Signs of AI Writing (and How to Fix Each One)

Readers spot AI writing long before any detector does. They do not run a classifier; they notice that a paragraph took 80 words to say nothing, that every list has exactly three items, and that the conclusion repeats the introduction. Once they notice, trust drops for the whole page, including the parts a human wrote.

This guide lists the signs of AI writing that editors catch most often, explains why each one happens, and shows a concrete before-and-after rewrite for each. The goal is not to “beat detectors”. It is to publish text that sounds like someone who knows the subject wrote it, which is what both readers and search engines reward.

Why AI text has recognisable tells

Large language models predict likely next words. Left to defaults, they drift toward the most statistically safe phrasing: generic openers, balanced hedges, familiar transitions and tidy structures. None of these is wrong on its own. The problem is density. A human writer uses “moreover” occasionally; an unedited model draft uses a transition at the start of nearly every paragraph.

So the tells below are patterns, not proof. One instance means nothing. Five in a single section mean the draft was not edited.

The 12 most common signs of AI writing

1. The throat-clearing opener

AI drafts often begin by announcing that the topic is important, instead of saying anything about it.

Before: “In today’s fast-paced digital landscape, content marketing has become more important than ever for businesses of all sizes.”

After: “Most B2B blogs publish weekly and still get no leads, because their posts answer questions nobody in their market is asking.”

Fix: delete the first sentence of the draft. If the piece still makes sense, it was throat-clearing.

2. Stock vocabulary

Some words appear far more often in model output than in ordinary professional writing: “delve”, “tapestry”, “landscape”, “realm”, “leverage”, “robust”, “seamless”, “navigate the complexities”, “unlock the power of”, “elevate”. Wikipedia’s editor guide on signs of AI-generated text lists many of these, and editors who review a lot of drafts learn them quickly.

Before: “Let’s delve into how you can leverage robust analytics to unlock the full potential of your content.”

After: “Here is how to use your analytics to find which posts actually bring in sign-ups.”

Fix: search the draft for a short banned-word list and replace each hit with the plain word you would say out loud.

3. Everything comes in threes

“Fast, reliable, and scalable.” “Plan, create, and optimise.” Models love triplets because they sound complete. Humans use them too, but not in every sentence.

Before: “Our approach is simple, effective, and scalable, helping teams save time, reduce costs, and drive growth.”

After: “The approach cut our editing time roughly in half. It has not changed costs much.”

Fix: when you see a triplet, ask which one item is actually true and specific, keep it, and support it.

4. Hedging without information

“It’s important to note that results may vary depending on various factors.” This sentence is always true and never useful.

Before: “The ideal posting frequency may vary depending on several factors, including your industry, audience, and goals.”

After: “If you have one writer, publish one well-researched post a week. Two rushed posts will not beat it.”

Fix: replace each hedge with a recommendation and the condition under which it changes.

5. Uniform paragraph rhythm

AI drafts often have paragraphs of nearly identical length, each with a topic sentence, two supporting sentences and a summary sentence. Read it aloud and it drones.

Fix: vary deliberately. Put a one-line paragraph after a long one. Break up a list. Let an important point stand alone.

6. The “not just X, but Y” construction

“This isn’t just a tool, it’s a complete transformation of your workflow.” Used once, fine. Used every few paragraphs, it is a signature.

After: “It replaces the three spreadsheets we used to track briefs.”

7. Summaries that repeat the introduction

Unedited model output tends to end with “In conclusion” followed by a paraphrase of the opening. It adds length and no information.

Fix: end with the next action, a decision rule, or the edge case the reader should watch for. If there is nothing new to say, end at the last useful section.

8. Missing specifics: no names, numbers, or examples

The most reliable sign of all. AI drafts describe categories (“various tools”, “many companies”, “industry experts”) because the model has no first-hand experience to draw on.

Before: “Many tools can help you track content performance.”

After: “We check Google Search Console’s Pages report every Monday, filtered to posts published in the last 90 days, and look for anything with impressions but a position worse than 20.”

Fix: for every general claim, add one concrete instance: a tool, a setting, a number you measured, a mistake you made.

9. Fabricated or unverifiable facts

Models can produce plausible statistics, studies and quotes that do not exist. “According to a recent study, 73% of marketers…” with no source is a red flag to careful readers and a real liability for you.

Fix: every number gets a source link or gets cut. If you cannot find the study in two minutes, assume it does not exist.

10. Over-formatted structure

Bolded phrases in every paragraph, emoji in headings, a bullet list where a sentence would do. Formatting is meant to help scanning; in unedited AI drafts it decorates instead.

Fix: keep bold for the one phrase per section a skimmer must not miss.

11. Even-handedness on everything

Ask a model whether something is a good idea and it will list pros and cons and conclude that it depends. Expert writers take positions.

Before: “Both gated and ungated content have their advantages and disadvantages.”

After: “Ungate your guides. Gating made sense when the PDF was the product; for most B2B blogs it now trades search traffic for a few low-intent emails.” (See our view on gated vs ungated content.)

12. Answers to questions nobody asked

Models pad with definitions the audience already knows. An article for content managers does not need to explain what a blog is.

Fix: write down who the reader is before editing, then cut anything that person would skip.

A 10-minute editing pass for AI drafts

Use this in order. It catches most of the tells above without rewriting from scratch.

1. Delete the first paragraph. Read from the second. Restore only what is necessary.
2. Run a find for your banned list: delve, landscape, leverage, robust, seamless, crucial, “it’s important to note”, “in conclusion”, “not just”.
3. Mark every unsupported number. Source it or cut it.
4. Find three general claims and make each specific with a tool, a setting, a number or an example from your own work.
5. Break one triplet per section. Keep the strongest item.
6. Replace one hedge per section with a recommendation.
7. Rewrite the ending so it gives the reader a next step.
8. Read one section aloud. Fix whatever makes you stumble.

For a fuller process, see our AI content editing workflow and AI content quality control.

What about AI detectors?

AI detectors estimate how statistically predictable a text is. They have well-documented limits:

• False positives. Plain, formulaic human writing, including technical documentation and text by non-native English speakers, can score as “AI”. OpenAI withdrew its own AI text classifier in 2023, citing its low rate of accuracy.
• Easy to shift. Light paraphrasing changes scores substantially without making the text any better.
• No quality signal. A detector score says nothing about whether the content is accurate, useful or original.

Google’s published guidance on AI-generated content focuses on whether content is helpful and made for people, not on how it was produced. It treats using automation primarily to manipulate rankings as spam, whatever the tool. Our guide to low-quality AI content quotes the relevant Google Search Central pages and turns them into a pre-publish checklist.

So the question to ask about a draft is not “will a detector flag this?” It is “would an expert reader learn something here, and would they trust it?” Text that passes that test rarely shows the signs above. For more on the detection side, see our guide to AI content detection.

What most people get wrong

The common reaction to AI tells is to run the draft through a “humaniser” that swaps synonyms and varies sentence length. That treats the symptom. The root cause of AI-sounding text is almost always missing information: no first-hand example, no position, no specific number. A humaniser cannot add what the draft does not contain.

The better workflow is to put the expertise in before generation: a brief with the angle, the examples, the data and the opinion, so the model is drafting around real substance. That is the logic behind detailed content briefs, and it is why heavily briefed AI drafts need far less cleanup than drafts generated from a one-line prompt.

Frequently asked questions

What are the most obvious signs of AI writing?

Generic openers about the importance of the topic, stock words such as “delve” and “landscape”, lists that always have three items, hedges that say nothing, and an absence of specific names, numbers and examples. A single instance proves nothing; a cluster usually means unedited AI output.

Can AI detectors reliably tell if text was written by AI?

No. Detectors can flag human writing as AI and can be shifted by light paraphrasing. OpenAI retired its own classifier in 2023 because of low accuracy. Use them as one weak signal at most, never as proof.

Does Google penalise AI-written content?

Google’s guidance says it rewards helpful, people-first content however it is produced, and treats content made mainly to manipulate rankings as spam regardless of whether AI was used. Thin, unedited AI content tends to perform poorly because it is unhelpful, not because of how it was made.

How do I make AI writing sound more human?

Add what only you know: first-hand examples, specific numbers from your own work, a clear opinion and the edge cases you have run into. Then cut generic openers, hedges and repeated summaries. Synonym-swapping tools change the surface without fixing the underlying emptiness.

For a worked example of fixing generic AI output in one of the most common formats, see our guide to AI letter writing, which includes prompts and an edit checklist for cover letters, complaints, and business letters.