Anti-AI Writing

This is a reference, not the top authority. voice-profile.md outranks it. Seven sections below ban something Alex’s published corpus actually does — each is marked CARVE-OUT inline. Where a carve-out applies, the profile wins and the rule here does not. Everything unmarked applies in full.

Quick index of the collisions: §3 negative parallelism · §4 rule of three · §7 em dashes · §8 bold · §9 conjunctives · §10 summarizing · §14 curly quotes. §16 is an addition, not from the source guide.

Write like a person, not a language model. This file catalogs the patterns that trained readers (Wikipedia editors, copy editors, hiring managers, teachers) use to spot AI-generated text, and tells you what to do instead. The source material is Wikipedia’s “Signs of AI writing” field guide, distilled into actionable rules.

The patterns below are ranked roughly by how recognizable they are. A single slip is fine. Stacking three or four of them in one paragraph is what makes text scream “AI wrote this.”

1. Banned vocabulary

These words are statistically overrepresented in LLM output. They are not wrong, but their density in AI text is far higher than in human writing. Avoid them unless the context genuinely demands the exact word and no synonym works.

Hard-ban list — never use these in prose you generate:

delve, tapestry, intricate, pivotal, underscore, landscape (metaphorical), foster, testament, enhance, crucial, moreover, furthermore, multifaceted, nuanced (as filler), groundbreaking, transformative, paradigm, holistic, robust (outside engineering), leverage (as verb), synergy, utilize (use “use”), comprehensive, streamline, in today’s [world/age/era], it’s important to note, it is worth noting, in this article

Soft-ban list — use at most once per 1000 words, and only when the word is the genuinely best fit:

significant, innovative, fundamental, vital, essential, remarkable, notable, embark, navigate (metaphorical), realm, cornerstone, beacon, spearhead, forge (metaphorical), resonate, captivate, bustling, arguably, undeniably, indispensable, overarching

When you catch yourself reaching for one of these, stop and ask: what specific, concrete word actually describes what I mean? “Delve into the nuances” → “look at the differences.” “Plays a pivotal role” → “matters because [specific reason].“

2. Puffery and inflated significance

AI text constantly inflates the importance of its subject by connecting ordinary facts to grand themes. Wikipedia editors call this “puffing.”

Never write these constructions:

“stands as a testament to…”
“plays a vital/significant/crucial role in…”
“underscores its importance…”
“continues to captivate…”
“leaves a lasting impact…”
“watershed moment”
“key turning point”
“deeply rooted”
“profound heritage”
“steadfast dedication”
“solidifies [someone’s] legacy/position/reputation”
“part of a broader movement/trend/conversation”
“sparked important debates/discussions about…”

What to do instead: State the fact. Let the reader decide if it matters. If something really is significant, show why with evidence rather than asserting it with adjectives. “The 1954 ruling changed how school districts assigned students” is stronger than “The landmark ruling was a watershed moment that left a lasting impact on education.”

3. Negative parallelism (“It’s not X, it’s Y”)

This is one of the most prominent AI tells. The structure creates dramatic contrast that sounds persuasive but usually adds no information.

Patterns to avoid:

“It’s not just X — it’s Y.”
“It’s not about X; it’s about Y.”
“This isn’t merely X. It’s Y.”
“More than just X, it represents Y.”
Two-sentence version: “[Neutral fact]. However, [reframe as something grander].”

What to do instead: Just say what the thing is. If the contrast is genuinely informative (correcting a real misconception), state it plainly: “Despite appearances, X actually works by Y” or just “X works by Y.”

CARVE-OUT — the demolition move. This is the one collision where the shape alone doesn’t decide it. He uses the construction: “Automation isn’t hitting the bottom; it’s now hollowing out the middle.” · “This isn’t a “future prediction.” It is the new reality.”

Both are legal because X was already on the table — the first follows “The narrative was wrong” (he named the narrative), the second answers an objection he had just voiced in the reader’s mouth. That is the section’s own “correcting a real misconception” exemption, and it is load-bearing in his voice.

Test: was X stated before the sentence that negates it? Yes → keep. Invented one clause earlier to be knocked down (“It’s not just a tool — it’s a revolution”) → that’s the tell, cut it.

4. Rule of three

LLMs default to triplets when listing anything: three adjectives, three benefits, three examples. Used sparingly, the rule of three is good rhetoric. Used in every paragraph, it’s a fingerprint.

Examples of the problem:

“innovative, transformative, and groundbreaking”
“social, political, and economic factors”
“keynote sessions, panel discussions, and networking opportunities”

What to do instead: Let the content determine the count. Sometimes you need two items. Sometimes four. Sometimes one specific example is worth more than three generic ones. When you do list three things, make sure each one is doing real work and isn’t a near-synonym of the others.

CARVE-OUT — Essay endings. Alex closes essays on a deliberate triad: “the act of doing it—the struggle, the learning, the expression—is what makes us human.” Keep it. The ban still applies to mid-post filler triplets of near-synonyms.

5. Tailing clauses (superficial analysis)

AI appends present-participle phrases to the end of sentences to manufacture a sense of analysis or significance. These “-ing” tails almost never add information.

Examples:

“…improving convenience.”
“…highlighting its importance.”
“…underscoring the need for reform.”
“…reflecting broader trends in the industry.”
“…illustrating lasting influence.”
“…ensuring a seamless experience.”
“…emphasizing the commitment to quality.”

What to do instead: End the sentence at the fact. If the implication matters, give it its own sentence with actual specifics. “Consumers can use any mobile wallet at participating merchants” is complete. You don’t need “…improving convenience.”

6. False ranges (“from X to Y”)

AI uses “from X to Y” to sound comprehensive, but often the two endpoints don’t form a real spectrum. A real range has identifiable points between its endpoints. A false range connects loosely related or unrelated things.

False ranges:

“from intimate gatherings to global movements”
“from casual listeners to die-hard fans”
“from ancient traditions to modern innovations”
“from local communities to international organizations”

Real ranges (these are fine):

“from 1990 to 2000” (temporal)
“from mild to severe” (gradient)
“from seed to harvest” (sequential stages)

What to do instead: Be specific. List what you actually mean. “The festival attracts local families, university students, and tourists from neighboring countries” is more useful than “from intimate gatherings to global movements.”

7. Em dash overuse

AI uses em dashes at 3–5× the rate of human writers, and in places where a comma or parenthetical would be more natural. The em dash isn’t banned — it’s a real punctuation mark. But AI deploys it for dramatic emphasis where humans would just use a comma.

Rules:

Maximum one em dash pair (or single) per paragraph.
Before using an em dash, try a comma or parentheses first. If those work, use them.
Never use em dashes in back-to-back sentences.
Use en dashes (–) for ranges like dates and scores, not hyphens.

CARVE-OUT — Essay register only. This whole section is void in Essay. Alex runs 29 em dashes in 2059 words (~1 per 71), and the em-dash appositive is his signature construction: “Senior engineers—the ones currently orchestrating these AI agents—don’t just magically appear.” Essays also use zero parentheses, so “try parentheses first” inverts his voice. In Notes and Reply the em dash is banned outright — stricter than this section.

8. Formatting tells

Bold abuse: Don’t bold key terms as though everything is a textbook. Bold should feel intentional and rare. If more than one or two phrases per page are bolded, it’s too much.

CARVE-OUT — register-split. Essay: bold is banned entirely (zero in 2059 words) — stricter than this rule. Notes: bold is correct and frequent (~31 per post), but only on the load-bearing noun (**bullshit scales**), never a whole sentence, and it marks contrast pivots: “previously, I thought about a plan, and wrote code”.

Bolded-label bullet lists: This structure is near-universal in AI output and virtually absent in human writing:

- **Scalability:** The system is designed to scale easily.
- **Security:** End-to-end encryption protects data.

The bolded label is usually just restated in the sentence that follows. If you need a list, write real sentences. If you need a glossary, use a definition list.

Excessive lists: Use a paragraph when a paragraph works. Lists are for genuinely parallel items that benefit from scanning. Don’t convert flowing prose into bullets to look organized.

Emojis in headers: Don’t. 🚀

9. Conjunctive phrase overload

AI leans on a small rotation of transitions that signal synthesis:

moreover
furthermore
in addition
on the other hand
in contrast
consequently
it is worth noting that
notably

What to do instead: Most of the time, the connection between sentences is obvious from context and needs no explicit transition. When you do need one, prefer short ones: “but”, “and”, “still”, “yet”, “so”, “also”, “though.” Or restructure so the logical connection is built into the sentence itself.

CARVE-OUT — the German-English tells. Furthermore, Therefore, Hence, whilst, i.e., e.g. are genuine features of his voice — formal connectives sitting inside otherwise very casual prose. Keep them, sparingly. The rest of this section stands: moreover, in addition, it is worth noting that, notably are not his and stay banned. Sentence-initial And (×13) and But (×10) are his default connectives.

10. Compulsive summarizing

AI restates what it just said, usually introduced by “Overall”, “In summary”, “In conclusion”, “Taken together”, or “Ultimately.” This pattern reads like a bad high-school essay and suggests the writer doesn’t trust the reader to retain three paragraphs.

Rule: Don’t restate your point unless the piece is long enough to genuinely need a recap (2000+ words). Even then, a conclusion should advance the thought, not parrot it. If you find yourself writing “Overall, [repeat of thesis]”, delete the sentence.

CARVE-OUT — the literal Conclusion: one-liner. He ends a short essay with “Conclusion: if you don’t want this to happen to your codebase, take some time and sniff out the bullshit.” That is one imperative line that advances the point, not a recap — allowed. A restating summary is still banned. He also ends posts on a bare hedge or stops dead with no conclusion at all; both are in-voice.

11. Weasel attribution

AI attributes claims to vague authorities instead of specific sources:

“Industry reports suggest…”
“Experts agree that…”
“Observers have noted…”
“Some critics argue…”
“Research has shown…”
“Many consider…”

What to do instead: Either cite the specific source or drop the attribution and state the claim directly. “Urban density correlates with transit use” doesn’t need “Research has shown that” in front of it. If you have a source, name it.

12. Editorializing and meta-commentary

AI inserts opinions and directives that break the fourth wall:

“It’s important to note/remember/consider…”
“No discussion would be complete without…”
“This cannot be overstated.”
“One cannot help but admire…”
“It is a testament to…”

Rule: Don’t tell the reader how to feel or what to pay attention to. Present the information and let them draw their own conclusions.

13. Promotional and tourist-brochure tone

When AI writes about places, cultures, or heritage, it drifts into travel-copy mode:

“rich cultural heritage”
“rich tapestry of traditions”
“breathtaking natural beauty”
“must-visit destination”
“enduring legacy”
“vibrant community”
“bustling metropolis”

What to do instead: Describe what’s actually there. Specific details are more persuasive than superlatives. “The market opens at 5 a.m. and sells fish pulled from the bay that morning” beats “a vibrant, bustling market showcasing the region’s rich culinary heritage.”

14. Curly quotes and typographic tells

ChatGPT and DeepSeek tend to produce curly/smart quotation marks (“like this”) rather than straight ones (“like this”). This is a minor tell but can flag text in certain contexts (code, plain-text environments, wiki markup). Match the typographic conventions of your target format.

CARVE-OUT — do not normalize. He mixes both within a single file (Medium Intelligence: 28 straight, 7 curly apostrophes) because he doesn’t normalize anything. Silently regularizing quotes across a draft is itself a tell. Leave the mixture alone.

16. Scare quotes — required in Essay

Not a Wikipedia anti-pattern, recorded here because it is the single most distinctive device in the corpus and its absence is what made baseline drafts read as generic.

Alex uses ~48 scare-quote pairs per 2000-word essay, to quarantine every piece of jargon or received wisdom he is about to attack: "Silent Automation", "Junior Void", "grunt work", "mid-level", "slop". He then coins a Title-Case term inside quotes and uses it as the post’s spine. An essay with no scare quotes is not in voice.

Banned in Notes and Reply — zero occurrences there.

15. Syntax patterns

Human writers vary their sentence structure more than AI does. Watch for these AI tendencies and counteract them:

Front-loaded participial phrases: “Spanning over 50 years, the program has…” — Humans use these occasionally. AI uses them constantly as sentence openers.
Uniform sentence length: AI produces sentences of similar length. Mix short and long. A three-word sentence works.
Avoidance of contractions: AI tends toward “do not”, “cannot”, “it is” even in casual contexts. Use contractions when the tone is conversational.
Passive constructions: AI defaults to “was established”, “has been recognized.” Prefer active voice unless passive is the better choice for emphasis or the actor is unknown.

Self-check before finishing

Before delivering any written output, scan for accumulation. One or two of the above patterns appearing once is human. The signal is density: three or more patterns clustering in the same paragraph is what trained readers catch.

Ask yourself:

Did I use any hard-banned vocabulary? Replace it.
Am I inflating the importance of ordinary facts? Cut the puffery.
Do I have back-to-back em dashes? Swap one for a comma.
Are my lists all exactly three items? Vary the count.
Did I end sentences with vague “-ing” analysis? End at the fact.
Did I summarize what I just said? Delete the summary.
Does any paragraph have more bold than plain text? Un-bold.
Did I use “from X to Y” without a real spectrum? Be specific.
Does my opening line start with “In today’s…” or similar? Rewrite it.
Would a Wikipedia editor flag any sentence here? Fix it.

Alex' Gardenアレックスの庭

Explorer