The last few months have been marked by the arrival of ChatGPT with its GPT3 language model, the conversational artificial intelligence developed by OpenAI.

The availability of ChatGPT to the general public has created a great deal of excitement in many areas, including teaching, training, research, writing, etc.

ChatGPT can generate texts in several languages: French, English, Italian, Spanish, Arabic... A question must be asked to the AI tool to generate these answers.

So how do you know if a content is authentic, written by a human?

To indicate the nature of an author, several detectors have been developed: they can identify whether a text is generated by an AI such as ChatGPT (GPT-3), YouChat... or if it is written by a human.

In this article, we identify and compare the best AI detectors for you: 

At the end of the article, you will also find the visual summary of the comparative analysis.

best AI detectors comparison - ChatGPT (GPT-3)
authenticity originality contents

When should you ensure the authenticity of the content?

In this context of universal accessibility to AI tools such as ChatGPT (GPT-3) or YouChat; authenticity and originality of content is an issue.

For example, in the fields of education, it is important to ensure that a student's work (homework, essays, dissertations and theses) is original, with properly cited references.

If the teacher cannot ensure the originality of the students' work, he or she cannot ensure that the students have learned their lessons well.

From the student's point of view, the search for ease in writing can sometimes lead to fraudulent practices: plagiarism, ghostwriting by agencies, ghostwriting by an AI...

Behind this ease, the student misses out on the :

  • important knowledge to be acquired according to their specialisation
  • crucial skills to be developed in preparation for professional life: searching for the right sources of information, synthesising, schematising, personal reflection and writing a report.


AI tools can generate texts in abundance. ChatGPT (GPT3), for example, finds the most likely words to continue an initiated conversation. In most cases, the most likely continuation of a text is the correct answer to the question asked. However, this is not always the case and can contribute to misinformation.

 

Top 11 AI detectors in 2023 compared

 

Focus on our comparison methodology

Early 2023, we tested several AI and ChatGPT detectors with the aim of assessing the reliability of each tool. 

We analysed 125 English texts, ranging from 250 to 2500 characters, in the 11 AI detectors. The texts analysed contain human-written texts, texts generated by ChatGPT and texts generated by YouChat.

The final result of our tests is a reliability rate of analysis for each tool, based on the same contents.

Artificial Intelligence Detector Compilatio

The Compilatio AI Detector (POC*)

 

Presentation of the POC Compilatio AI Detector

  • French company specialised in plagiarism prevention and detection
  • ChatGPT detector based on GPT2 and GPT3, Youchat.
  • Multilingual detector: French, English, Italian, Spanish, Portuguese, German... 
  • Results displayed as % of reliability and emoji: šŸ§‘ or šŸ¤–.

Results of our tests on the POC* Compilatio AI Detector

  • 125 English texts analysed.
  • 46/50 human-written texts recognised, i.e. 92% reliability on human text recognition.
  • 66/75 AI-generated texts recognised, i.e. 88% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 89.60%.
Draft & Goal AI Content Detector

Draft & Goal AI Detector

 

Presentation of Draft & Goal AI Detector

  • Canadian company specialised in the development of marketing tools.
  • ChatGPT detector, based on GPT2 and GPT3.
  • Bilingual detector: French and English.
  • Results displayed as % of AI content in text.

Results of our tests on Draft & Goal AI Detector

  • 118 English texts analysed.
  • 43/50 human-written texts recognised, i.e. 86% reliability on human text recognition.
  • 59/68 AI-generated texts recognised, i.e. 86.76% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 86.44%.
Giant Language Model Test Room - GLTR Test-2

GLTR Test-2 (Giant Language model Test Room)

 

Presentation of GLTR Test-2

  • AI detector developed by researchers at MIT-IBM Watson AI lab and Harvard NLP.
  • ChatGPT detector, based on GPT2.
  • No information provided on the languages processed.
  • Results are displayed as different colours for each word according to the probability of the AI predicting the word.

Results of our tests on GLTR Test-2 

  • 125 English texts analysed.
  • 49/50 human-written texts recognised, i.e. 98% reliability on the recognition of human texts.
  • 55/75 AI-generated texts recognised, i.e. 73.33% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 83.20%.
gptzero chatgpt

GPTZero

 

Presentation of GPTZero

  • AI detector created by Edward Tian, a 22-year-old student at Princeton University (USA).
  • ChatGPT detector, based on the GPT-3 model.
  • No information provided on the languages processed.
  • Results are displayed as a final rating: human or AI text.

Results of our tests on GPTZero

  • 125 English texts analysed.
  • 42/50 human-written texts recognised, i.e. 84% reliability on the recognition of human texts.
  • 59/75 AI-generated texts recognised, i.e. 78.67% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 80.80%.
content at scale AI Detector

Content at Scale

 

Presentation of the AI sensor of Content at Scale

  • Company specialised in the development of marketing solutions.
  • ChatGPT detector, based on GPT-3.
  • No information provided on languages processed.
  • Display of results as % of AI and Human content.

Results of our tests on the AI detector of Content at Scale

  • 125 English texts analysed.
  • 43/50 human-generated texts recognised, i.e. 86% reliability on human text recognition.
  • 57/75 AI-generated texts recognised, i.e. 76% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 80%.
detectorgpt from sokat

DetectGPT from Sokat

 

Presentation of DetectGPT from Sokat

  • American company specialised in the development of AI solutions.
  • No information on the detected artificial intelligences.
  • No information on the languages processed.
  • Results displayed as % of AI content.

Results of our tests on DetectGPT from Sokat

  • 123 English texts analysed.
  • 49/50 human-written texts recognised, i.e. 98% reliability on human text recognition.
  • 46/75 AI-generated texts recognised, i.e. 61.33% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 77.24%.
crossplag ai content detector

Crossplag AI Content Detector


Presentation of Crossplag AI Content Detector

  • American company specialising in similarity detection
  • No information on the detected artificial intelligences
  • AI detector for English texts only
  • Results displayed as % of AI content.

Results of our tests on Crossplag AI Content Detector

  • 125 English texts analysed.
  • 43/50 human-written texts recognised, i.e. 86% reliability on human text recognition.
  • 52/75 AI-generated texts recognised, i.e. 69.33% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 76.61%.
kazan seo ai content detector

Kazan SEO

 

Presentation of the AI detector of Kazan SEO

  • Company specialised in the development of tools to help optimize the natural referencing of websites.
  • No information on the artificial intelligences detected.
  • No information on the languages processed.
  • Results displayed as % of real or artificial text.

Results of our tests on the Kazan SEO AI detector

  • 124 English texts analysed.
  • 40/50 human-generated texts recognised, i.e. 80% reliability on human text recognition.
  • 52/75 AI-generated texts recognised, i.e. 69.33% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 74.19%.
AI Text Classifier - Open AI ChatGPT

AI Text Classifier - Open AI (ChatGPT)

 

Introducing AI Text Classifier - Open AI (ChatGPT)

  • AI detector developed by Open AI, the company behind ChatGPT.
  • ChatGPT detector, based on GPT2 and GPT3.
  • Text detector in English only.
  • Results are displayed as a rating: AI-generated text very unlikely, unlikely or uncertain.

Results of our tests on AI Text Classifier from Open AI (ChatGPT)

  • 45 English texts analysed.
  • 22/28 human-generated texts recognised, i.e. 78.57% reliability on human text recognition.
  • 11/17 AI-generated texts recognised, i.e. 64.70% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 73.33%.
Copyleaks AI Content Detector

Copyleaks AI Content Detector

 

Presentation of Copyleaks AI Content Detector

  • American company specialised in the development of similarity detection software.
  • Chat GPT Detector, based on GPT-2 and GPT-3.
  • Multilingual detector: English, French, Spanish, German and Portuguese.
  • Display of results as % of text written by AI or human.


Results of our tests on Copyleaks AI Content Detector

  • 125 English texts analysed.
  • 48/50 human-written texts recognised, i.e. 96% reliability on human text recognition.
  • 42/75 AI-generated texts recognised, i.e. 56% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 72%.
Writer AI Content Detector

Writer AI Content Detector

 

Presentation of Writer AI Content Detector

  • Company specializ-sed in the development of AI solutions for web editors.
  • No information on the detected artificial intelligences.
  • No information on languages processed.
  • Results displayed as % of AI and Human content.

Results of our tests on the Writer AI Content Detector

  • 112 English texts analysed.
  • 39/41 human-written texts recognised, i.e. 95.12% reliability on human text recognition.
  • 38/73 AI-generated texts recognised, i.e. 52.05% reliability on the recognition of AI-generated texts.
  • Overall reliability rate: 68.75%.

Summary of the comparative analysis of the top 11 AI detectors

Here is a graph showing the reliability rates of each AI detector in relation to the number of tests performed:

comparison of the best AI detectors
Discover Compilatio AI Detector

In conclusion, we would like to remind you that artificial intelligence is part of our lives and that we need to adapt to this change, to learn to work with these tools

In education, for example, artificial intelligence can be useful to : 

  • improve the understanding of concepts by providing clear and concise explanations with concrete examples 
  • create interactive exercises that help students learn in a fun way
  • ... 


*POC: Proof Of Concept is a realisation whose purpose is to show the feasibility of a process or an innovation.


To go further :