Asking ChatGPT how to recognize articles written by ChatGPT…

Yes, it’s a bit like asking a snake to bite its own tail. But don’t worry I won’t only copy past its answer but give you my own opinion about it.

As for why I am interested by this subject it is because I might use those kind of tools myself in the future to help me reduce the time to write new articles.

Photo by Solen Feyissa on Unsplash

What makes me think an article was written using ChatGPT

I’m not saying that I’m always right when I try to determine whether or not a text has been generated by a Large Language Model, like what’s behind ChatGPT, I’m just saying that some things are rather representative.

Lists, lists, always lists

Screenshot of the start of the answer to “how to recognize an article written by chatgpt?”

I think it is the first thing we notice while using ChatGPT, it LOVES making lists.

Well, it does admits it itself but only as it’s 8th point.

8. Formatting Patterns
AI often structures text into clear sections with headings, bullet points, or numbered lists, aiming for clarity and conciseness.

It often repeats itself!

Even we humans have a tendency to repeat ourselves, but we usually correct ourselves by rereading.

What it wrote about it:

7. Repetition
AI may repeat ideas, rephrase points, or circle back to earlier topics unnecessarily.pl

The repetition that was made during the answer…

1. Consistent and Polished Tone
...
Lack of emotional depth or personal anecdotes might suggest AI authorship.

6. Lack of Personal Insight
Absence of first-hand experiences, specific examples, or unique perspectives that a human writer might include.

As you can see in the first and sixth point it was noted that ChatGPT does not add personal elements.

To go further on this subject, we humans sometimes tend to leap from one subject to another, we have an expression for it in French “passer du coq à l’âne”. That’s something that a LLM can’t do as it construct a text by appending a token and starting again and again, the next token being determined by the current semantic context.

The LLM’s tendency to simply build on an idea by appending token by token means that it can tend to go round in circles and so go back to earlier topics.

It’s far too self-confident

A human can write in an article that it does not know about a subject, ChatGPT can’t and will often “think” of what it is saying as being the truth (hallucinations)

You can see it’s self-confidence in the fifth point of its answer.

5. Factually Correct but General
The article may include accurate information but rarely dives deeply into niche or highly specialized topics unless specifically prompted.
Overuse of definitions or background information for simple topics.

ChatGPT assume he is always factually correct, it does not mention hallucinations!

It is also “hard” for a LLM that there can be multiple point of views around a single topic.

4. High Predictability
Sentences may feel predictable or structured in a way that flows logically but lacks creativity.
Limited use of idiomatic expressions, slang, or region-specific humor.

The other points listed by ChatGPT itself

About the language level

1. Consistent and Polished Tone
The tone may feel neutral, overly formal, or consistently polished without the natural variation typical of human writing.

2. Overuse of Common Phrases
AI tends to use generic or repetitive phrases like "it is important to note," "in conclusion," or "on the other hand."

9. Language and Grammar
Grammatically flawless but occasionally slightly awkward sentence constructions, such as:
Overuse of conjunctions like "however" or "furthermore."
Odd transitions between ideas.

Last point the tools that can be used

10. Detection Tools
Use AI detection tools like OpenAI AI Text Classifier (https://platform.openai.com/ai-text-classifier) or others such as GPTZero, Copyleaks, or Turnitin, which are designed to identify AI-generated text.
However, as AI-generated writing becomes more sophisticated, distinguishing it from human writing will require deeper analysis or specialized detection tools.

To conclude

While I think ChatGPT and Large Language Models in general can be great tools to help write articles, I think that as for now they must remain what they are: tools, not writers.

You can use them to help find topics, help you correct your sentences, but you shouldn’t rely directly on their content.

That’s all folks!

Feel free to clap or follow me!