brycewang-stanford/ai-writing-humanizer

AI Writing Humanizer

A skill for identifying and removing characteristic patterns of AI-generated text to produce natural, authentic academic writing. Designed for researchers who use AI tools for drafting and want to ensure the final output reads as genuine scholarly prose.

Common AI Writing Patterns

Lexical Patterns to Identify and Replace

AI-generated text frequently overuses certain words and phrases:

def identify_ai_patterns(text: str) -> dict:
    """
    Scan text for common AI-generated writing patterns.

    Returns a report of detected patterns with suggested replacements.
    """
    overused_phrases = {
        # Hedging/filler phrases AI overuses
        'it is important to note that': 'Note that',
        'it is worth mentioning that': '[delete or rephrase]',
        'it should be noted that': '[delete or rephrase]',
        'in the realm of': 'in',
        'in the context of': 'in / for / regarding',
        'a testament to': '[rephrase with specific evidence]',
        'the landscape of': '[delete -- be specific]',
        'a nuanced understanding': '[delete or specify what nuance]',
        'shed light on': 'clarified / revealed / explained',
        'delve into': 'examined / analyzed / investigated',
        'furthermore': '[vary: also, additionally, moreover, or restructure]',
        'moreover': '[vary: in addition, also, or restructure]',
        'utilizing': 'using',
        'leverage': 'use / apply / employ',
        'facilitate': 'enable / support / help',
        'a myriad of': 'many / numerous / various',
        'plays a crucial role': 'is important for / contributes to',
        'in conclusion': '[often unnecessary -- just conclude]',
        'overall': '[often unnecessary filler]',
        'comprehensive': '[usually vague -- be specific about scope]',
        'robust': '[overused -- specify what makes it strong]',
        'multifaceted': '[specify the actual facets]',
        'notably': '[usually filler -- delete or restructure]'
    }

    results = {'detected': [], 'total_flags': 0}

    text_lower = text.lower()
    for phrase, suggestion in overused_phrases.items():
        count = text_lower.count(phrase.lower())
        if count > 0:
            results['detected'].append({
                'phrase': phrase,
                'count': count,
                'suggestion': suggestion
            })
            results['total_flags'] += count

    return results

Structural Patterns

AI text tends to exhibit predictable structural patterns:

AI Pattern: Formulaic paragraph structure
  - Topic sentence (broad claim)
  - Supporting point 1
  - Supporting point 2
  - Concluding/transition sentence
  Every paragraph follows this exact template.

Human Fix: Vary paragraph structure
  - Sometimes lead with evidence, then interpret
  - Sometimes pose a question, then answer it
  - Sometimes use a single punchy sentence as a paragraph
  - Let paragraph length vary naturally (2-8 sentences)

AI Pattern: Excessive parallel construction
  "The study examined X, analyzed Y, and evaluated Z."
  "This approach enhances accuracy, improves efficiency, and reduces cost."

Human Fix: Break parallelism occasionally
  "The study examined X. For Y, a different analytical lens was required,
   so we turned to Z for comparison."

Revision Strategies

Sentence-Level Humanization

def humanize_sentence_variety(sentences: list[str]) -> dict:
    """
    Analyze sentence variety -- AI text often has uniform sentence lengths
    and structures.
    """
    lengths = [len(s.split()) for s in sentences]
    avg_length = sum(lengths) / len(lengths)
    std_length = (sum((l - avg_length)**2 for l in lengths) / len(lengths)) ** 0.5

    # Check first word variety
    first_words = [s.split()[0].lower() if s.split() else '' for s in sentences]
    unique_first_words = len(set(first_words)) / len(first_words)

    issues = []

    if std_length < 3:
        issues.append(
            f"Sentence lengths are too uniform (avg={avg_length:.0f}, "
            f"std={std_length:.1f}). Mix short (5-10 words) and long "
            f"(20-30 words) sentences."
        )

    if unique_first_words < 0.5:
        repeated = [w for w in set(first_words) if first_words.count(w) > 2]
        issues.append(
            f"Too many sentences start with the same word: {repeated}. "
            f"Vary sentence openings."
        )

    # Check for consecutive similar-length sentences
    uniform_runs = 0
    for i in range(1, len(lengths)):
        if abs(lengths[i] - lengths[i-1]) < 3:
            uniform_runs += 1

    if uniform_runs > len(lengths) * 0.6:
        issues.append("Too many consecutive sentences with similar lengths.")

    return {
        'avg_sentence_length': round(avg_length, 1),
        'length_std': round(std_length, 1),
        'first_word_variety': round(unique_first_words, 2),
        'issues': issues,
        'assessment': 'natural' if not issues else 'needs_revision'
    }

Voice and Perspective

AI text often defaults to an impersonal, overly balanced voice. Academic writing benefits from:

1. Authorial voice: Use "we" in multi-author papers. Take clear positions.
2. Disciplinary conventions: Match the register of your target journal (some are more formal, others more conversational).
3. Specific over general: Replace "many researchers have studied X" with "Smith (2020), Jones (2021), and Lee (2023) each approached X differently."
4. Genuine hedging: Use hedging when genuinely uncertain, not as a default.

Workflow for AI-Assisted Writing

Step 1: Draft with AI assistance (outline, first draft)
Step 2: Print the draft and read aloud -- mark anything that sounds generic
Step 3: Replace flagged phrases with your natural voice
Step 4: Add personal scholarly judgment (interpretations, critiques)
Step 5: Insert discipline-specific terminology and citations
Step 6: Vary sentence structure and paragraph length
Step 7: Run the pattern detector to catch remaining AI fingerprints
Step 8: Final read-aloud check

Ethical Considerations

Using AI for writing assistance is increasingly accepted in academia, but transparency is essential. Many journals now require disclosure of AI tool usage. The key ethical principle: you must deeply understand and stand behind every claim in the final text. AI is a drafting tool; scholarly judgment and intellectual ownership remain yours.