How to Detect AI-Written Content

I wanted to know if AI detection tools actually work. So I tested five of them.

Spoiler: they’re not as reliable as they claim.

The Experiment

I submitted three types of content to five AI detection tools:

Content A: Pure ChatGPT output, no editing Content B: ChatGPT output, heavily edited by me Content C: Something I wrote entirely myself, no AI

Then I checked what the tools said.

The Tools I Tested

1. GPTZero
2. Originality.ai
3. Copyleaks
4. ZeroGPT
5. Writer.com AI Detector

The Results (Messy)

Content A: Pure AI Output

This should be easy. Straight ChatGPT with no editing.

Results:

• GPTZero: 95% AI (correct)
• Originality.ai: 100% AI (correct)
• Copyleaks: 89% AI (correct)
• ZeroGPT: 78% AI (correct)
• Writer.com: 92% AI (correct)

All five caught it. Good start.

Content B: AI + Heavy Human Editing

ChatGPT draft, but I rewrote about 40%, added personal examples, changed structure.

Results:

• GPTZero: 62% AI (uncertain)
• Originality.ai: 45% AI (uncertain)
• Copyleaks: 71% AI (thinks AI)
• ZeroGPT: 33% AI (thinks human)
• Writer.com: 58% AI (uncertain)

All over the map. Same content, different verdicts.

Content C: Written Entirely by Me

No AI involvement. Just me typing.

Results:

• GPTZero: 12% AI (correct, thinks human)
• Originality.ai: 8% AI (correct)
• Copyleaks: 23% AI (correct)
• ZeroGPT: 6% AI (correct)
• Writer.com: 31% AI (mostly correct)

Mostly correct, but notice the variation. Writer.com thought 31% of my human writing looked AI-generated.

The Problem: False Positives

I ran another test with famous writing - published authors, classic prose. Some tools flagged portions as AI-generated.

Why? Because AI was trained on good writing. Good writing can look like AI because AI learned from good writing.

This is a fundamental problem these tools can’t solve.

What Actually Gives AI Away

Forget the tools. Here’s what I’ve noticed after reading thousands of pieces of AI content:

1. Generic openings

AI loves:

• “In today’s fast-paced world…”
• “In the ever-evolving landscape of…”
• “When it comes to…”
• “Let’s dive into…”

Real writers usually just start.

2. No actual opinions

AI hedges. Everything is “can be,” “may help,” “often provides.” Real writers take positions. “This tool sucks.” “Don’t buy this.” “I hate how this works.”

AI doesn’t have opinions because it doesn’t have experiences.

3. Suspiciously perfect structure

AI loves:

• Intro paragraph
• Exactly 3-5 main points
• Each point the same length
• Tidy conclusion

Real writing is messier. Some sections longer than others. Tangents. Points that flow naturally rather than being mechanically structured.

4. No specific examples

AI gives generic examples: “For example, businesses can use this to improve efficiency.”

Humans give specific examples: “Last month I used this for a client project and saved 3 hours.”

5. Perfect grammar, no voice

AI doesn’t make typos (usually). It doesn’t have verbal tics. It doesn’t use expressions unique to a person.

Real writing has personality. Fragments. Casual asides. Ways of phrasing things that are distinctly yours.

6. List parallelism

AI makes every list item the same structure:

• “Improving your workflow”
• “Enhancing your productivity”
• “Maximizing your efficiency”

Humans vary it naturally:

• “Better workflows”
• “Actually get stuff done”
• “Stop wasting time on BS”

Why Detection Tools Struggle

Problem 1: AI keeps improving

Every GPT update makes output more human-like. Detection tools play catch-up and always lag behind.

Problem 2: The training data paradox

AI learned from human writing. Good human writing and good AI writing can be indistinguishable because AI learned what “good” looks like from humans.

Problem 3: Editing defeats detection

Light editing makes AI content harder to detect. Heavy editing makes it nearly impossible. The tools can’t distinguish “AI wrote this, human edited” from “human wrote this.”

Problem 4: False positives destroy trust

When tools flag genuine human writing as AI, people stop trusting them. I’ve seen writers falsely accused of using AI based on these tools.

My Approach

For evaluating content professionally:

1. Read it critically - does it have specific examples, genuine opinions, personality?
2. Check multiple detection tools (if needed)
3. Use tools as one data point, not gospel
4. Trust your reading over algorithmic detection

For making your AI content undetectable:

1. Edit heavily - change structure, add specifics
2. Add personal experiences and examples
3. Include opinions and takes
4. Make it messier - vary paragraph lengths, be less formal
5. Write like you talk

The Honest Truth

AI detection is unreliable. The tools help sometimes but fail often.

The better approach: write content that’s genuinely valuable regardless of how it was created. Add your experience. Have opinions. Include specifics.

If you’re editing AI output into something with genuine value and personality, detection becomes irrelevant. You’ve created something worth reading.

If you’re publishing raw AI output with no editing, it probably deserves to be flagged - not because tools can detect it, but because it’s probably not that good.

Bottom Line

Don’t trust AI detection tools blindly. They’re 60-80% accurate on their best days and generate false positives regularly.

The best detector is still human judgment: Does this writing have personality, specifics, and genuine insight? Or does it read like a Wikipedia article about the topic?

That question matters more than any algorithm.