A recent discussion with my son prompted me to experiment (again) by putting Zero GPT Detector to the test to see if it could spot AI text.

By “putting Zero GPT Detector to the test”, I mean putting ChatGPT’s own “AI content detection assistant” to the test, which was prompted by my son Carter – a senior in high school – telling me that ChatGPT’s own AI detector was very accurate and that it will spot any AI generated text – which I’m happy to let him continue to believe. 😉

I conducted a test back in June of last year on ZeroGPT, (not to be confused with ChatGPT’s Zero GPT Detector) which still touts itself as “the most Advanced and Reliable Chat GPT, GPT4 & AI Content Detector” and wrote about it here, proving that it’s not that accurate.

I decided to use the same three examples from that test in putting Zero GPT Detector to the test. Here are the results. I’ll explain why they’re out of order in a bit. My analysis is in italics below.

Test #2: Content written by NotebookLM from Google

For this test, I chose the Briefing Document portion of this blog post about DeepSeek that I wrote back in January. This was a post where I loaded 15 articles into NotebookLM and had it generate a briefing document. The text I loaded for this one was the paragraph starting with “DeepSeek: AI’s Sputnik Moment” (but not the info in parenthesis in the first paragraph) down through the “Conclusion” paragraph. As I noted in the post, this text is all AI-generated. This is what Zero GPT Detector said:

AI-Generated Likelihood: 88%

Summary

The document shows a high likelihood of AI-generated or AI-assisted content.

---

Key Indicators

1. Structured “Briefing Document” Format
The text follows a rigid, almost templated structure:

• Executive summary
• Thematic sections with headers
• Bullet-pointed evidence and quotes
• Formal conclusion

This format is common in AI outputs designed for summarization and synthesis tasks.

---

2. Dense Information Compression
The writing efficiently packs:

• Technical details
• Quotes
• Analysis
• Implications

into tight bullet points. This level of high-density summarization with minimal redundancy is a strong AI signal.

---

3. Uniform Tone Across Sections
Despite covering multiple domains (technical, geopolitical, economic), the tone remains:

• Consistent
• Neutral
• Evenly formal

Human-authored reports often show subtle tonal shifts depending on the topic or emphasis.

---

4. Repetitive Analytical Framing
Patterns such as:

• “Implications:”
• “This development could…”
• “This demonstrates…”

appear frequently, suggesting template-driven reasoning structures.

---

5. Balanced, Multi-Angle Coverage
The document systematically presents:

• Benefits
• Risks
• Market reactions
• Geopolitical context

This comprehensive, evenly weighted coverage is typical of AI-generated synthesis rather than opinionated human writing.

---

6. Source-Like Quotation Integration
Quotes are included in a consistent format and style, but without deeper narrative integration—another indicator of automated aggregation or summarization.

---

Notable Human-Like Elements

• The content is coherent and logically organized for a professional audience
• Topic selection and framing could plausibly come from a human analyst

---

Bottom Line

The combination of:

• Highly structured formatting
• Uniform tone
• Dense, balanced synthesis

indicates the document is very likely AI-assisted or AI-generated, though possibly refined or guided by a human.

Well, that’s pretty much spot on! Maybe Carter is right – maybe this AI detector is a LOT better, after all, it’s a ChatGPT-related product! Or not.

Test #1: Content written by me.

Now to test #1. For this test, I chose this blog post I wrote close to three years ago comparing the rise and fall of the BlackBerry to the emergence of LLMs and GenAI in eDiscovery. I loaded in the plain text of the title and almost the entire blog post – down to (but not including) the paragraph beginning with “P.S.”, which seemed like a large clue that this was written by a person. Again, this was totally written by me – no AI*. What did Zero GPT Detector say about this test?

AI-Generated Likelihood: 92%

Summary

The text shows a high probability of AI involvement, based on its structured argumentation, consistent tone, and polished narrative flow.

---

Key Indicators

1. Highly Structured Narrative Arc
The piece follows a very deliberate progression:

• Historical setup (BlackBerry rise)
• Turning point (iPhone disruption)
• Failure case (BlackBerry Storm)
• Modern parallel (genAI in eDiscovery)
• Prescriptive takeaway (two-part lesson)

This kind of clean, almost formulaic structure is common in AI-generated long-form content.

---

2. Consistent Tone and Linguistic Uniformity
The writing maintains:

• Even pacing
• Balanced sentence length
• Minimal stylistic deviation

Human writing often includes more variability (e.g., abrupt phrasing, uneven emphasis), while this remains uniform throughout.

---

3. High Clarity with Low Friction
The text is:

• Exceptionally clear
• Free of ambiguity or digressions
• Lacks personal anecdotes or imperfect transitions

This “frictionless clarity” is a hallmark of AI-assisted writing.

---

4. Repetitive Framing Patterns
Certain rhetorical patterns repeat:

• “Which brings me to…”
• “In my opinion…”
• “The other part of the lesson…”

These are template-like transitions frequently seen in AI outputs.

---

5. Balanced, Risk-Aware Conclusion
The ending delivers a measured, dual-sided recommendation:

• Don’t ignore genAI
• Don’t rush implementation

This type of symmetrical reasoning is characteristic of AI-generated business analysis.

---

Bottom Line

While a skilled human could produce this, the combination of:

• Structural precision
• Stylistic consistency
• Predictable rhetorical patterns

suggests the text is very likely AI-generated or heavily AI-assisted.

Say what?!? “Clean, almost formulaic structure”? “Consistent Tone and Linguistic Uniformity”? “Frictionless clarity”? “Template-like transitions”? “Measured, dual-sided recommendation” that shows “symmetrical reasoning”?

Wow. I guess you can start calling me “the machine”. Or at least a “skilled human”. 😁 This is why I showed this test second even though I did it first – after all, I had to draw out the suspense!

Zero GPT Detector said it could “compare [the second] document directly against your BlackBerry piece to show which one is more likely AI-generated and why.” So, I had it do that and this is what it said:

AI-Generated Likelihood Comparison

• BlackBerry Piece: 92%
• DeepSeek Briefing Document: 88%

👉 Conclusion: The BlackBerry piece is more likely AI-generated than the DeepSeek document.

---

Why the BlackBerry Piece Ranks Higher

1. Narrative Smoothness & Storytelling Precision

The BlackBerry text reads like a polished business essay with a cinematic arc:

• Setup → disruption → failure → lesson → application
• Transitions are seamless and almost too clean

This kind of highly optimized storytelling flow is a strong AI indicator.

---

2. Rhetorical Framing Patterns

It repeatedly uses structured persuasion:

• “Which brings me to…”
• “In my opinion…”
• “The other part of the lesson…”

These are classic AI scaffolding phrases, used to organize long-form reasoning.

---

3. Balanced, “Consultant-Style” Insight

The conclusion lands on a perfectly balanced takeaway:

• Don’t ignore genAI
• Don’t rush genAI

This symmetrical, risk-aware framing is extremely typical of AI-generated business analysis.

---

Why the DeepSeek Document Scores Slightly Lower

1. Format Explains the Structure

The DeepSeek text is explicitly a:

“Briefing Document”

That format naturally justifies:

• Bullet points
• Section headers
• Neutral tone

So its structure is less suspicious, even though it’s still highly systematic.

---

2. More “Compiled” Than “Composed”

The document feels like:

• Aggregated research
• Summarized inputs
• Compiled notes

Rather than a single flowing narrative. This reduces the “AI voice” signal slightly.

---

3. Slight Variability in Content Density

Some sections are:

• Highly technical (e.g., FP8, DualPipe)
• Others more general (market reactions, geopolitics)

This introduces minor natural inconsistency, which is more human-like.

---

Side-by-Side Signal Comparison

Signal	BlackBerry Piece	DeepSeek Document
Narrative flow	Very smooth, linear	Fragmented, sectional
Tone consistency	Extremely uniform	Uniform but format-driven
Structure type	Persuasive essay	Research briefing
Rhetorical patterns	Repetitive & templated	Minimal
Information style	Explanatory storytelling	Dense summarization
AI likelihood	Higher	High, but lower

---

Final Takeaway

• The BlackBerry piece shows stronger signs of AI-style composition and persuasion patterns.
• The DeepSeek document shows signs of AI-assisted summarization, but its format makes that less definitive.

👉 In simple terms:

• BlackBerry = AI “writing”
• DeepSeek = AI “compiling”

Wow. Putting Zero GPT Detector to the test is eye-opening. It couldn’t be more wrong about the comparison of the two pieces.

Test #3: Content written by ChatGPT

I almost hesitate to run this third test, but I wanted to see what happens when I’m putting Zero GPT Detector on something ChatGPT actually wrote. This was the validation protocol example in the In re Broiler Chicken Antitrust Litigation case from 2018 that I used in the June test (text of the ChatGPT output available here). So, this text is also all AI-generated. What happened here?

AI-Generated Likelihood: 35%

Summary

This document shows a low-to-moderate likelihood of AI involvement, and is much more likely human-authored, possibly with light editing or formatting assistance.

---

Key Indicators

1. Legal/Procedural Writing Style (Strong Human Signal)

The document reflects:

• Formal legal structure
• Precise procedural language
• Domain-specific terminology

This type of writing is typically:

• Produced by legal professionals
• Constrained by court standards and precedent

AI can mimic this, but the consistency with real-world legal conventions strongly suggests human authorship.

---

2. Irregular Sentence Construction

Unlike the previous two documents, this text includes:

• Long, sometimes dense sentences
• Occasional redundancy or awkward phrasing
• Less “polished” transitions

These are natural human imperfections that AI tends to smooth out.

---

3. Lack of Narrative or Persuasive Framing

There is:

• No storytelling arc
• No rhetorical transitions (“which brings me to…”)
• No balanced “insight conclusion”

Instead, the document is:

• Functional
• Instructional
• Procedural

This is not typical of AI-generated explanatory content.

---

4. Uneven Emphasis and Detail

Some sections:

• Go very deep into specifics
• Others remain brief or highly technical

This unevenness reflects human drafting priorities, not AI’s tendency toward balanced coverage.

---

5. Purpose-Driven Language

The writing is clearly designed to:

• Define processes
• Establish validation standards
• Support defensibility in litigation

AI-generated text usually optimizes for readability, while this prioritizes precision and compliance.

---

Comparison to Your Previous Documents

Document	AI Likelihood	Primary Style
BlackBerry piece	92%	Narrative / persuasive
DeepSeek briefing	88%	Structured synthesis
Legal PDF (this)	35%	Procedural / legal

---

Final Takeaway

• This document is significantly more human-authored than the other two.

If AI was used at all, it was likely limited to:

• Minor editing
• Formatting
• Possibly summarization of small sections

👉 In simple terms:

• BlackBerry = AI-written
• DeepSeek = AI-compiled
• This PDF = Human-written (with possible light AI assistance)

Wow (again). The model simply couldn’t be more wrong – and this is ChatGPT analyzing its own output. What else can I say?

Once again, this was, admittedly, not a scientific case of putting Zero GPT detector to the test, but it does hopefully reiterate how poor AI detectors are at accurately differentiating human generated text from AI generated text. Just don’t tell Carter. 🤣

So, what do you think? Have any of you been putting Zero GPT detector to the test? If so, what did you find? Please share any comments you might have or if you’d like to know more about a particular topic.

*I swear it was all me!

Disclaimer: The views represented herein are exclusively the views of the authors and speakers themselves, and do not necessarily represent the views held by my employer, my partners or my clients. eDiscovery Today is made available solely for educational purposes to provide general information about general eDiscovery principles and not to provide specific legal advice applicable to any particular circumstance. eDiscovery Today should not be used as a substitute for competent legal advice from a lawyer you have retained and who has agreed to represent you.