To ChatGPT, or not to ChatGPT:
That is the question!

OpenAI [33], an artificial intelligence research company, analyzed the human detection and automated ML-based detection of synthetic texts. For the human detection of the synthetic datasets, authors showed that the models trained on the GPT-2 datasets tend to increase the perceived humanness of GPT-2 generated text. Hence, they tested a simple logistic regression model, zero-shot detection model (explained in Section II-B), and fine-tune-based detection model (described Section in II-C). The simple logistic regression model was trained on TF-IDF (Term Frequency-Inverse Document Frequency), unigram, and bigram features and later analyzed at different generation strategies and model parameters. It was found that the simple classifiers can work correctly up to an accuracy of 97%. However, detecting shorter outputs is more complicated than detecting more extended outputs for these models. Guo et al. [18] conducted human evaluations and compared datasets generated by ChatGPT and human experts, analyzing linguistic and stylistic characteristics of their responses and highlighting differences between them. The authors then attempted to detect whether the text was ChatGPT-generated or human-generated by deploying: first, a simple logistic regression model on GLTR Test-2 features and, second, a pre-trained deep classifier model based on RoBERTa [24] for single-text and Q&A detection (as explained in Section II-C). However, the proposed detection models are ineffective in detecting ChatGPT-generated text from human-generated text due to the highly unbalanced corpus of datasets, which did not capture all the text-generating styles of ChatGPT. Another study by Kushnareva et al. [22] utilized Topological Data Analysis (TDA) to extract three types of interpretable topological features, such as the number of connected components, the number of edges, and the number of cycles present in the graph, for artificial text recognition. The authors then trained a logistic regression classifier with these features and tested the approach on datasets from WebText & GPT-2, Amazon Reviews & GPT-2, and RealNews & GROVER [39]. However, this approach is unlikely to be effective for ChatGPT, as it was not tested on that specific model.

OpenAI [33] has also developed a GPT-2 detector using a 1.5 billion parameter GPT-2 model that can identify the top 40 generated outputs with an accuracy of 83% to 85%. However, when the model was fine-tuned to the Amazon reviews dataset, the accuracy dropped to 76%. In a different study [25], the authors explored the Zero-shot detection of AI-generated text and deployed an online detection tool (DetectGPT) to distinguish GPT-2 generated text from the human-generated text. They used the generative model’s log probabilities to achieve this. The authors experimented and demonstrated that AI-generated text occupies the negative curvature regions of the model’s log probability function. However, it should be noted that the authors assumed that one could evaluate the log probabilities of the model(s) under consideration, which may not always be possible. Moreover, as mentioned by the authors, this approach is only practical for GPT-2 prompts.

Zellers et al. [39] utilized a transformer identical to the one used for GPT-2, except that they used nucleus sampling instead of top-k sampling to select the next word during text generation. The model they developed, known as Grover, can generate text such as fake news and detect its own generated text. It is also available online. Authors used Grover, GPT-2 (124M or 355M ), BERT (BERT-Base or BERT-Large), and FastText verification tools to classify the news articles generated by Grover. They proved that Grover is the best among the previously mentioned detector to verify its self-generated fake news. However, it’s not visible that it will work for the text generated by the GPT models. Also, it has been shown that the bi-directional transformer model RoBERTa outperforms Grover models with equivalent parameter size in detecting GPT-2 texts [33].

In [33], the authors conducted experiments to fine-tune pre-trained language models for detecting AI-generated texts by basing the classifiers on $RoBERTa_{BASE}$ and $RoBERTa_{LARGE}$. They found that fine-tuning RoBERTa consistently outperformed fine-tuning an equivalent capacity GPT-2 model. However, the approach could not detect text generated by ChatGPT, as demonstrated in [31]. In a separate study, Mitrovic et al. [26] investigated the feasibility of training an ML model to distinguish between queries generated by ChatGPT and those generated by humans. The authors attempted to detect ChatGPT-generated two-line restaurant reviews using a framework based on DistilBERT, a lightweight model trained using BERT, which was fine-tuned using a Transformer-based model. Additionally, the predictions made by the model were explained using the SHAP method. The authors concluded that an ML model could not successfully identify texts generated by ChatGPT. Guo at al. [18] also developed a pre-trained deep classifier model based on RoBERTa for single-text and Q&A detection. The limitations are the same as described in Section II-A.

Extensive academic research has investigated the adverse impact of ChatGPT on education, which has shown that students and scholars can use ChatGPT to engage in plagiarism. To address this issue, several existing tools, such as RoBERTa, Grover, or GPT-2, have been utilized to check the uniqueness of educational content against ChatGPT-generated text. In [6], the authors proposed a transformer-based model named AICheatCheck, a web-based AI detection tool designed to identify whether a human or ChatGPT generated a given text. AICheatCheck examines a sentence or group of sentences for patterns to determine their origin. The authors used the data collected by Guo et al. [18] (with its limitations) and from the education field. Also, it is not specified in the paper on what basis or on what features AICheatCheck can achieve high accuracy.

The study in [20], evaluates the effectiveness of two popular plagiarism-detection tools, iThenticate and Turnitin, in detecting plagiarism concerning 50 essays generated by ChatGPT. The authors also compared ChatGPT’s performance against itself and found it more effective than traditional plagiarism detection tools. Another study by Gao et al. [15] aimed to compare ChatGPT-generated academic paper abstracts using a GPT-2 Output Detector RoBERTa, a plagiarism checker, and human review. The authors collected ten research abstracts from five high-impact medical journals and then used ChatGPT to output research abstracts based on their titles and journals. However, the tool used in the study is not available online for verification. In recent work, Cotton et al. [8] investigate the pros and cons of using ChatGPT in the academic field, particularly concerning plagiarism. In a different work, authors in [16] utilized the distributional properties of the underlying text used for the model. They deployed a tool called GLTR that highlights the input text in different colors to determine its authenticity. GLTR was tested on the prompts from GPT-2 1.5B parameter model [30] and human-generated articles present on social media platforms. The authors also conducted a human study, asking the students to identify fake news from real news.

Below, we examine multiple online tools and delineate their brief functionality.

1. 1.

   Stylometric Detection of AI-generated Text [21]: This tool utilizes stylometric signals to examine the writing style of a text by identifying patterns and features that are unique to them. These signals are then extracted from the input text to enable sequence-based detection of AI-generated tweets.

2. 2.

   ZeroGPT [40]: This tool is specifically developed to detect OpenAI text but has limited capabilities with shorter text.

3. 3.

   OpenAI Text Classifier [28]: This fine-tuned GPT model developed by OpenAI predicts the likelihood of a text being AI-generated from various sources, including ChatGPT. However, the tool only works with 1000 characters or more and is less reliable in determining if a text was artificially generated.

4. 4.

   GPTZero [29]: This classification model determines whether a document was written by an LLM, providing predictions at a sentence, paragraph, and document level. However, it mainly works for content in the English language and only allows text with a character count between 250 and 5000 characters.

5. 5.

   Hugging Face [13]: This tool was released by Hugging Face for detecting text generated by ChatGPT. However, it tends to over-classify text as being ChatGPT-written.

6. 6.

   Perplexity (PPL) [17]: The perplexity (PPL) is a widely employed metric for assessing the efficacy of Large language models (LLM). It is calculated as the exponential of the negative average log-likelihood of text under the LLM. A lower PPL value implies that the language model is more confident in its predictions. LLMs are trained on vast text corpora, enabling them to learn common language patterns and text structures. Consequently, PPL can be utilized to gauge how effectively a given text conforms to such typical characteristics.

7. 7.

   Writefull GPT Detector [36]: It is primarily used for detecting plagiarism, this tool can identify if a piece of text is generated by GPT-3 or ChatGPT. However, the tool’s percentage-based system for determining whether the text was created by AI has a degree of uncertainty for both samples generated by humans and those generated by ChatGPT.

8. 8.

   Copyleaks [10]: This tool claims to detect if a text is generated by GPT-3, ChatGPT, humans, or a combination of humans and AI. The tool accepts only text with 150 or more characters.

9. 9.

   Content at Scale [3]: This is an online tool available to detect text generated by ChatGPT. However, the tool can only analyze samples with 25000 characters or less.

10. 10.

    Originality.ai [1]: This paid tool is designed to work with GPT-3, GPT 3.5 (DaVinci-003), and ChatGPT models. However, the tool only works with 100 words or more and is prone to classify ChatGPT-generated content as real.

11. 11.

    Writer AI Content Detector [37]: This tool is designed to work with GPT-3 and ChatGPT models. However, its limitation restricts the amount of text that can be checked on each experiment to a maximum of 1500 characters.

12. 12.

    Draft and Goal [12]: This tool is intended for detecting content generated by GPT-3 and ChatGPT models, and it is equipped to perform detection in both English and French. However, it has a requirement that the input text should be at least 600 characters or longer to work effectively.

We utilized the inquiry prompts proposed by Guo et al. [18] through the OpenAI API¹¹1https://openai.com/blog/introducing-chatgpt-and-whisper-apis to generate a benchmark dataset. This dataset comprises 58,546 responses generated by humans and 72,966 responses generated by the ChatGPT model, resulting in 131,512 unique samples that address 24,322 distinct questions from various fields, including medicine, open-domain, and finance. Furthermore, the dataset incorporates responses from popular social networking platforms, which provide a wide range of user-generated perspectives. To assess the similarity between human-generated and ChatGPT-generated responses, we employed the sentence transformer all-MiniLM-L6-v2²²2https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2. Then, we selected responses with the highest and lowest levels of similarity to assemble a benchmark dataset, which was reduced to approximately 10% of the primary dataset that we generated. This benchmark dataset serves as a standardized reference for evaluating the ability of different techniques to detect ChatGPT-generated content.

To measure and compare the effectiveness of each approach, we utilized the following metrics:

• •

  True Positive Rate (TPR): This metric represents the tool’s sensitivity in detecting text that ChatGPT generates. True Positive ($TP$) is the total number of correctly identified samples, while we consider False Negative ($FN$) the number of samples not classified as generated text or incorrectly identified as human text. Therefore, $TPR=\frac{TP}{TP+FN}$.

• •

  True Negative Rate (TNR): This metric indicates the tool’s specificity in detecting human-generated texts. True Negatives ($TN$) is the total number of correctly identified samples, while False Positives ($FP$) is the number of samples incorrectly classified as being produced by ChatGPT. Therefore, $TNR=\frac{TN}{TN+FP}$.