Explainable Deep Learning Model for ChatGPT-Rephrased Fake Review Detection Using DistilBERT

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Overall

The authors have identified a worthy niche. However, they did not show why they developed both the dataset and the algorithm. Datasets already exist. The authors did not compare their results to other models. This reduces faith in the claims of the authors. The paper overall has lots of merit, but the shortcomings listed below need to be addressed.

Issues

1. The authors should benchmark their proposed models against existing models.

2. There are few details about their newly created dataset. First and foremost, why do the authors not use existing datasets. Second, authors should share details regarding prompt generation and the model used. The authors only state ChatGPT. If they used ChatGPT-2 or ChatGPT-4o the impact on the resultant quality of the dataset is huge.

3. Related work section is difficult to understand. Commonalities and differences are not drawn out. The lack of paragraphing makes it harder to follow, too.

4. Figure 1 is overly simple (and unnecessarily large given the lack of detail). Readers want to know the details of the pipeline. The figure needs more annotation, since the screenshots of the LIME and SHAP results are only understandable to readers who have used those XAI models before.

5. Equation 4 is incorrect.

6. Section 4 presents Experimental results. However, Subsection 4.1 appears to be the author’s observations of the contents of the dataset. There is no methodogy described for the observations.There are varies protocols for the  various types of observation, e.g syntactic analysis, semantic analysis, coherence assessment, style and tone evaluation, and contextual relevance checks. Additionally, many of the statements made do not hold true in my experience, which is why I worry about the veracity of the statements.

7. The authors make claims about ChatGPT’s style of writing without noticing the impact of prompts. Telling ChatGPT to adopt a particular style greatly impacts writing. I assume the authors mean the default setting. If so, this should be stated.

 

Comments on the Quality of English Language

Language

1. The paper is riddled with numerous language errors, many of which are unintrusive, but some block understanding of the meaning.

2. Presence of tortured phrases but no acknowledgement of the use of generative AI, e.g. “big language models” instead of “large language models”.

3. Lack of paragraphing

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper addresses the  important challenge of identifying AI-generated or AI-rephrased fake reviews, a challenge that has gained prominence in the digital age.  Leveraging DistilBERT and combining it with xAI techniques like LIME and SHAP is a good idea. However I have some remarks :

1) The findings may not generalize well to other domains such as product or travel reviews. Including datasets from diverse sources would make the conclusions more robust.

2) The paper briefly compares its results to state-of-the-art methods but does not dive deeply into why DistilBERT performs better or worse than other models. 

3)The study specifically targets ChatGPT-generated reviews. It would be beneficial to test whether the proposed model can detect fake reviews generated by other AI tools  to assess its broader applicability.

4)The societal implications of detecting fake reviews and the potential biases in dataset labeling and model predictions deserve more attention. 

I suggest some improvements:

 

1) Include data from various platforms  to evaluate the model’s performance across different contexts and review styles.

2)Test the proposed approach against more state-of-the-art models and explore hybrid architectures or ensemble methods to strengthen the comparative analysis.

3) Discuss how the model might evolve to detect fake reviews generated by more advanced generative AI systems.

4) Offer specific recommendations for integrating the model into real-world systems.

5)Add a section exploring the ethical implications of using AI to detect fake reviews.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Fake Reviews and ChatGPT’s Role: Can you elaborate on how ChatGPT-generated reviews differ from genuine user reviews in practical scenarios? Are there specific industries or product categories where these fake reviews are more prevalent?

 

Impact on Consumers: Do you have data or examples showing how these fake reviews influence customer decisions? Could you quantify the extent of harm or misinformation caused by these reviews?

 

Scope of the Study: Does the paper focus only on detecting rephrased reviews by ChatGPT, or does it also address reviews written from scratch?

 

Dataset Details: Can you provide more details about the dataset used? How was the balanced dataset constructed, and what criteria were used to label reviews as genuine or ChatGPT-rephrased?

 

Linguistic Patterns: What are the key linguistic features used in distinguishing ChatGPT-generated reviews from genuine ones? Did the linguistic differences vary based on the domain (e.g., product type)?

 

Pre-processing Techniques: Could you clarify how the pre-processing phase (e.g., POS tagging, tokenization) contributed to improving classification performance?

 

Fine-tuned DistilBERT: Why was DistilBERT chosen over other transformer-based models like BERT or GPT? How was the model fine-tuned, and what hyperparameters were used?

 

Real-World Applications: How can businesses or review platforms implement your model? What are the computational or logistical challenges in deploying it at scale?

 

Model Limitations: What are the limitations of the proposed approach? For example, can it detect reviews generated by future versions of ChatGPT or other AI tools?

 

Ethical Considerations: What ethical considerations were addressed when developing a model for detecting AI-generated reviews? For instance, how do you ensure no bias against legitimate, well-written reviews by humans?

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This paper discusses a new model designed to detect fake reviews, particularly those generated or rephrased by ChatGPT, which can distort perceptions of product quality. The model utilizes a balanced dataset to analyze sentiment and linguistic patterns in both real and fake reviews. It incorporates preprocessing steps like POS tagging, lemmatization, and tokenization, and employs a fine-tuned DistilBERT transformer model for predictions. Additionally, the model leverages explainable AI techniques (LIME and SHAP) to provide insights into its classification decisions, ensuring a clearer understanding of its logic and predictions.

I believe this paper can be of value for readers who are keen on explainable AI techniques. However, the paper has several issues that must be addressed by the authors. As such,

1- How does the proposed model distinguish itself from existing methods for detecting ChatGPT-generated or rephrased fake reviews? What specific advancements or unique techniques does it introduce compared to other studies? Please discuss these issues more clearly in Abstract, Introduction, and the main body of the paper.

2- Given the sensitive nature of detecting fake reviews and potential ethical implications, how does the paper address concerns related to privacy, potential misuse of the model, or the risks of over-reliance on automated systems for review verification?

3- Since the paper focuses on detecting fake reviews generated or rephrased by ChatGPT, how does the model perform against outputs from different versions of ChatGPT? For example, was it tested on both GPT-3.5 and GPT-4 outputs to ensure robustness across versions?

4- How does the proposed model distinguish between reviews written by ChatGPT and those written by humans but heavily influenced by templates or AI suggestions? What safeguards are in place to minimize false positives?

5- While the use of LIME and SHAP for interpretability is commendable, how reliable are these techniques in explaining the nuanced linguistic patterns specific to ChatGPT-generated text? Were there specific cases where these techniques failed to provide meaningful insights?

6- As ChatGPT evolves and generates increasingly sophisticated outputs, how does the proposed detection model adapt to these improvements? Does the model require regular retraining to remain effective, and how feasible is this in real-world applications?

7- Combining literature such as (1) "Deep Metric Learning with Soft Orthogonal Proxies, arXiv preprint arXiv:2306.13055 (2023)," (2) "Detecting ChatGPT: A Survey of the State of Detecting ChatGPT-Generated Text, arXiv preprint arXiv:2309.07689 (2023)," and (3) "Fake News Detection: Comparative Evaluation of BERT-Like Models and Large Language Models with Generative AI-Annotated Data, Knowledge and Information Systems (2025): 1-26" is recommended, as they can enhance the quality and depth of the review in the current research.

Comments on the Quality of English Language

The manuscript shows promise; however, the English language quality needs enhancement to improve clarity and readability. Certain sentences are challenging to comprehend due to grammatical errors and awkward phrasing. I suggest the authors consider professional language editing services or assistance from a fluent English speaker to refine the manuscript's overall quality. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Overall

The content of the article is timely and of interest to readers. Although I believe the authors have a motivation for their study it is not conveyed clearly to the readers.  The paper is well-written with no intrusive language issues. The authors have made an attempt to address the reviewer’s concerns. The paper, however, still falls short of expectations in terms of methodology rigour. I detail some major issues below

Major issue

1. After the review of related works, there is no segway to the proposed work of the author. The idea of the literature review/description of related works is to show how your research fills a research niche, i.e. a gap in the current research. The gap is not shown clearly.
2. 3.1 “Existing datasets for fake review detection primarily focus on either human-written deceptive reviews or fully AI-generated text and do not focus on rephrased versions of real human reviews, which pose a unique challenge due to their hybrid nature.” This claim is not supported and based on quick search appears false. State specifically which datasets, and then you can point to the gap. However, please check common dataset sources carefully.
3. 3.2 Preprocessing. The order of pre-processing steps is unusual. Why would you tokenize after word lemmatization?
4. LLM details – there is still no detail in the method on the ChatGPT model used. If this was through an API, there are multiple other parameters that deserve mention. Currently, this is not replicable.
5. Benchmarking is attempted, but not successfully. The authors created a dataset and a method but do not sun other methods on the same dataset and do not run their method on other datasets, so it is impossible to declare SOTA or even compare.

Currently, the biggest weakness in the paper is the opaqueness of the method.

Minor issue

1. Figure 1: A horizonal work flow is much more logical. This is a clockwise work flow ignoring the XAI block.
2. Selection of SHAP and LIME. There are many other libraries for interpreting ML models, so why were these selected?

Although not part of the reviewer remit, I also noted some language issues that the authors may want to address.

Research writing issues

1. Abstract “The obtained experimentalresults indicate” - - > include space between words
2. Abstract: (LIME) Local Interpretable Model-agnostic Explanations 9 and (SHAP) Shapley Additive Explanations – The abbreviation is placed after the full form, not before.
3. Introduction: “The remaining of this paper” - -> “The remainder of this paper”
4. Define both terms: “convolutional neural network (CNN) and RNN,”
5. Don’t define the same term twice “convolutional neural networks (CNNs).”
6. Don’t define terms you do not reuse e.g. “Amazon reviews (OR)” (and why select OR?) and GPT-2 model (CG).
7. 3.4 Explainable AI XAI à Explainable AI (XAI)
8. Define LIME and SHAP on first usage

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

It can be accepted

Author Response

Thank you

Reviewer 4 Report

Comments and Suggestions for Authors

The paper is ready for publication.

Author Response

thank you

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have tried to minimally address most of the previous comments by reviewers. However, when reviewers suggest "reject" rather than "major revision", minimal revisions are unlikely to and in this case do not persuade reviewers to shift their opinion.

To me this "improved" version is somewhat more readable than both version 1 and version 2, but without some substantial change, I fail to see how the authors have addressed the key concerns previously raised. I understand that the authors may not want to act on suggestions for multiple reasons, and so in hope of appeasing reviewers rather than taking action, simply move them to "future work", but this does not address the key issue. What does this paper contribute to the current body of research? The authors state that they have created something that performs better than a pre-trained base model. However, that is not difficult and certainly does not make a paper worthy of publication. At the very least, with this very minimal claim, the authors should test on multiple datasets against multiple base models. Even this, though, is not what many NLP academics would accept as a claim to SOTA. The aim is to produce a system that breaks the state of the art, not one whose results exceed the performance of a basic out-of-the-box model. 

I have no new concerns, but all the concerns raised in earlier reviews continue to exist. 

Comments on the Quality of English Language

no new comments

Author Response

please see the attachment

Author Response File: Author Response.pdf