The Anti-AI Writing Style Guide

Your readers can smell AI copy in two sentences. Here’s how to avoid it.

The 40 Banned Phrases

If any of these appear in your newsletter, delete them immediately.

The Original 25 (Still Deadly)

1. “Delve” / “delve into”
2. “Landscape” (as in “the AI landscape”)
3. “Tapestry”
4. “It’s worth noting”
5. “In conclusion”
6. “Let’s dive in”
7. “Game-changer”
8. “Revolutionize”
9. “Leverage” (as a verb)
10. “Synergy”
11. “In today’s rapidly evolving”
12. “At the forefront”
13. “In the realm of”
14. “It is important to”
15. “Cutting-edge”
16. “Groundbreaking”
17. “Seamlessly”
18. “Robust”
19. “Holistic”
20. “Unlock the power”
21. “Harness the potential”
22. “Take it to the next level”
23. “Furthermore”
24. “Navigate” (as in “navigate the complexities”)
25. “Paramount”

The New 15 (Added 2025-2026)

These became AI tells as models got better at avoiding the original list. The newer models stopped saying “delve” but picked up new crutches.

26. “Showcasing” (LLMs use this 4x more than humans)
27. “Fostering” (especially “fostering innovation” or “fostering collaboration”)
28. “Pivotal” (replaced “crucial” as the go-to importance word)
29. “Underscore” / “underscores” (as in “this underscores the need for”)
30. “Bolstered” (nobody says this out loud)
31. “Intricate” / “intricacies” (AI loves making things sound complicated)
32. “Vibrant” (especially “vibrant ecosystem” or “vibrant community”)
33. “Interplay” (as in “the interplay between X and Y”)
34. “Meticulous” / “meticulously” (AI overrates carefulness)
35. “Testament” (as in “a testament to”)
36. “Multifaceted”
37. “Noteworthy”
38. “Comprehensive” (as in “a comprehensive guide to”)
39. “Spearheading”
40. “Realm” (as in “in the realm of”)

The Pattern Behind the Patterns

Stanford and UT researchers found that AI systems repeat connective and transitional phrases up to 6x more frequently than human writers in similar contexts. The words themselves change every 6-12 months as models get retrained. The underlying habit doesn’t.

Here’s the rule: if a word sounds like it belongs in a press release or a corporate earnings call, cut it. Real people don’t talk like that.

AI Syntax Patterns (The Grammar-Level Tells)

Banned phrases are easy to catch. Grammar patterns are harder. These are the structural habits that make text feel AI-generated even when every individual word seems fine.

1. The Triple Parallel List

AI defaults to three-item comma lists constantly: “fast, efficient, and reliable” or “innovative, scalable, and user-friendly.”

Humans do use lists. But AI does it 3-5x per paragraph. One per section max. And vary the count. Two items. Four items. Never always three.

2. The Participial Phrase Closer

AI loves ending sentences with “…ing” phrases:

• “The company released the tool, hoping to attract developers.”
• “They redesigned the API, making it easier to integrate.”
• “The update shipped Tuesday, bringing new features to all users.”

Humans do this occasionally. AI does it in 30-40% of sentences. If you spot two in the same paragraph, rewrite one.

Fix: Just split it into two sentences. “The company released the tool. They’re trying to attract developers.” Simpler. Sounds like a person.

3. The Topic-Comment Structure

Almost every AI paragraph follows this pattern: state the topic, explain it, give evidence, summarize. Every single time. Humans jump around. We start with an example sometimes. We bury the point at the end of a paragraph for effect. We write paragraphs that are just one sentence.

Like this.

4. The Hedging Stack

AI hedges constantly because it’s trained to be safe:

• “It’s worth noting that this may potentially impact…”
• “This could arguably be considered…”
• “It remains to be seen whether…”

Pick a side. Say “this will hurt small companies” not “this could potentially have implications for smaller organizations.” Wrong is more interesting than vague.

5. The Symmetric Section Problem

AI writes sections that mirror each other in structure. Three paragraphs, then three paragraphs, then three paragraphs. Each paragraph roughly the same length. Each section with a similar opening.

Humans are messy. One section gets four paragraphs because you had more to say. Another gets two. A third is just a sentence and a code block. That asymmetry reads as human.

6. The Absent “I”

AI almost never uses first person in informational writing unless explicitly told to. Humans naturally say “I think,” “I’ve seen,” “I tried this.” If your piece has zero first-person pronouns, it reads like a textbook.

7. The Nominalization Habit

AI turns verbs into nouns: “implementation” instead of “implement,” “utilization” instead of “use,” “the facilitation of” instead of “help with.” This is called nominalization. It makes sentences longer and more abstract.

Bad: “The implementation of the new framework led to an improvement in performance.” Good: “We implemented the new framework. Performance improved.”

8. The Adverb Overload

“Significantly,” “substantially,” “fundamentally,” “inherently,” “remarkably.” AI sprinkles these like seasoning. Count the adverbs in your draft. If there’s more than one per 200 words, you’re in AI territory.

The 6 Writers to Study

Your voice should be the intersection of these people:

Paul Graham (Essays)

• Steal: Short sentences. Strong opinions. Contrarian takes.
• Example: “The most dangerous thing about bad investors is not the money they burn, but the time.”
• Rule: If you can say it in 5 words, don’t use 15.

Packy McCormick (Not Boring)

• Steal: Conversational tone. Analogies from unexpected places. Builds excitement without hype.
• Example: “Stripe is basically doing to financial infrastructure what AWS did to computing infrastructure, except financial infrastructure is a much bigger market.”
• Rule: Make complex topics feel like a story you can’t put down.

Lenny Rachitsky (Lenny’s Newsletter)

• Steal: Direct. Opinionated. Data-backed. No filler.
• Example: “The #1 mistake PMs make is building features nobody asked for.”
• Rule: Every sentence earns its place. If it doesn’t add value, cut it.

Ben Thompson (Stratechery)

• Steal: Analytical frameworks. Connects dots others miss.
• Example: “The Internet transformed distribution. AI transforms creation. They are complementary, not competitive.”
• Rule: Have a unique lens. Don’t just report. Explain why it matters.

Trung Phan (SatPost)

• Steal: Witty. Pop culture references. Makes business fun.
• Example: “Mark Zuckerberg’s pivot to the metaverse was like Blockbuster launching a streaming service in 2010. Right idea, wrong execution, wrong time.”
• Rule: Humor makes dense topics accessible.

Simon Willison (Blog)

• Steal: Hands-on. “I tried this and here’s what happened.”
• Example: “I ran this prompt against Claude, GPT-4, and Gemini. Here’s what each got wrong.”
• Rule: Show, don’t tell. Credibility comes from doing, not theorizing.

Humor and Wit

The voice is Paul Graham’s directness meets Trung Phan’s comedic timing. That means one earned wry observation per section — not a punchline, not a bit. A dry aside that makes the reader think “yeah, exactly.” Humor that performs is cringe. Humor that just lands is credibility.

Earned vs. Forced

Forced: “OpenAI just dropped a new model. Classic OpenAI, am I right? 😂”

Earned: “OpenAI dropped a new model. The benchmark numbers are great. The benchmark numbers are always great.”

Earned humor comes from noticing something true and understating it. Forced humor announces itself. The difference: forced humor has a setup that signals “here comes a joke.” Earned humor just says the thing.

Where Humor Lives

The dry aside. A parenthetical that punctures the spin. “(The ‘reasoning model’ that got confused by a grocery list shipped last week.)”

The sharp analogy. Makes the reader laugh by connecting two things they’d never connect. “Using LLMs as a drop-in replacement for an API is like hiring a novelist to write your error messages. Technically works. Not what either of you should be doing.”

The understatement. Describe something bad with clinical calm. “The model hallucinated the API endpoint. The endpoint does not exist. The company raised $200M.”

The callback. Reference something from earlier in the piece in a new context. Works when it surprises the reader. Doesn’t work when it’s just repetition.

What Not to Do

• Don’t announce the joke. “Here’s the funny part:” kills every joke.
• Don’t use emoji as a punchline delivery mechanism. Ever.
• Don’t be sarcastic about the reader’s competitors or peers. Punching sideways reads as insecure.
• Don’t do topical humor about things that will age badly in 6 months.
• Don’t apologize for the joke with hedges. “If you’ll allow me a small bit of humor here…” No. Just write the thing.
• Smug is not funny. Smug is someone explaining the joke they’re too clever to tell.

The Test

Read the “funny” sentence out loud. Does it make you smile slightly? Or does it make you cringe? If cringe: cut it. One less joke and a tighter piece is always better than a joke that doesn’t land.

Contrarian is fine. “The consensus is wrong about this and here’s why” is great content. Smug is not — “the consensus is wrong, as usual, which I’ve been saying for years.” The first makes a point. The second is just status signaling.

The Formula

Every article summary in your newsletter should follow this structure:

1. Hook — One sentence that makes them stop scrolling
2. Context — What is this, specifically? (Not vague)
3. So what? — Why should a builder care?
4. Edge — What do you know that they don’t?
5. Action — What should they do about it?

Example (Bad, AI Slop)

“In today’s rapidly evolving AI landscape, a groundbreaking new framework has emerged that promises to revolutionize how developers leverage large language models. It’s worth noting that this cutting-edge tool seamlessly integrates with existing infrastructure, showcasing the vibrant interplay between innovation and practical implementation.”

Example (Good, Human)

“GitAgent packages AI agents into portable containers. Like Docker but for agents. If you’re shipping agents to customers and tired of ‘works on my machine,’ this is the fix. LangChain and CrewAI agents both work. Try it before your competitors do.”

Same information. One sounds like a robot. The other sounds like a friend who just found something useful.

More Before/After Examples

Bad: “Additionally, the platform offers a comprehensive suite of tools that enable developers to build and deploy AI applications with unprecedented efficiency.”

Good: “It also ships a debugger, a deployment CLI, and a cost tracker. The debugger alone saved us two hours last week.”

Bad: “The integration of AI into software development workflows represents a paradigm shift that will fundamentally reshape how engineering teams approach problem-solving.”

Good: “AI coding tools are changing how teams write software. Not in theory. Cursor users at our company ship PRs 40% faster. The junior devs improved most.”

Bad: “This innovative solution addresses the growing need for scalable, enterprise-ready AI infrastructure that can adapt to evolving business requirements.”

Good: “It handles 10K requests/second on a single node. Most AI wrappers fall over at 500. If you’re past the prototype stage, look at this.”

Subject Line Writing for Newsletters

Your subject line decides if anyone reads the rest. Average newsletter open rate is 35-45%. Bad subject lines drop you to 15%. Here’s what works.

The Rules

1. 28-50 characters. Mobile shows 28-50 characters before cutting off. 68% of emails are opened on phones. If your subject line gets truncated, you lose.
2. Specifics beat generics. “Claude now runs on your desktop” beats “Exciting AI updates this week.”
3. Numbers work. “3 tools that replaced our QA team” beats “AI tools for testing.”
4. Questions work (sometimes). “Are coding agents actually faster?” works. “Have you heard about the latest AI developments?” does not.
5. No hype words. “Revolutionary,” “game-changing,” “must-read.” These trigger spam filters AND reader skepticism.
6. Front-load the interesting part. The first 3 words decide if they keep reading. “OpenAI just shipped…” vs “This week in our newsletter about…”

Subject Line Formulas That Work

Subject Lines to Avoid

• “AI Newsletter #47” (zero information, zero curiosity)
• “This week’s top AI stories” (could be any newsletter)
• “You won’t believe what happened in AI” (clickbait, kills trust)
• Anything with “BREAKING” in caps (you’re not CNN)
• Emoji-heavy subject lines (work for B2C, hurt credibility in tech)

Send Time

Send-time optimization is a 15-23% open rate improvement for zero creative effort. For AI/tech newsletters:

• Tuesday-Thursday, 7-9am in your reader’s timezone tends to work best
• Monday mornings compete with too much inbox catchup
• Friday afternoon is a graveyard
• Test your own list. The “best” time varies by audience.

The Structural Tells (What Detection Tools Actually Measure)

AI text gets flagged on two main metrics: perplexity (how predictable each word is) and burstiness (how much sentence length varies). Here’s how to beat both.

How the Detectors Work in 2026

GPTZero is the most accurate detector right now. 99.3% accuracy across 3,000 test samples, with a false positive rate of just 0.24% (about 1 in 400 documents wrongly flagged). Originality.ai came in at 83% accuracy with a 4.79% false positive rate. Copyleaks and Grammarly’s detector are in between.

They all measure similar things:

• Perplexity: How surprising is each word given the previous words? AI picks the most probable next word. Humans are weirder.
• Burstiness: How much does sentence length vary? AI writes sentences of roughly equal length. Humans write a 4-word sentence followed by a 30-word one.
• Vocabulary distribution: AI uses a narrower band of vocabulary. Fewer very common words, fewer very rare words. Everything clusters in the middle.
• Structural consistency: AI paragraphs have uniform structure. Same pattern, same rhythm, same length. Humans don’t do that.

The detectors get updated regularly to catch newer models. GPTZero can detect Claude, GPT-4/5, Gemini, and LLaMA output. But they all still struggle with heavily edited AI text. That’s the gap you exploit.

Sentence Structure

• Never start two consecutive sentences with the same word
• Mix 4-word sentences with 25-word ones. Three sentences of similar length in a row = AI flag
• Use fragments. For emphasis. Like this.
• Start sentences with “And” or “But” sometimes. AI rarely does this.
• Stop ending sentences with participial phrases (”, hoping to…” or ”, making it…”). AI does this 2-5x more than humans.
• Don’t do three-part parallel lists (“fast, efficient, and reliable”) more than once per section

Paragraph Structure

• Vary paragraph length wildly. One-sentence paragraphs exist. Use them.
• Don’t follow topic-sentence > evidence > summary for every paragraph. Lead with the example sometimes. Let the point come last.
• Never write a conclusion that restates what you said. End on a detail, a question, or just stop.

Punctuation

• NEVER use em dashes. AI uses them 10x more than modern human writers. They are the #1 punctuation-level AI tell. Use periods. Use commas. Use parentheses if you must. But not em dashes.
• Semicolons: max one per 500 words
• Don’t bold random words. Don’t use emoji in prose. Don’t use arrows or decorative unicode.

Tone

• Don’t hedge everything. “Try this” beats “it may be advisable to consider implementing”
• Say things are bad when they’re bad. AI is relentlessly positive. Genuine negative opinions signal humanity.
• Mix contractions inconsistently. Sometimes “it’s,” sometimes “it is” in the same piece. Humans do this naturally.
• Include specific names, numbers, dates, dollar amounts. AI writes in plausible generalities. “$0.003 per token” beats “affordable pricing.”
• Swear occasionally if it fits your brand. AI won’t. One well-placed “this is bullshit” signals authenticity stronger than 500 words of careful prose.

The Editing Checklist

Run this on every piece before publishing:

1. Ctrl+F the banned list. All 40 phrases. Kill every hit.
2. Count your adverbs. More than 1 per 200 words? Cut half of them.
3. Check sentence starts. Do any two consecutive sentences start with the same word? Fix it.
4. Measure paragraph lengths. Are they all roughly the same? Break one up. Merge two short ones.
5. Look for participial closers. More than one per section? Rewrite as two sentences.
6. Find the hedges. “Could potentially” becomes “might.” “It remains to be seen” becomes “nobody knows yet.”
7. Verify specifics. Every claim should have a name, number, date, or link. “Many companies” becomes “Stripe, Vercel, and Linear.”
8. Read it out loud. If you wouldn’t say it to a friend over coffee, rewrite it.

The Test

Read every sentence out loud. If you wouldn’t say it to a friend over coffee, rewrite it. If a sentence could appear in any article on any topic, it’s too generic. Kill it.

The goal isn’t to “beat” AI detectors. The goal is to write things only you would write. The detector evasion comes free.