AI writing sounds fake because it repeats a small set of phrasing habits at high density. In 29 drafts graded by the ForgeRank published dataset, reviewers logged 594 phrasing hits across 23 distinct families, with em dashes alone accounting for 182 hits in 19 drafts. The tell is repetition, not vocabulary.
Why This Matters
The ForgeRank dataset (https://forgerankai.com/static/data/ai-content-quality-29-drafts.json) shows how concentrated the problem is. One draft carried 73 phrasing hits by itself. That draft didn't sound fake because it used one bad word. It sounded fake because it kept reaching for the same handful of constructions until the reader noticed the pattern before the argument.
Here's the sharper finding. Eight of the nine drafts that scored 7.0 or above still received a heavy-AI-phrasing verdict. The only draft whose phrasing came back clean scored 4.0. Sounding fake and scoring badly are separate problems, and the dataset proves it: a draft can clear the quality bar and still read like a machine wrote it.
That split matters for anyone publishing AI-assisted work. Fixing the score does not fix the voice, and fixing the voice does not fix the score. You have to work both levers.
What Do the Four Scoring Dimensions Actually Measure?
The ForgeRank dataset scores every draft on four dimensions, each on a 10-point scale. Tabulating the tiers behind all 29 rows shows that each step in the distribution is a single dimension moving, not all four rising together. The ten drafts at 4.0 scored 4.0 on all four dimensions. The ten at 4.8 scored 4.0/4.0/7.0/4.0.
Here is how to read that pattern. A draft that lifts one dimension from 4.0 to 7.0 lands at 4.8 overall, because the other three dimensions hold the total down. The three drafts at 7.0 were flat at 7.0 across the board. The four at 7.2 scored 7.0/7.0/8.0/7.0. The single 7.8 draft scored 8.0/7.0/9.0/7.0, and the single 8.5 draft scored 8.0/9.0/9.0/8.0.
The consequence for a writer is blunt: patching your worst dimension moves the number by less than you expect. You need two dimensions at 8.0 or above before the total clears 8.0. The boundary is that this holds for the phrasing and structure dimensions specifically. A single dimension can carry a draft past 7.0 only when the other three already sit at 7.0 or higher, which is why the 7.2 cluster exists at all.
Why Do Em Dashes Get Flagged So Heavily?
Em dashes are the single most-flagged phrasing family in the dataset: 182 hits across 19 of the 29 drafts. That is roughly one in three of all 594 recorded hits coming from one punctuation mark. Reviewers treat the em dash as a tell because AI drafts deploy it to do the same job every time.
The mechanism is substitution. A human writer uses an em dash when a sentence genuinely needs a hard turn mid-clause. An AI draft uses it to bolt an explanatory clause onto a sentence that would read cleaner as two sentences. Because the pattern repeats, the reader starts predicting the dash before it arrives, and prediction is what registers as fake.
For the writer, the fix is mechanical. Count the em dashes in a draft. If they cluster in the middle third, the draft is using the dash as a default connector rather than a deliberate choice. The boundary: this rule stops applying in fiction and long-form narrative, where em dashes carry dialogue rhythm that a period cannot. In a product page or a report, they are almost always replaceable.
What Separates a Clean-Phrasing Draft From a High-Scoring One?
The dataset's clearest result is that phrasing and quality diverge. Eight of the nine drafts scoring 7.0 or above still drew a heavy-AI-phrasing verdict. The one draft with clean phrasing scored 4.0. Two rival beliefs explain this, and only one survives contact with the numbers.
The common belief is that good writing naturally sheds AI phrasing. If the argument is strong and the structure is sound, the voice takes care of itself. The evidence-backed belief is that phrasing is an independent variable. A draft can be well-argued, well-organized, and still carry 40 phrasing hits, because the habits that produce those hits live below the level of argument.
The dataset resolves it. The 4.0 draft with clean phrasing had no argument to carry. The 7.0-and-above drafts had arguments but kept the habits. Fixing one axis does not move the other, which is why the ForgeRank methodology page treats them as separate passes. Run a phrasing pass and a quality pass, and expect them to disagree.
Why Do Detectors Flag Human Writers Too?
Detector false positives land hardest on non-native English writers. Research by Weixin Liang and co-authors, published on arXiv as 2304.02819 and later in Patterns, found that roughly 61 percent of TOEFL essays written by humans were flagged as AI-generated. That number sits above what any honest detector should produce on genuine human work.
The mechanism is the same density logic that flags AI drafts. Detectors score surface patterns, and a writer working in a second language reaches for safe, repeated constructions for the same reason an AI does: predictability. Because the detector cannot tell the source of the predictability, it attributes it to a machine.
The consequence for a writer is that "sounds fake" is not a reliable signal on its own. If your draft reads flat, the cause could be a phrasing habit, a thin argument, or a style that a detector misreads. The boundary: this false-positive problem applies to detection tools, not to human readers. A person who reads 50 drafts a week catches the repetition directly, which is why the ForgeRank dataset uses reviewer verdicts rather than a detector score.
What Most People Get Wrong About AI Phrasing
The widespread assumption is that AI writing sounds fake because of the words it picks. Swap the vocabulary, the thinking goes, and the draft reads human. The ForgeRank dataset contradicts this. The 594 hits span 23 distinct families, and the flagged constructions are ordinary English: em dashes, transition phrases, parallel clause shapes. None of them are exotic.
The misconception persists because vocabulary is the visible layer. A reader notices "delve" or "tapestry" and blames the word. The real signal is frequency. A draft with 73 hits is not using 73 unusual words. It is using a small set of common constructions 73 times, and the repetition is what the reader's brain catches. Google's own guidance in Search Central frames the adjacent problem in production terms, noting that scaled content abuse covers pages generated "primarily for the purpose of manipulating search rankings" rather than helping people, which is a volume problem, not a vocabulary one.
The correction is to audit for repetition, not for banned words. Count how many times a construction appears. That count is the tell.
Key Takeaways
- The ForgeRank published dataset recorded 594 phrasing hits across 23 families in 29 drafts, with em dashes at 182 hits in 19 drafts.
- Eight of the nine drafts scoring 7.0 or above still drew a heavy-AI-phrasing verdict, and the only clean-phrasing draft scored 4.0.
- A single dimension moving from 4.0 to 7.0 lands the total at 4.8, because the other three dimensions hold it down.
- Detector false positives hit non-native writers hard: roughly 61 percent of human-written TOEFL essays were flagged, per Liang et al. on arXiv.
- Fixing the score and fixing the voice are separate passes, and they will disagree.
Frequently Asked Questions
Why does AI writing sound fake even when the content is accurate?
Accuracy and voice are separate dimensions. The ForgeRank dataset shows eight of nine drafts scoring 7.0 or above still carried a heavy-AI-phrasing verdict. The content held up; the repeated constructions did not.
What is the most common AI phrasing tell?
Em dashes. The ForgeRank published dataset logged 182 em-dash hits across 19 of 29 drafts, the largest single family out of 594 total hits. The tell is the frequency, not the punctuation itself.
Can a draft score well and still sound like AI?
Yes, and the dataset proves it. Every draft scoring 7.0 or above except one still drew a heavy-AI-phrasing verdict. The single clean-phrasing draft scored 4.0, which means the two problems move independently.
Do AI detectors wrongly flag human writing?
They do, and the effect falls hardest on non-native English speakers. Liang et al., published on arXiv as 2304.02819, found roughly 61 percent of human-written TOEFL essays were flagged as AI-generated.
Does Google penalize AI-written content?
Google's Search Central guidance targets scaled content abuse, meaning pages generated "primarily for the purpose of manipulating search rankings" rather than helping people. AI authorship alone is not the violation.
How do I find AI phrasing in my own draft?
Count repeated constructions rather than hunting for banned words. The dataset's flagged families are ordinary English, so the signal is how many times each one appears, not whether it appears.
Open your current draft and count every em dash. If the number is above five, replace each one with a period or a comma and read the paragraph aloud.
Why does AI-generated content sound generic?
Because the phrasing families repeat at a rate a human writer does not reach. Across the 29 drafts in the ForgeRank published dataset the phrasing pass recorded 594 hits spanning 23 families, and em dashes alone accounted for 182 of them across 19 drafts. A generic read comes from that concentration rather than from any single phrase, which is why hunting for one banned word misses it.
The measurement also separates generic from weak. Eight of the nine drafts scoring 7.0 or above still came back flagged heavy on AI phrasing, so a draft can carry real substance and still read generic. The one draft whose phrasing came back clean scored 4.0.