brycewang-stanford/humanizer

Humanizer

Strip AI writing patterns from academic text while preserving economics content and formal structure.

Input: $ARGUMENTS — path to file to humanize.

---

Workflow

Step 1: Read the File

Read the target file from $ARGUMENTS. Support .tex, .md, .txt, and .qmd files.

Step 2: Scan for AI Patterns

Check all 24 patterns across 4 categories:

Category 1: Structural Tics (6 patterns)

1. Triplet lists — "X, Y, and Z" appearing 3+ times in a section
2. Formulaic transitions — "Moreover", "Furthermore", "Additionally" as sentence starters
3. Echo conclusions — final paragraph restates every point
4. Uniform paragraph length — all paragraphs suspiciously similar in length
5. Topic sentence + support — every paragraph follows identical structure
6. Numbered/bulleted reasoning — "First... Second... Third..." in prose

Category 2: Lexical Tells (6 patterns)

7. "Delve" — almost never used by humans in academic writing
8. "Landscape" — as metaphor ("the policy landscape")
9. "Crucial/pivotal/vital" — overused intensifiers
10. "Multifaceted" — AI favourite
11. "Underscores" — as verb ("this underscores the importance")
12. "Navigate" — metaphorical ("navigate the challenges")

Category 3: Rhetorical Patterns (6 patterns)

13. Excessive hedging — "it is worth noting", "interestingly", "arguably"
14. Performative enthusiasm — "fascinating", "remarkable", "exciting"
15. False balance — "while X, it is also true that Y" for every claim
16. Hollow acknowledgment — "this raises important questions" without answering them
17. Premature synthesis — summarising before enough evidence is presented
18. Universal agreement — "scholars agree", "it is widely recognised"

Category 4: Formatting Tells (6 patterns)

19. Over-signposting — "In this section, we will discuss..."
20. Excessive parallelism — every sentence in a list has identical structure
21. Definitional opening — starting sections with dictionary-style definitions
22. Disclaimer stacking — multiple caveats before making a point
23. Summary before content — "This section covers X, Y, and Z" at the start
24. Colon-list pattern — "There are three key factors: (1)... (2)... (3)..."

Step 3: Apply Fixes

For each detected pattern:

1. Identify the specific instance with line number
2. Rewrite to sound more natural/human
3. Preserve the academic content and meaning

Academic Economics Adaptation Rules

• Preserve formal structure where genuinely needed (equations, proofs, identification arguments)
• Keep technical terms — do not simplify econometric vocabulary (heteroskedasticity-robust, LSOA fixed effects, hedonic pricing, etc.)
• Maintain citation density — do not remove scholarly references
• Vary sentence structure — mix short and long, simple and complex
• Allow imperfection — real writing has slight asymmetries and personality
• Respect economics conventions — "we estimate", "we find", "column (3) shows" are standard, not AI patterns
• Keep LaTeX intact — do not modify \textcite{}, \parencite{}, \label{}, \ref{}, or equation environments

Step 4: Report Changes

Present a summary of changes:

## Humanizer Report: [filename]
**Patterns found:** N / 24
**Changes made:** N

| # | Pattern | Category | Location | Change |
|---|---------|----------|----------|--------|
| 1 | Triplet lists | Structural | Section 3, para 2 | Varied list lengths |
| 2 | "Delve" | Lexical | Line 47 | Replaced with "examine" |
...

Step 5: Save

Overwrite the original file with the humanised version (the user can review via git diff).

---

Principles

• Preserve meaning. The content must be identical — only the expression changes.
• Do not over-correct. Some formal academic patterns are normal, not AI tells.
• Context-aware. "Moreover" in a proof derivation is fine. "Moreover" as every paragraph opener is not.
• Reversible. User can always git checkout to undo.
• Light touch for good writing. If only 2-3 patterns found, the text is probably fine. Focus effort on heavily AI-patterned text.
• Economics-aware. Standard economics phrasing ("we exploit variation in...", "we instrument...") is conventional, not AI-generated.