DEV Community: Skryvo AI

Signs of AI Writing: 12 Tells and How to Fix Them

Skryvo AI — Tue, 20 Oct 2026 13:00:00 +0000

The signs of AI writing are patterns a model falls into when no one edits it: stock vocabulary, "not just X, but Y" sentences, everything in threes, inflated claims of importance, and nothing the author could only know first-hand. None of them proves a machine wrote the text. Together they are what readers notice.

I keep a list of words that get drafts rejected. Not a style preference: an actual blocklist, wired into Skryvo's generation pipeline, built up over months of reading AI drafts and noticing the same tells over and over. When a draft trips the list, it does not ship.

That list exists because readers have learned to smell AI writing in about two sentences. They are not psychic. The tells are specific and countable, and once you know them you will start seeing them everywhere, including, uncomfortably, in things you already published. This article is the list written down: what gives AI writing away, why models do it, and the edits that actually fix it.

The short version: the twelve signs below fall into four groups: vocabulary, sentence patterns, structure, and missing substance. Prompting reduces them, a five-step editing pass removes them, and a count of the tells before and after your edits tells you whether the edits worked. Treat the signs as a reason to edit, never as proof of who wrote something.

What are the signs of AI writing?

These twelve come from my own blocklist and from Wikipedia's guide to signs of AI writing, which its volunteer editors maintain while cleaning up machine-written articles.

Sign	What it looks like	The fix
1. Stock vocabulary	delve, leverage, seamless, robust, crucial, showcase, tapestry	Use the plain word, or cut the sentence
2. "Serves as" for "is"	"The tool serves as a hub for…"	"The tool is…"
3. Not just X, but Y	"It's not just a scheduler, it's a strategy."	Say what it is, once
4. The rule of three	"fast, simple and powerful", in every paragraph	Keep the one item that matters
5. Inflated significance	"marks a pivotal moment in the broader landscape"	State what changed and for whom
6. Vague attribution	"Experts say…", "Studies show…"	Name the source and link it
7. Empty "-ing" analysis	"…, highlighting its importance"	Delete the clause
8. The "despite challenges" ending	"Despite challenges, the future looks bright."	End on the last real point
9. Flat rhythm	Every sentence medium length	Mix short and long sentences
10. Formatting tics	Title Case headings, bold labels on every bullet, emoji as bullets	Sentence-case headings, plain lists
11. Hedging	"It's important to note…", "results may vary"	Make the claim or cut it
12. Nothing only you could say	Advice that fits any company on any day	Add a number, story or opinion from your work

One sign proves nothing; people wrote "crucial" long before chatbots did. Several together, in one paragraph, are what make readers stop trusting a page.

Are these signs proof that AI wrote something?

No. Wikipedia's own guide says so: the signs are potential indicators, many human writers show them too, and fixing the surface can hide deeper problems such as invented sources. People are also poor judges; in one study academics called human writing AI 72% of the time. Use the signs to edit your own drafts, not to accuse anyone else.

Why do these signs show up? The four that matter most

Four of the twelve do most of the damage, in rough order of how fast a reader catches them.

1. Watermark words

Some words are so overrepresented in model output that they work like watermarks. The worst offenders on my blocklist: delve, leverage, seamless, robust, elevate, unlock, game-changer, landscape, navigate, tapestry, harness, unleash. Add the stock openers ("In today's fast-paced world", "In the ever-evolving landscape of") and the stock closer, "In conclusion".

None of these words is wrong on its own. The tell is density. A human might use one of them in an article. A model will use six, because they are the statistically safest choices in its training data, and safety is exactly what averaged writing sounds like.

2. Symmetrical structure

Model drafts share a skeleton: an intro that restates the question you asked, three points of nearly identical length, a bullet list whether the content needs one or not, and a summary that repeats the intro. Everything comes in threes. Human structure is lumpier: one point gets four paragraphs because it is the point that matters, another gets one sentence. If every section of a draft weighs the same, readers flag it before they can say why.

3. Flat rhythm

Read the draft aloud. Model sentences cluster around the same length (medium, medium, medium) with a dependent clause in the same position each time. Human writing swerves. Short sentence. Then a longer one that takes its time getting where it is going because the idea needs the room.

4. Nothing only you could say

This is the deep tell. A model writes the consensus of everything it has read, so a pure model draft contains no surprise: no number from your own results, no story from your own customer, no position anyone could disagree with. It hedges instead: "it's important to note", "results may vary". Readers detect this. So do search engines that reward structured author credibility and AI answer engines, because content with nothing new in it earns no citation.

Why does AI write like that?

Two reasons, and neither is a bug. First, a language model predicts the most probable next word, and the most probable word across billions of documents is by definition the most average one. Ask it to write about marketing and you get the mean of every marketing article ever written. Second, the fine-tuning that makes models polite also makes them agreeable: trained to satisfy everyone, they hedge, balance every claim, and sand off anything that might read as an opinion. Average plus agreeable is precisely the voice you recognize as "AI". The New York Times dug into this exact question and landed in the same place: the voice is a product of how models are trained, not a glitch.

Understanding this matters because it tells you what prompting can and cannot fix. You can prompt away some vocabulary and some structure. You cannot prompt a model into having your experience.

How to make AI content sound human

Here is the editing pass I actually run, in order. It is also the free manual humanizer no tool page will tell you about. On a 1,200-word draft it takes me 20–25 minutes.

Cut the watermark words. Keep a real list and search the draft against it. Two minutes, and it removes the loudest signal.
Break the symmetry. Merge two sections, expand the one carrying the argument, and delete the summary paragraph; the article already made its point.
Vary the rhythm. Split every third long sentence. Let one sentence stand alone.
Add one thing per section that only you know: a number from your own work, a failure, an opinion someone could argue with. This is the edit that turns a draft into an article, and no model can do it for you.
Kill the hedges. Every "it's important to note" becomes either a claim or a deletion.

What that looks like in practice:

Before: "In today's fast-paced digital landscape, leveraging AI tools can help businesses unlock seamless content creation and elevate their brand presence."

After: "AI tools write faster than you. That is the whole pitch. The catch is what they write."

Same idea. One of them you have read a thousand times; the other one you have not.

How to prompt AI so it doesn't sound like AI

Prompting gets you maybe 60% of the way, and it is worth taking. The instruction block I use as a base:

"Write in plain, direct English. Vary sentence length; keep some short. Never use these words: delve, leverage, seamless, robust, elevate, unlock, game-changer, landscape, tapestry, harness. No intro that restates the question, no summary paragraph, no 'In conclusion'. Take a clear position instead of presenting both sides. Do not hedge."

It helps. It reliably fails in two places: the model drifts back to its averages after a few hundred words, and no instruction can inject the experience-based specifics from step four above. Prompting reduces the tells; editing removes them.

How do I check if my writing sounds like AI?

Run it through an AI detector, but read the result as a signal, not a verdict. Detectors are genuinely noisy on any single sentence and they false-flag skilled human writers, so never treat one score as truth. What they are good at is measuring change: a draft that scores 95% AI before your editing pass and 30% after tells you the edits landed. That before/after delta is the honest way to use them, and it is how I gate my own posts. Keep the goal straight too: Google's own guidance rewards people-first content however it was produced, you are editing for readers, and the score is just a proxy for them.

The faster check costs nothing: read it aloud. Your ear catches flat rhythm and stock phrasing before any detector does.

Where a tool fits

Everything above can be done by hand, and for one article a week it should be. I publish more than that, which is why I built the checklist into Skryvo itself: the blocklist is enforced at generation time so watermark words never appear, a structure check flags flat rhythm, and every draft ships with the list of tells it found plus a fact-check workflow for AI drafts. The machine does the mechanical passes. The judgment edits, the things only you know, are still yours, which is exactly how it should be.

AI-sounding content is not an AI problem. It is an editing problem with a known checklist. Work the checklist, and the "written by a bot" feeling disappears, because by the time you publish, a bot did not finish it. You did. Fixing these tells removes one signal of four; the other three are about the publisher, not the prose.

First published on skryvo.ai.

AI Makes Up Statistics: We Checked 186 of Its Claims

Skryvo AI — Mon, 19 Oct 2026 13:00:00 +0000

Six in ten. That is how many of the checkable claims in 20 AI-written blog posts I could not confirm, even with a fact-checking pipeline built for exactly that job. And the claims that failed were not the vague ones. They were the precise ones: "47% of email recipients open based on the subject line (Mailchimp)", "91.5% of web pages get zero traffic from Google (Ahrefs)".

The short version: AI makes up statistics, and it does it in a recognisable way: it attaches a plausible number to a real, trusted source. In our test, 105 of 186 checkable claims (56.5%) could not be traced to any source, 8 more were contradicted by one, and every one of the 20 articles contained at least one of them. The fix is not avoiding AI. It is never publishing a number you have not traced to its original page.

Does AI make up statistics?

Yes. A language model writes the most likely next words, and in a post about email marketing, "according to Campaign Monitor" followed by a percentage is very likely text. Whether that study exists is a separate question the model is not answering. Ask it for "statistics and expert insights" and it supplies the shape of evidence, with or without the evidence.

This is not a fringe problem. When the European Broadcasting Union and the BBC tested AI assistants on news questions, 45% of answers had at least one significant issue and 31% had serious sourcing problems. In 2025, Deloitte's Australian firm agreed to a partial refund on a government report after reviewers found references to research that does not exist and a misquoted court judgment.

What percentage of AI-generated claims can you trust?

We wanted a number for the kind of content people actually publish, so we ran the test ourselves. The prompt was the one most people use: "Write a 700-word blog post about [topic]. Include statistics and expert insights." Ten everyday business topics (email marketing, conversion rates, SEO for new websites, SaaS retention and six more), run twice, six days apart, on September 18 and 24, 2026, with DeepSeek V4 Flash, the model Skryvo uses as its main writer. Every draft then went through our production fact-check pass, which pulls out each checkable claim and searches for a source that supports it.

	Run 1 (Sep 18)	Run 2 (Sep 24)	Total
Articles (words)	10 (9,286)	10 (9,359)	20 (18,645)
Checkable claims	88	98	186
Confirmed by a source	28	45	73 (39.2%)
No source found	56	49	105 (56.5%)
Contradicted by a source	4	4	8 (4.3%)
Articles with at least one of those	10 of 10	10 of 10	20 of 20

Two details matter more than the headline. First, 110 of the 113 unconfirmed or contradicted claims contained a number. The model rarely invents opinions; it invents figures. Second, 54 of those 113 arrived with an attribution: a named company, "according to", or "a study found". The claims that looked the most sourced were among the least verifiable.

Read the limits before you quote this. "No source found" is not proof a claim is false. It means an automated search could not confirm it, and one of our lookup services timed out during the second run, so some true claims sit in that bucket. It is one model, one prompt and 20 articles: directional, not a census. That is why I traced a sample by hand.

AI hallucination examples: seven claims we traced by hand

I took seven flagged claims and searched for the original source of each one, the way an editor would.

What the AI wrote	What the source actually says	Verdict
"91.5% of web pages get zero organic traffic from Google (Ahrefs)"	Ahrefs published 90.63% in its original study and 96.55% in its current one. Neither is 91.5%.	Real source, invented number
"According to a 2023 McKinsey Global Survey, 60% of organizations now use AI in at least one business function"	In McKinsey's 2023 report, 60% is the share of organisations already using AI that had adopted generative AI. One third of respondents used generative AI regularly.	Real number, wrong meaning
"Walking meetings boost divergent thinking by up to 81%"	Stanford's 2014 study (Oppezzo and Schwartz): walking raised creativity for 81% of participants on one test; the average gain was 60%. Nobody measured meetings.	Real study, wrong meaning and context
"47% of email recipients open an email based solely on the subject line (Mailchimp)"	The figure circulates as a chart on email-vendor blogs. We found no study behind it and nothing from Mailchimp.	Zombie stat, borrowed source
"According to Intuit, owners who use AI accounting tools save an average of $4,200 per year"	A search of Intuit and QuickBooks pages turned up nothing matching it.	Untraceable
"It takes an average of 23 minutes to refocus after a distraction (UC Irvine)"	Gloria Mark's research found people take about 23 minutes to resume an interrupted task.	Real, slightly reworded
"The ROI on email marketing is often $36 for every $1 spent, compared to $2 for paid social"	Litmus does publish $36 for every $1. We found no source for the $2 comparison.	Real headline, invented comparison

Only two of the seven held up, and both were bent: one reworded, one with an unsourced comparison bolted on. Our automated checker had also missed the Litmus figure, which is exactly why the hand trace matters. Look at the failures: real brands, real studies and real-sounding numbers, recombined. Nothing looks invented, which is why these survive a quick skim.

Why does AI make up facts?

Because it is rewarded for answering, not for knowing. A model predicts text; it has no internal flag that separates "I read this in a study" from "this sounds like something a study would say". Asking for statistics raises the pressure: the prompt says a good answer contains numbers, so it produces numbers. We covered how language models fabricate with confidence — and the workflow for catching it before you publish.

Is AI-generated content accurate enough to publish?

Not by default — but the problem is specific. AI-generated content is accurate enough to publish for structure, explanation and first drafts. It is not accurate enough to publish when it contains a number, a date, a quote or a named source. In our sample, fewer than four in ten checkable claims could be confirmed. A post with ten statistics and that miss rate is not "mostly accurate"; it is a post with six problems, and readers remember the one they catch. Google evaluates authorship and substance, not the model that produced the draft — what matters is whether the page is worth trusting, and one fabricated figure is enough to lose that trust. It is also the fastest way to produce content that fails every signal of the four-signal slop test. For a related question — whether AI detection is accurate — the answer is also nuanced: see what two studies actually found about AI detectors.

How to catch a made-up statistic in ten minutes

Highlight every number, date and named source in the draft. Search the exact figure alongside the source's name — if the only results are other blogs repeating it, you have a zombie stat. Open the original page, check the number against what it actually measured, and cut anything you cannot verify. The full workflow, including the kill-list of viral stats we keep, is in the fact-check guide.

What we do about it in Skryvo

This test is the reason Skryvo's blog writer runs a fact-check pass by default: each checkable claim in the draft is pulled out, searched, and marked confirmed or unconfirmed, with the source attached. Unconfirmed claims are not hidden; they are flagged so you decide whether to cut, fix or keep them. It does not catch everything (it missed the Litmus figure here), which is why it shows you its sources instead of asking for your trust. You can see how it works on the features page.

Use AI for the draft. Trace every number before you publish. If a statistic cannot be found at its source, it does not go in.

First published on skryvo.ai.