How to Detect AI Writing (And Beat Detection)

AI detection has become an industry. Schools use it. Clients ask about it. Everyone worries about it.

Here’s the truth about what detection tools can and can’t do.

How AI Detection Actually Works

The Pattern Matching

AI detectors look for:

Statistical patterns:

• Word frequency distributions
• Sentence length uniformity
• Predictable word choices
• Low “perplexity” (too smooth/predictable)

Stylistic markers:

• Perfect grammar throughout
• Lack of personal voice
• Formulaic structures
• Missing colloquialisms

The Problem

Human writing is diverse. Some humans write “AI-like”:

• Technical writers
• Non-native speakers
• Formal academic writers
• Anyone editing heavily

Detection tools can’t distinguish “writes like AI” from “is AI.”

Testing the Detectors

I ran controlled tests across major detection tools.

The Test

AI content: ChatGPT and Claude outputs Human content: My actual writing, plus samples from published authors Mixed content: AI drafts I heavily edited

Results

Translation: Every detector flagged some human writing as AI. Every detector missed some AI writing.

The Worst Part

Non-native English speakers got flagged at much higher rates. Their human writing triggered “AI” flags constantly.

Signs of AI Writing (Real Ones)

If you’re trying to spot AI, look for:

1. The Opener Patterns

AI loves certain openings:

• “In today’s fast-paced world…”
• “When it comes to…”
• “It’s important to note that…”
• “[Topic] has become increasingly…”

2. The Perfect Balance

AI text often has:

• Extremely consistent paragraph lengths
• Suspiciously balanced pro/con lists
• No tangents or personality
• Relentless structure

3. The Vocabulary Tells

Overused AI words:

• “Delve”
• “Leverage”
• “Facilitate”
• “Comprehensive”
• “Streamline”
• “Multifaceted”

4. Missing Human Elements

AI rarely includes:

• Specific personal stories
• Precise numbers/dates
• Unique opinions
• Self-deprecation
• Genuine uncertainty

5. The “Too Helpful” Problem

AI tries to be complete and balanced. Humans have opinions and biases they don’t always mask.

Making AI Content Undetectable

Not for cheating - for creating genuinely good content with AI assistance.

1. Start Different

Don’t accept the first output. AI patterns are strongest in default responses.

Better approach:

Write about [topic].

Important: Don't start with common AI openings. Begin with a specific example,
question, or surprising statement. Avoid words like "delve," "comprehensive,"
and "facilitate."

2. Add Your Actual Experiences

AI can’t make up your experiences. Add:

• Real projects you’ve done
• Specific outcomes you’ve seen
• Mistakes you’ve made
• Opinions you actually hold

This is why edited AI content passes detection - human elements change the patterns.

3. Vary Structure

AI output tends toward predictable structures. Mix it up:

• Some short paragraphs, some long
• Questions without answers
• Tangential asides
• Incomplete thoughts occasionally

4. Remove AI Vocabulary

Search and replace:

• “Delve” → “look at” or specific verb
• “Comprehensive” → cut or be specific
• “Leverage” → “use”
• “Facilitate” → “help” or “enable”
• “It’s important to note” → delete entirely

5. Read Aloud

AI text often sounds unnatural spoken. Read your content aloud. If you’d never say it that way, rewrite it.

6. Edit Significantly

Light editing keeps AI patterns. Heavy editing creates human patterns.

Target: Change 40-60% of words/structure. At that point, it’s genuinely your writing with AI assistance.

The Ethical Framework

Where AI Writing Is Fine

• Blog posts (disclosed or not)
• Marketing copy
• Internal documents
• First drafts you’ll edit
• Brainstorming and outlines

Where It’s Problematic

• Academic work (where prohibited)
• Bylined journalism
• Personal communications pretending to be manual
• Any context where authenticity is expected and AI use isn’t disclosed

The Disclosure Question

You don’t need to disclose “I used AI” on every blog post any more than “I used spellcheck.”

You probably should disclose AI use when:

• Authenticity is the point (memoirs, personal essays)
• Professional standards require it
• Readers would feel deceived if they knew
• It’s legally required

For Those Checking Others’ Work

Teachers/Professors

Detection tools are unreliable. Better approaches:

• In-class writing samples for comparison
• Process documentation (drafts, notes)
• Oral explanation of written work
• Assignments that require personal experience

Editors/Employers

Focus on:

• Does the content serve its purpose?
• Is the quality acceptable?
• Do claims check out?
• Does voice match other work from this person?

AI detection as gatekeeping creates more problems than it solves.

The Future

Detection and generation are in an arms race. Detection will improve. Generation will improve to beat detection. Neither side wins permanently.

Better approach: Accept that AI-assisted writing is normal. Focus on output quality and appropriate use rather than detection theater.

Practical Takeaways

If you use AI for writing:

• Edit substantially (40%+ changes)
• Add genuine personal content
• Vary structure and vocabulary
• Don’t worry excessively about detection

If you’re checking for AI:

• Don’t trust any single detector
• Look for quality and accuracy instead
• Consider false positive harm
• Focus on the work serving its purpose

For everyone:

• AI writing assistance is a tool
• Quality and honesty matter more than detection
• The conversation should be about appropriate use, not gatekeeping