English Text Humanization

# English Text Humanization Skill

## When to Use

- Writing blog posts, articles, marketing copy in English
- Polishing AI-generated drafts to sound human
- Any English text that needs to pass AI detectors
- Content that must feel authentic to readers

## Core Principle

**Specificity vs generality.** Human text references exact things, takes positions, makes mistakes, shows personality. AI text covers all bases, hedges everything, uses the safest phrasings, produces text that could apply to any similar topic. "Surface polish with nothing underneath" is AI's signature.

## Core Concepts

### Burstiness (sentence rhythm variation)

AI writes flat, uniform sentences (15-25 words, clustered). Humans alternate wildly - a 5-word sentence next to a 40-word one.

**Rules:**
- Mix short (3-7 words), medium (10-18 words), and long (20-30 words) sentences
- NEVER write 3+ consecutive sentences of similar length
- Use one-sentence paragraphs for emphasis
- Use sentence fragments deliberately. Like this.
- Vary paragraph length: 1-sentence → 3-sentence → 2-sentence → 4-sentence

**Before (low burstiness - AI-like):**
```
AI retouching has become an important tool for photographers. It allows them to
enhance their images quickly and efficiently. Many professionals are now adopting
this technology in their workflow. The results are consistently impressive and
save considerable time.
```

**After (high burstiness - human):**
```
AI retouching changed everything. Not gradually - overnight. Photographers who
spent three hours on a single portrait now finish in seconds. The quality? Better
than most manual edits. Not all of them, sure. But enough that the 23 retouchers
I talked to last month all said the same thing: "I can't go back."
```

### Perplexity (word unpredictability)

AI picks the most probable next word - consistently low perplexity, every word is the safest statistical choice. Human writing has perplexity spikes: unexpected metaphors, unusual word pairings, surprising verbs.

**Rules:**
- Replace the first word that comes to mind with the second or third
- Use specific nouns: "23 retouchers" not "many professionals"
- Use unexpected verbs: "grab" not "utilize", "ship" not "implement", "kill" not "eliminate"
- Drop in concrete details the AI wouldn't know: names, dates, specific tools
- Numbers > vague quantities. Always.

**Weak (predictable):** "Many users find the tool helpful for their work."
**Strong (surprising):** "84% of beta testers shipped their first edit in under 90 seconds."

---

## Banned Words & Phrases

### Tier 1 - NEVER use (research-confirmed AI markers)

Based on Liang et al. (arxiv 2406.07016) - 15M+ PubMed abstracts analysis. The 10 most effective marker words (highest excess ratio post-ChatGPT):

**across, additionally, comprehensive, crucial, enhancing, exhibited, insights, notably, particularly, within**

Other high-excess words (>5x frequency spike):

| Kill this | Replace with |
|-----------|-------------|
| delve (25.2x spike) | dig into, explore, look at |
| showcasing (9.2x) | showing, demo |
| underscores (9.1x) | highlights, shows |
| tapestry | mix, blend, combination |
| leverage | use, grab, tap into |
| utilize | use |
| enhance | boost, sharpen, improve |
| elevate | raise, lift, push |
| embark | start, kick off, jump into |
| resonate | click, land, hit home |
| landscape | space, market, field |
| multifaceted | complex, layered |
| intricate | detailed, complex |
| interplay | relationship, tension, push-pull |
| cutting-edge | latest, newest, bleeding-edge |
| game-changer | shift, breakthrough |
| revolutionize | reshape, rethink, flip |
| seamlessly | smoothly, without friction |
| illuminate | show, reveal, expose |
| unveil | launch, drop, release |
| remarkable | striking, sharp, wild |

**Key finding from Liang et al.:** unlike content shifts (COVID terms = nouns), LLM excess is almost entirely STYLE words - 66% verbs, 18% adjectives. This is what makes them detectable.

**Note:** "delve" showed 25.2x spike but dropped sharply in 2025 after being called out. LLMs and humans co-evolve - static word lists go stale. The principles matter more than specific words.

### Tier 2 - Avoid (padding / filler)

| Kill this | Replace with |
|-----------|-------------|
| In order to | To |
| Due to the fact that | Because |
| It is important to note | (delete entirely) |
| It is worth mentioning | (delete entirely) |
| At the end of the day | (delete entirely) |
| In today's world | (delete entirely) |
| As a matter of fact | Actually / In fact |
| For all intents and purposes | (delete entirely) |
| needless to say | (delete - if needless, don't say it) |
| a wide range of | many / various |
| in the realm of | in |

### Tier 3 - Banned transitions

| Kill this | Replace with |
|-----------|-------------|
| Furthermore | And / Plus / On top of that |
| Moreover | And / What's more |
| Additionally | Also / And |
| In conclusion | (use "Conclusion" as H2, never "In conclusion") |
| To summarize | (just summarize, don't announce it) |
| It's worth noting | (delete - just note it) |
| That being said | But / Still / That said |
| On the other hand | But / Then again |

---

## Structural Anti-Patterns to Avoid

### The "AI Shape" of Text

- **Symmetrical paragraphs** - every paragraph roughly the same length
- **Symmetrical lists** - every bullet the same structure/length
- **Predictable headers** - "Understanding X", "The Role of Y", "Why Z Matters"
- **The sandwich** - broad claim → supporting detail → restatement, repeated identically
- **Over-structuring** - not everything needs H2 → H3 → bullet list. Sometimes a paragraph is just a paragraph

### Tone Traps

- **Excessive enthusiasm** - "Fascinating!", "Remarkable!", unearned excitement about mundane topics
- **Hedge-heavy** - "It's worth noting...", "While it may seem...", never taking a definitive position
- **Balanced-to-a-fault** - always "On the one hand... On the other hand..." even when one side i