safe.

7. everything is perfectly balanced

on one hand, x. on the other hand, y.

while x has advantages, it also has limitations.

real writing is usually messier. a human might spend 700 words annoyed about one small feature and barely mention another. when every argument gets a perfectly polished counterargument, the prose starts to feel generated.

8. the conclusion repeats the article

intro says what the piece will explain. body explains it. conclusion summarizes it. then the last line summarizes the summary.

ultimately, the key is...

at the end of the day...

the bottom line is...

a conclusion is fine it just shouldn't exist purely to remind you what you already read.

what about em dashes and "ai words"?

the internet loves lists of supposed ai tells delve, nuanced, landscape, robust, leverage, the em dash itself.

none of it proves anything. humans use these too. vocabulary and punctuation can feed into a broader stylistic pattern, but treating any single word as a detector doesn't hold up the differences between human and llm writing show up across vocabulary, syntax, and repetition together, and they shift by model, genre, and context.

so what should you actually look for?

don't hunt for one forbidden word. look for combinations:

repeated "not x, but y" constructions
constant groups of three
the same idea restated in different words
generic transitions
broad claims with few concrete details
unsupported precision
perfectly balanced sections
polished but strangely impersonal writing

one of these means nothing. several together are worth noticing and even then, you're recognizing a style, not proving authorship.

the real opposite of ai slop is specificity

you don't make writing human by adding typos or deleting every em dash. you make it useful by having something specific to say.

tell me which feature broke. show me the actual error. tell me what surprised you. say which part you dislike and why.

give me the detail that wouldn't survive being copied into ten other articles. that's usually what generic ai writing is missing not perfect grammar, but a reason to believe someone actually thought about the thing they're writing about.

sources:

russell, karpinska & iyyer, "people who frequently use chatgpt for writing tasks are accurate and robust detectors of ai-generated text" (acl 2025).
zanotto & aroyehun, "linguistic and embedding-based profiling of texts generated by humans and large language models" (emnlp 2025).