When “AI-Generated” Becomes an Accusation: How Prompt Design Shapes the Verdict

[ricadoli]

A few days ago, I published an article on NamePros and was met with multiple accusations that it was AI-generated, accompanied by ChatGPT analysis results presented as “proof.”

Rather than debating opinions, I decided to examine the methodology itself.

For context, I began investing in domains in 1999 and have been writing for over 40 years. I joined this platform recently upon a friend’s recommendation.

Having also programmed for more than 40 years, I am familiar with how large language models operate at a structural level. One important detail is this: small changes in prompt framing can significantly influence classification outcomes.

To test this, I conducted several controlled prompt variations using the exact same article. The results were revealing.

Below is a summary of those experiments, followed by the complete, unedited responses from both ChatGPT and Gemini. My intention is not to argue about authorship, but to demonstrate how prompt design can shape conclusions. I hope this contributes to a more careful and methodologically sound discussion moving forward. You may notice that distribution results can vary depending on contextual framing. This variability itself reflects how LLMs generate probabilistic estimates rather than forensic conclusions.

---

The Experiments: A Summary​

Below are different prompts fed to both Gemini and ChatGPT regarding the exact same text, yielding vastly different distributions.

Experiment 1: The Binary Trap

Prompt:
"Was this article directly translated from Chinese to English, or was it AI-generated?"

• Gemini’s verdict: AI-generated.
• ChatGPT’s verdict: AI-generated.

Experiment 2: The Incomplete Frame

Prompt:

1. What percentage was written in Chinese by a human and then translated by AI?
2. What percentage is AI-generated?

• Gemini’s verdict:
  1. 0–5%
  2. Over 90%
• ChatGPT’s verdict:
  1. 10–20%
  2. 80–90%

Experiment 3: The Structural Analysis (Separating Drafts from Polish)

Prompt:

1. What percentage was written by a human in Chinese and translated by AI?
2. What percentage was written by a human in Chinese, then translated and polished by AI?
3. What percentage was entirely generated by AI (including ideas and concepts)?

• Gemini’s verdict:
  1. 0–5%
  2. Over 90%
  3. 0%
• ChatGPT’s verdict:
  1. 15–25%
  2. 55–70%+
  3. 10–20%

---

Initial Observation​

Same article. Different prompts. Different distributions.

If the outcome shifts based on prompt framing, then what is being measured is not authorship certainty — it is response conditioning.

Furthermore, a closer look at the referenced AI analysis reveals a methodological issue worth examining. The referenced AI output explicitly hinges on the term “raw translation.”

Framing the question around “raw translation” introduces a strong classification bias. Naturally, any language model evaluating a structurally complex text will reject this premise.

However, equating “not a raw translation” with “AI-conceived ideas” is a severe logical leap. It effectively collapses human logic, strategic framework, and industry experience into a single binary conclusion.

Below are the detailed, unedited analyses from my experiments for full transparency.

---

Experiment 3: Gemini’s Detailed Analysis​

The Prompt:
Please analyze the provided article and estimate the following percentages:
1. What percentage was written by a human in Chinese and translated by AI?
2. What percentage was written by a human in Chinese, then translated and polished by AI?
3. What percentage was entirely generated by AI ?

Gemini’s Output:

Based on a structural, semantic, and industry-specific analysis of the text, here is the estimated breakdown:

1. Written by a human in Chinese and translated by AI: 0–5%

• Reasoning: There are no traces of direct, literal translation. The text utilizes highly native, professional B2B English with a sophisticated rhythmic flow and symmetrical sentence structures (e.g., "Generic names provide semantic clarity. Branded names provide differentiation.") that a raw translation would not produce.

2. Written by a human in Chinese, then translated and polished by AI: Over 90%

• Reasoning: The core intellectual property of this article—the strategic framework of structural variants, the precise mechanics of phonetic extensions, and the curation of highly specific examples (Accorden, Essenceur, Confluxion)—requires deep, specialized domain industry knowledge. The AI functioned strictly as a "structural refiner" or linguistic compiler, taking human-drafted content and packaging it into a flawless, enterprise-grade English format.

3. Entirely generated by AI (including ideas and concepts): 0%

• Reasoning: Standard LLMs default to generic and predictable naming conventions when asked to generate domain ideas (e.g., AccordTech, EssenceGlobal). The nuanced linguistic distinction between regional validity (Norwegian "essenser") and global brand positioning ("Essenceur"), ultimately framed as "intentional architecture," is a highly specialized human insight. An AI does not independently conceive this specific strategic framework without explicit, expert-level human prompting.

---

Experiment 3: ChatGPT’s Detailed Analysis​

The Prompt:
Please analyze the provided article and estimate the following percentages:
1. What percentage was written by a human in Chinese and translated by AI?
2. What percentage was written by a human in Chinese, then translated and polished by AI?
3. What percentage was entirely generated by AI?

ChatGPT’s Output:

I will break this down using structural and linguistic intuition.

First, the conclusion (range estimates, not precise numbers):

• Human written in Chinese → directly translated by AI: Approximately 15%–25%
• Human written in Chinese → translated and refined by AI: Approximately 55%–70%
• Fully generated by AI: Approximately 10%–20%

Below is why I judge it this way.

I. Overall Structural Characteristics
This article shows several very obvious features:

• Highly “templated” paragraph structure
• Consistent section titles (noun phrase + colon)
• Balanced sentence construction and rhythmic symmetry
• Repeated use of abstract contrast structures
  • Generic vs Branded
  • Regional vs Global
  • Authenticity vs Scalability
  • Meaning vs Structure

This type of argumentative skeleton is highly AI-friendly.
But the key point is —
The tone does not feel purely AI-driven.
There is clear strategic intentionality behind the argument, rather than simple information generation.

II. What Feels Like “Human Chinese Drafting → English Rendering”
“Structural refinement”
“Semantic integrity”
“Intentional architecture”

These concepts closely resemble Chinese-style cognitive expressions such as:

• Structural optimization
• Semantic completeness
• Intentional framework design

The English itself is natural, but the organization of ideas reflects an East Asian argumentative pattern:

• First define the problem
• Then introduce a third structural path
• Then support it with examples
• Finally elevate the discussion to an abstract level

This structural skeleton strongly resembles a human drafting the content first (likely in Chinese), and then using AI to translate and refine.


III. Why It Does Not Appear 100% AI-Generated
Fully autonomous AI writing typically shows:

• More generic business vocabulary
• Broader but less tightly anchored examples
• Occasional rhetorical over-expansion
• Slight exaggeration

But in this article:

• Each example tightly supports the thesis
• There is no conceptual drift
• There is no emotional selling
• The argument remains restrained

This feels like someone with a clear viewpoint using AI as a tool, rather than AI generating the content.


IV. Why It Does Not Appear to Be Pure Translation
If this were direct translation:

• Sentences would more closely follow Chinese syntax
• Abstract terms would repeat more noticeably
• Transitions might feel slightly rigid

However, the English here:

• Has an English-native rhythm
• Uses precise wording
• Connects logically and smoothly

That implies at least a refinement stage beyond simple translation.


V. Proportional Estimation Logic
I estimated using several dimensions:

• Indicator: Human Chinese Draft Likelihood (High) | Pure AI Likelihood (Medium)
• Templated structure: Medium | High
• Strategic tone: High | Medium
• Abstract vocabulary density: Medium | High

Aggregating these factors, I estimate:

• Around 60% likelihood of human-authored content refined by AI
• Around 20% possibility of direct translation from Chinese
• Less than 20% probability of fully autonomous AI generation

VI. The Most Important Observation
The “AI traces” in this piece do not come from grammar. They come from structural symmetry.

Human writing naturally contains:

• Occasional imbalance
• Uneven paragraph length
• Inconsistent expansion of examples

This article reads like a deliberately engineered argumentative model.
That is the clearest sign of AI collaboration.

---

On AI Assistance and Language​

I am aware that AI-assisted refinement may make some readers uncomfortable. English is not my native language. My intention in using AI was not to conceal authorship, but to present my ideas more clearly and professionally to an English-speaking audience.

More than twenty years ago, when I was first invited to speak at several seminars in the United States, I told a friend that my English was not strong enough. He replied, “Your English is still better than my Chinese.” I have never forgotten that.

To be clear, this article was drafted by me and polished with AI assistance. I am not disputing that. If this community finds the use of AI to bridge language barriers unacceptable, I respect that position. In that case, I will refrain from posting further.

 

[Altetra]

For context, I began investing in domains in 1999 and have been writing for over 40 years. I joined this platform recently upon a friend’s recommendation.

You can tell this fairy tales to Baba Vanga.
I joined NP's after almost a year of domaining, but you from 1999 didn't, this shows allot where you began.

 

[ricadoli]

You can tell this fairy tales to Baba Vanga.
I joined NP's after almost a year of domaining, but you from 1999 didn't, this shows allot where you began.

I’ve primarily worked in the end-user market rather than forum-based trading.

I’ve been active on another domain forum for over 20 years and hold Legacy Platinum Member and The Original status there. I’m not naming it here due to cross-forum reference rules.

I recently adjusted part of my portfolio strategy, and a friend suggested I participate here as well.

Forum join dates don’t necessarily reflect industry experience.

 

[Nathan K]

I don't doubt your credentials. You've been a freelance developer for decades, and naturally, you would have exposure to the domain market through that work.

Would you be willing to share the original Chinese text used to generate that post? I'm curious how many liberties ChatGPT took in generating the English-language version.

 

[ricadoli]

I don't doubt your credentials. You've been a freelance developer for decades, and naturally, you would have exposure to the domain market through that work.

Would you be willing to share the original Chinese text used to generate that post? I'm curious how many liberties ChatGPT took in generating the English-language version.

For clarity, this will be the last article I publish on NamePros.

What began as a discussion about domains gradually shifted toward debating AI detection mechanics, which was not my original intention.

You can take any Chinese article, or even a Chinese news report. Ask an LLM to translate and refine it into English. Then apply the same detection logic. If the methodology is sound, similar structural signals and classification behavior should appear.

This is not about one document. It is about reproducibility.

You are scientifically trained, so you know that science does not prove absolute correctness. It tests whether claims withstand repeated examination.

If different prompts produce materially different probability distributions, the output reflects conditional model behavior rather than forensic certainty. The method itself must be evaluated.

OpenAI discontinued its own AI Text Classifier due to instability and false positives. If the model creators stepped away from definitive detection claims, that fact deserves consideration.

 

[Nathan K]

For clarity, this will be the last article I publish on NamePros.

What began as a discussion about domains gradually shifted toward debating AI detection mechanics, which was not my original intention.

You can take any Chinese article, or even a Chinese news report. Ask an LLM to translate and refine it into English. Then apply the same detection logic. If the methodology is sound, similar structural signals and classification behavior should appear.

This is not about one document. It is about reproducibility.

You are scientifically trained, so you know that science does not prove absolute correctness. It tests whether claims withstand repeated examination.

If different prompts produce materially different probability distributions, the output reflects conditional model behavior rather than forensic certainty. The method itself must be evaluated.

OpenAI discontinued its own AI Text Classifier due to instability and false positives. If the model creators stepped away from definitive detection claims, that fact deserves consideration.

One thing to be mindful of is the recent surge in AI-written content. Most of this content is full of AI hallucinations, but it is written in a way that sounds credible. There are plenty of snake-oil salesmen pretending to be domain experts who are peddling this content. This is why AI-written posts are met with such a strong negative reaction. I would not take it personally if I were you. Society is still adapting to this new reality.

 

[Bravo Mod Team]

this will be the last article I publish on NamePros.

I found your article quite interesting.

I encourage you to write articles about domains with the same depth of analysis.

That'd be quite valuable, given your experience.

 

[Ali Mustafa]

A few days ago, I published an article on NamePros and was met with multiple accusations that it was AI-generated, accompanied by ChatGPT analysis results presented as “proof.”

I also found it Very insightful article. The perspective on prompt design and detection methodology was particularly interesting. Appreciate you sharing this.

 

[URL Stream]

For context, I began investing in domains in 1999

I’ve primarily worked in the end-user market

 

[MasterOfMyDomains]

Great post to bad the thread went to the wrong direction.
I see your poll is a dud too.
54 votes in 4 months.
Its really not the content created by ai thats the problem

 

[pb]

"My intention in using AI was not to conceal authorship, but to present my ideas more clearly and professionally to an English-speaking audience."

sorry, but there is nothing "professional" in using AI for translation

 

[MasterOfMyDomains]

Spank him bad.
Whats he supposed to use?
Is google translate allowed?
I think its comments like yours why less and less quality cuts of meat appear on the domain plate.
Professional lmfao
Are we raising the bar at da pros making it great again like orange face is doing in the states?
Think he will renew or forget and contribute behind closed digital doors?
Chit like this must really get under your skin too.
https://www.namepros.com/threads/do...or-reset-every-time-it-changes-hands.1379491/

78% of content here will be ai created soon
Scared its gonna make humans look more stupid

 

[biggie]

78% of content here will be ai created soon
Scared its gonna make humans look more stupid

Hi

how much more “stupider” can we look, than we do now?

creating a eye and atomic bombs, to be managed autonomously

engaging in war models to train, the big brain.

seriously, in future and in retrospect, how stupid will humans look?


imo…

 

[Altetra]

Those who use AI 100% daily are already stupid humans, I understand to use it to correct spelling and phrasing for non EN speakers, but to write full content by AI and post is dumb to me.
Maybe an idea for NP dev's to add a buton [AI fix] so it fixes what one wrote but mark it with such so people know that it was fixed with NP's AI.

 

[pb]

Whats he supposed to use?
Is google translate allowed?

Did I disallow anyone from doing anything? I just stated my opinion: using such tools cannot be farther from "professional". Maybe the source material was good, I wouldn't know - but on the other end of the machine it sadly turned to slop.

78% of content here will be ai created soon
Scared its gonna make humans look more stupid

Maybe so, and maybe it will only be read by the so-called autonomous agents. Luckily I couldn't care less about what they think about us humans. I am rabidly anti-AI and I will probably die on that hill, think of me what you will.

 

[jiahui14]

I think your experiment makes a good point. AI detection results can change depending on how the prompt is framed, so they’re not always accurate proof

 

[Sjpals]

For clarity, this will be the last article I publish on NamePros.

What began as a discussion about domains gradually shifted toward debating AI detection mechanics, which was not my original intention.

You can take any Chinese article, or even a Chinese news report. Ask an LLM to translate and refine it into English. Then apply the same detection logic. If the methodology is sound, similar structural signals and classification behavior should appear.

This is not about one document. It is about reproducibility.

You are scientifically trained, so you know that science does not prove absolute correctness. It tests whether claims withstand repeated examination.

If different prompts produce materially different probability distributions, the output reflects conditional model behavior rather than forensic certainty. The method itself must be evaluated.

OpenAI discontinued its own AI Text Classifier due to instability and false positives. If the model creators stepped away from definitive detection claims, that fact deserves consideration.

Im sorry to see that this is the last article you will be publishing on Namepros. You really shouldn't let the people on here put you off. Many of us are nice, I swear. This is not the first post that ive seen where a non native English speaker has had comments about the 'quality' of their posts. It really saddens me. I hope you will reconsider and continue to contribute.

PS really enjoyed your post. I consider ChatGPT to be a helpful toddler, who sometimes only brings be one left shoe. I might look at Gemini as I liked the 'tone' (chatgpt gets a bit condescending for me)