Does your writing read as machine-written?

Why this is not an AI detector

Because AI detection does not work well enough to be used on anyone, and this tool refuses to pretend otherwise.

Detectors classify text as machine-written or human-written, and they get it wrong often enough to do real damage. They flag non-native English writers disproportionately — measured repeatedly, and a predictable consequence of how the classifiers work, since more predictable word choice looks machine-like to a model whether it comes from an algorithm or from someone writing in a second language. Students have been accused on that basis. OpenAI withdrew its own classifier for poor accuracy.

So this tool makes no claim about authorship. It finds specific phrasings, shows you where they are, and suggests an edit. Every flag is a string you can look at and judge for yourself. Nothing is inferred, nothing is probabilistic, and there is no verdict to be wrong about.

The vocabulary tells

Certain words became statistically overrepresented in machine-generated prose, and readers picked up on them fast. Some are ordinary words used at unusual frequency; others are stock phrases that carry no information at all.

Every phrase this tool tracks, and what to write instead

The full list, with a replacement for each. Weight is how strongly the phrase signals machine drafting, not how bad the writing is.

The checker matches these in context rather than as bare strings, which matters more than it sounds. "Robustness of the design" is not the adjective robust; "she navigated to the settings page" is not "navigating the complexities"; "the conclusion of the match" is not "in conclusion,". Flagging those would train you to ignore the tool, which is worse than missing them.

It also matches across intervening words, because that is how people actually write. "In today’s fast-paced digital world" and "navigating the ever-evolving landscape" are the common forms, and a pattern requiring the words to be adjacent would never fire on either — a gap the first version of this tool shipped with and a test now guards.

The structural tells matter more

Vocabulary is what people list, but structure is what readers actually notice — and it survives a find-and-replace, which is why swapping synonyms rarely fixes anything.

Every sentence is the same length

This is the strongest signal there is. Human writing swings: a long sentence working through an idea, then a short one. Then another long one. Generated prose tends to settle around a consistent length and stay there, and the result reads as flat in a way most readers feel without being able to name.

The checker measures the variation in your sentence lengths rather than the average, and flags text where they are unusually uniform. The fix is mechanical: break two long sentences in half, and add a very short one somewhere.

Everything comes in threes

"Faster, cheaper, and more reliable." The rhythm of three is genuinely satisfying, which is why it becomes a reflex — and once you notice it, a page where every list has exactly three items reads as assembled rather than written. Keep the ones where all three items earn their place.

Paragraphs of identical size

Four sentences, four sentences, four sentences. Real arguments do not divide evenly. Let a paragraph be one sentence when the point is short.

The dutiful summary

A closing paragraph restating everything already said, usually opening "In conclusion". It announces the ending instead of writing one. Readers can see where they are on the page.

One paragraph, before and after

The example loaded by the button above. Paste it in and you will see these exact flags.

Before:

Fifty-eight words. It contains a stock opening, "delve into", the not-just-X seesaw, three empty adjectives, "unlocks the potential", "furthermore", "it’s worth noting", "leveraging", "plays a crucial role", "navigating" and "ever-evolving" — and communicates almost nothing.

After:

Forty-one words, no flags, and it actually says what the product does. Notice the fix was not synonyms — it was adding specifics. The phrases disappeared because there was finally something concrete to say instead of them.

What actually fixes machine-sounding prose

Removing flagged phrases is the easy half and the less important one. Text with every tell edited out can still read as generated, because the real problem is not vocabulary — it is that fluent, unspecific writing sounds like a machine because that is what machines produce most reliably.

1. Add something only you could have written. A real number, a specific example, a thing that happened, an opinion you would defend. This single change does more than every other item here combined.
2. Vary the sentence lengths deliberately. Read it aloud. Where you run out of breath, break it. Where it plods, cut a sentence to four words.
3. Cut the first paragraph. Generated drafts warm up. Your real opening is usually the second or third sentence.
4. Replace every praise adjective with a fact. Not "powerful platform" but "handles 40,000 rows without paging".
5. Take a position. Balanced on-the-one-hand prose is the safest thing a model can produce and the least memorable thing a person can write.
6. Delete the conclusion. Most drafts end one paragraph after they should have.

Why "delve", specifically?

It is a real question with a partial answer, and the answer is more interesting than the word.

These models are tuned on human preference ratings — people compare two outputs and pick the better one, and the model learns to produce more of what gets picked. That rating work is done by people, in particular places, with their own regional register. "Delve" is markedly more common in some varieties of English than in American usage, and one widely-discussed explanation is that the preference data carried those habits into the model.

Whether or not that account is complete, the mechanism it illustrates is real: these are learned stylistic habits, not technical necessities. That also means the list changes. As models are retrained and as writers consciously avoid the flagged words, new tells replace old ones. Any list of AI words — including the one on this page — is a snapshot, and treating a fixed list as permanent is how you end up avoiding the 2024 tells while writing the 2026 ones.

When this matters, and when it does not

It matters for anything with your name on it. Published writing, client work, cold outreach, a personal essay, a job application. Readers register "this was generated" faster than most writers expect, and the cost is credibility rather than comprehension.

It matters less for internal documents, first drafts, notes and anything where the job is to transmit information. A meeting summary that says "furthermore" is doing its job.

It is the wrong question for academic integrity. If you are trying to work out whether a student used AI, this tool will not tell you, and neither will any detector — reliably enough to accuse someone. That is not a gap waiting to be filled; it is a property of the problem.

Common questions

My text got zero flags. Does that mean it reads as human?

It means it avoids the specific phrasings on this list. Uniform, unspecific writing with none of the flagged words still reads as generated. Zero flags is a floor, not a finish line.

My human writing got flagged. Is that a bug?

No — it is the point. Human writers use these phrases constantly, especially in corporate and academic registers. The suggestions still apply, because the phrases weaken writing regardless of origin.

Will this help me get past an AI detector?

It is not built for that and we would not recommend relying on it. Detectors mostly key on statistical properties rather than the phrases here. If your goal is passing a detector rather than writing better, this is the wrong tool — and the honest advice is that the detector is probably wrong about you either way.

Why flag em-dashes? I like em-dashes.

So do we — clearly. It only flags unusual density, above roughly a dozen per thousand words, where they stop reading as choices and start reading as a tic.

Is my text stored?

No. It is analysed in memory and discarded — not logged, not saved, not used for training.

Glossary

• AI tell — a phrase or habit that makes prose read as machine-written. A style signal, not evidence.
• AI detector — a classifier claiming to identify machine-written text. Unreliable, and biased against non-native writers.
• Burstiness — variation in sentence length and complexity. Human writing has more of it.
• Perplexity — how surprising the word choices are to a model. Low perplexity reads as machine-like, and also as careful second-language writing.
• Tricolon — a three-part list. Effective once, a tic when every list has three items.
• RLHF — training on human preference ratings. The likely route by which particular stylistic habits entered these models.
• Register — the formality level of a piece of writing. Most tells are register mismatches: ornamental words in plain prose.