DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning

We would like to thank Reviewer Tv1c for the valuable feedback. The comments and suggestions have greatly helped us in improving the quality of our work. Please see below for our responses to your comments.

Baselines seem to be quite weak/over a year old.

1. We need to clarify that the comparison methods presented in our paper are commonly used baselines in the task of AI-generated text detection, and most of them are the latest state-of-the-art (SOTA) approaches as the table shows in the top-level Author Rebuttal. All the baseline schemes we identified are compared with DetectGPT. To align with related works in the field, we also include corresponding results.
2. Watermarking algorithms rely on embedding specific markers during the generation process. In contrast, our neural network based approach does not require any intervention in the generation process and can universally detect all texts. Since our goal is to provide a general detection method independent of the generation process, rather than intervening and controlling the model's generation process, our approach isn't suitable and feasible for comparison with watermarking algorithms. Moreover, watermarking algorithms are difficult to implement in closed-source models, and the flexible modifiability of open-source models also limits their widespread application.
3. The comparison with Binoculars and GPTZero can be respectively found at the Table 2 and Table 3 within the uploaded PDF file. The experimental results confirm that our method continues to outperform these two supplementary approaches.

Limited experiments on newer GPT-4 class LLMs.

In the out-of-distribution (OOD) detection experiment on the M4 dataset, a subset of the testing data is from GPT-4, and is utilized to evaluate the model's OOD detection capability. Our method demonstrates a performance advantage over comparative approaches on this dataset. Additionally, we supplement OOD experiments based on the "Unseen Domains & Unseen Models" dataset from the paper "MAGE: Machine-generated Text Detection in the Wild" presented at ACL 2024. The testing data is sourced entirely from GPT-4, and the results indicate that our method performs better than the comparative schemes. (Longformer† indicates that Longformer uses data from the testing set to determine the detection threshold.)

Limited emphasis on attack robustness.

As you mentioned, a portion of the M4 dataset has undergone paraphrasing enhancement by OUTFOX, and our method outperforms all comparative solutions on this dataset. Additionally, we have supplemented our experiments with assessments of attack robustness, with the results shown in the top-level Author Rebuttal and Table 4 within the uploaded PDF file. The results demonstrate that our approach possesses strong resistance to attacks, indicating robust performance.

It's not clear whether gains over baselines are coming from the SimCSE initialization or training.

As described in Section 4.2 of the paper, we conduct experiments with a variety of text encoders to verify that our approach can enhance the performance of AI-generated text detection. The results are already presented in Table 4 within the original submission. This indicates that our method is not limited to a single encoder like SimCSE but is applicable to various encoders, and the performance gains come from our training.

AI-generated text detectors typically need to operate in a low FPR setting, to minimize the risks of labeling innocent text as AI-generated.

Our classification approach is based on dense retrieval and KNN (K-Nearest Neighbors) classification. This approach differs from traditional probability-based classifiers in that it does not rely on probability outputs. Specifically, the KNN classifier makes decisions based on the proximity of neighboring samples, rather than estimating the probability distribution of categories. Due to the absence of probability outputs, our method cannot define and adjust the threshold for the False Positive Rate (FPR). Typically, a low FPR range (such as 0-1%) is calculated by selecting different thresholds on the model's output probabilities, thereby measuring the True Positive Rate (TPR) at a given FPR. In our setup, the classification outcome is derived from the distance measurement of the nearest neighbors, without any involvement of probability scoring or threshold adjustment, providing only a global FPR and TPR. Therefore, we are unable to provide the corresponding TPR data within a specific low FPR range, such as 0-1%.

Thank you for your response and acknowledgment of our supplementary experiments. We are currently conducting experiments on watermarking dataset, this experiment requires a certain amount of time because we need to conduct evaluations on watermarking specific dataset, which entails some preliminary data processing work. We will include the results in the final version of the supplementary material within our paper.

Suppose we liken our task to appreciating paintings, needing to categorize them into "Impressionist" and "non-Impressionist" genres. If we focus solely on these two broad categories, we might overlook the stylistic differences between works by Monet and Renoir within the Impressionist category. Fine-grained clustering, however, is akin to not only distinguishing between Impressionist and non-Impressionist works but also to further differentiating the styles of various artists within the Impressionist genre itself. When we can differentiate between Monet's and Renoir's works, it means we have gained a deeper understanding of Impressionist pieces, thus making us more adept at distinguishing between the "Impressionist" and "non-Impressionist" categories. This more nuanced classification aids in better understanding and appreciating the artwork, enhancing the precision of our discernment. Similarly, in AI-generated text detection, incorporating multi-level contrastive learning allows the model to more precisely identify and differentiate texts from various sources.

Our ablation study, as shown in Table 3 within the paper, also confirms the multi-level contrastive learning outperforms plain contrastive learning.

We would like to thank Reviewer W7ZQ for the valuable feedback. The comments and suggestions have greatly helped us in improving the quality of our work. Please see below for our responses to your comments.

The proposed multi-level contrastive loss shares some similarities with existing work so it is not exciting enough.

We would like to elaborate on the differences between our method and previous methods from the following aspects:

1. This paper utilizes a form of contrastive learning that adheres to the traditional definition of positive and negative sample pairs, which is also the plain contrastive learning(PCL) approach that we compared against in our paper. Their use of contrastive learning in conjunction with selective masking strategy aims to achieve better text representations. However, one of the core contributions of our work is the proposal of a multi-level contrastive learning(MCL) framework. We define multi-level positive and negative sample pairs and derive a loss function for multi-level contrastive learning, which is different from previous studies.
2. Additionally, we perform classification based on feature clustering, distinguishing it from traditional binary classification methods. Such a classification approach effectively enhances the model's detection capabilities and robustness.
3. Lastly, within the multi-level contrastive learning framework, our model also has the ability of Training-Free Incremental Adaptation, effectively improving out-of-distribution (OOD) detection capability. This is also an important capability that has not been explored or discovered in previous methods.

Comprehensive ablation studies on our multi-level contrastive learning framework are performed in Section 4.3. From the experimental results, it can be seen that our method surpasses plain contrastive learning framework. Each component improves the model's performance, thereby validating the effectiveness of our proposed framework.

Although the authors conduct extensive experiments on multiple datasets, only few of them are designed specifically for AI text detection (GLTR, DetectGPT). They should provide more comparisons with SOTA methods.

Thanks for your affirmation of the amount of our experiments. We want to clarify that methods we compared across each dataset are commonly used baseline methods in the field of AI-generated text detection. The two comparison methods you referenced, Ghostbuster and GPT-Sentinel, have been factored into our research. In our paper, we make a comparison against T5-Sentinel, a latest work proposed by the author of GPT-Sentinel, demonstrating superior performance relative to GPT-Sentinel. Detailed results can be observed in Table 1 of our paper. As for Ghostbuster, due to the necessary utilization of OpenAI API interface for testing and because our test dataset is quite voluminous, it is not an appropriate choice for comparison. Instead, we select Binoculars as a stand-in comparison method, which outperforms Ghostbuster in their paper. Our approach outperforms all compared methods; see Table2 in the uploaded PDF for details. Recently, an AI-generated text detection competition was held based on the M4 dataset. We also incorporate all the solutions from this competition into our list of comparison methods, details of which can be found in Table1 of the uploaded PDF file. In the monolingual test set of the M4 dataset, we achieve the top score, surpassing all other participating schemes. In the multilingual test set, we also secured the fifth place.

As a work towards detecting AI-generated texts, more analysis of the features of texts should be provided.

In Section 4.3 of the submission, we provide a visualization analysis of learned text embeddings. Through the UMAP dimensionality reduction visualization, it can be seen that text embeddings of different types have successfully clustered together, which demonstrates the effectiveness of our method.

I am wondering if this framework would do a good job in authorship attribution?

We have supplemented the performance of our proposed method on the task of authorship attribution detection on TuringBench dataset, the results can be found at the top-level Author Rebuttal. The experimental results indicate that our approach achieves state-of-the-art performance, demonstrating its ability to accomplish this task.

For the classification approach in ablation studies, is the classification head trained with your own data or initialized in other ways?

Our classification head is trained with the training data and then tested.

Could you please provide some explanation on the advantage of multi-level contrastive loss over pcl?

Please see more explanation in the comment below your review block.

The authors mention they do not explore the robustness of the proposed method. The validation could be conducted under different attacks like paraphrasing, editing, prompting.

Thanks for your suggestions. In this paper, our focus is on proposing a general method for AI-generated text detection. As introduced in Section 4.1 of the paper, the M4 dataset includes testing data that has undergone paraphrasing attack, which can be used to verify the model's ability to resist attacks. We have also conducted additional experiments regarding attack robustness which can be found at the top-level Author Rebuttal and the uploaded PDF file, and the experimental results demonstrate that our method possesses strong resilience against attacks.

We sincerely appreciate your acknowledgment of our proposed method as an inspiring work and for the commendable performance of our framework. This recognition serves as an encouragement in our research endeavors. Next, we will address your after-rebuttal questions in detail.

Q1

Thank you for your thoughtful and detailed feedback. We appreciate the points you've raised and would like to address your concerns directly from the following perspectives.

• The research objective and insights of our work.
  1. Performance vs. Insights: As stated in the Abstract and Introduction of the paper, our insight is that the key to accomplishing AI-generated text detection resides in distinguishing the multi-level writing styles of different "authors", rather than just modeling this task as a binary classification problem. Therefore, based on this key insight, we design the DeTeCtive framework. Specifically, we define multi-level positive and negative sample pairs according to the affinity relations between different language models and derive a multi-level contrastive learning loss function. Eventually, we accomplish the learning of multi-level features through contrastive learning. Moreover, under the DeTective framework, we do not focus on improving binary classification performance. Instead, within our framework, the model's learning objective is to capture multi-level features among different large language models. The principal objective function is the multi-level contrastive learning loss we designed, rather than a simple binary cross-entropy loss. Furthermore, our classification method is based on the K-Nearest Neighbors (KNN) clustering of embeddings. Therefore, the improvement in model performance results from successful learning of multi-level features rather than relying on the classification head. Hence, both the initial intent of framework design and the way of model learning reflect our insights. We believe that the end-to-end method we proposed, which outperforms baselines (e.g., SCL/PCL) and all the comparison methods (e.g., DetectGPT, GPT-Sentinal, CoCo) by a significant margin, is not just a trivial improvement. The substantial performance gains suggest that our method is capturing internal patterns and nuances that are specific to AI-generated text. These internal insights could offer valuable perspectives on how machines differentiate between human and AI-generated content effectively.
  2. Human-Friendly Insights vs. Machine Efficiency: As elaborated above, our understanding and insights of this task are reflected in the design of the method and framework. The visualization of learned text features also reveals that features at different levels (e.g., by model famliy, by individual model) could cluster well, which is something typical supervised contrastive learning (SCL) cannot achieve due to the lack of multi-level relationship constraints. We believe that the insights upheld by our paper, namely, that the key to accomplishing AI-generated text detection task resides in distinguishing the multi-level writing styles of different "authors", is intuitive and human-friendly. The effectiveness of our approach has also been proven through comprehensive experiments and the results are also consistent with our insights.
• Weaknesses of existing works.
  1. Unsatisfactory generalizability: The out-of-distribution (OOD) detection results shown in Table2 of our original submission, the performance of existing methods on OOD data is quite poor. This suggests that the handcrafted features and their insights discovered are challenging to generalize to OOD data, significantly hindering the practical application of the algorithm. Considering the rapid development of language models, this presents a problem.
  2. Performance bottlenecks and lack of comprehensive evaluations: Compared to our work, existing solutions lack a comprehensive evaluation on various benchmarks, testing scenarios and other applications. The experimental results of existing methods on multiple benchmarks are not ideal, indicating that there are performance bottlenecks. However, our method surpasses the existing state-of-the-art solutions by a large margin on each individual dataset.

Q2

Please allow me to clarify your misconception about our supplementary results for the task of authorship attribution detection. These corresponding experimental results are detailed in the section titled "Supplementary experiments on other applications. (Reviewer yy53, Reviewer W7ZQ)", which can be found at the top-level block of Author Rebuttal by Authors. While Table2 of the attched PDF file shows the results of the additional baseline schemes on several existing datasets on the task of AI-generated detection, so the results are consistent with that in Table1 of the original submission. Due to the length constraint of the uploaded PDF file, we only list the additional results of authorship attribution detection in the rebuttal reply. We commit to update these results in the final version afterward.

Thank you again for your comments. We sincerely hope that our responses can address your questions.

• The motivation and contributions of our research work.
  1. Motivation: In response to the aforementioned weaknesses of existing works, the motivation of our study is to propose an AI-generated text detection algorithm that can be widely applied to various language models and scenarios. It could effectively adapt to newly released large language models and other unseen domains, demonstrating good generalization and robustness.
  2. Contributions: Our work contributes to the field of AI-generated text detection by proposing a novel method that can effectively detect AI-generated text. Based on our framework, we introduce Training-Free Incremental Adaptation (TFIA), a scheme to enhance the model's out-of-distribution (OOD) detection capabilities. We validate our method on multiple benchmarks, where it consistently outperforms existing solutions, especially in terms of OOD detection performance. The visualization of text embeddings also validates that our model captures multi-level text features, which aligns perfectly with our insights. Additionally, we supplement experimental results on attack robustness and authorship attribution detection, further demonstrating our method's superior generalization and robustness surpassing existing solutions. We believe such a novel framework, paired with comprehensive experimental evaluations and state-of-the-art performance, contributes to the community, fostering the application of our algorithm in real-world scenarios.

In conclusion, we believe that our approach offers a complementary perspective that enhances the performance of AI-generated text detection and also aligns well with our key insights. Moreover, we have also conducted corresponding analysis and explorations based on our method, including extensive experiments under various testing scenarios (e.g., multiple benchmarks, OOD detection, authorship attribution detection, attack robustness) and the Training-Free Incremental Adaptation (TFIA) capabilities. These solid experimental results and findings are contributions of our work to the research community. Finally, we are open to incorporate further discussion on interpretability of our method into the final version.

Dear reviewer,

We kindly request your feedback on whether our response has addressed your questions. If you have any remaining questions or concerns, we are happy to address them. Thank you for your time and consideration!

We would like to thank Reviewer yy53 for the valuable feedback. The comments and suggestions have greatly helped us in improving the quality of our work. Please see below for our responses to your comments.

The paper's strong performance may not generalize to other datasets or applications.

We appreciate your acknowledgment towards our proposed framework and experimental results. In the submission, we conduct evaluations on three widely-used datasets for AI-generated text detection. These include a variety of testing scenarios, such as: multi-model, multi-domain, multi-language, text-paraphrase attack, and out-of-distribution (OOD) detection, among others. The experimental results demonstrate that our method achieves state-of-the-art performance in these various scenarios. Furthermore, we supplement our experiments on attack robustness and authorship attribution detection, and we also add performance comparisons with more baseline methods on existing datasets. Please refer to the tabels in the top-level Author Rebuttal and the uploaded PDF file for details. Additionally, we supplement OOD evaluation based on the "Unseen Domains & Unseen Models" dataset from the paper "MAGE: Machine-generated Text Detection in the Wild" presented at ACL 2024. The testing data is entirely derived from GPT-4, with the corresponding results presented in the table below. The results also indicate that our method performs better than other solutions. These additional experimental results will be incorporated into the final version of our paper.

To sum up, our approach exhibits state-of-the-art performance across various datasets, testing scenarios, and applications.

(Longformer† indicates that Longformer uses data from the testing set to determine the detection threshold.)

How does the quality and diversity of the training data affect the performance of DeTeCtive?

In the experiments, we use publicly available datasets and strictly follow the divisions of the training and testing sets without introducing any additional data to assist model training. Moreover, we believe that increasing data diversity and quality can effectively enhance the performance of our method. Under the proposed multi-level contrastive learning framework, diverse data compose more positive and negative sample pairs, enabling the model to learn more fine-grained writing-style features and relationships, thereby improving the model's performance. We believe that scaling up the data volume within our proposed framework holds promise as a future research direction.

Are there specific datasets that are more beneficial for training the model?

Within our proposed framework, the process is streamlined, requiring only the texts and labeling of different large language models. Following this, our method facilitates the construction of multi-level contrastive learning sample pairs. Consequently, there is no necessity for specific datasets to train our model.

The focus on distinguishing writing styles might overlook content-based cues that could also indicate AI-generated text. This might lead to missed detections if an AI successfully mimics human writing style while generating misleading content.

We appreciate you pointing out the potential limitation of our method. We will address this issue from the following two aspects, hoping to alleviate your concerns:

1. As our approach is based on fine-tuning pre-trained text encoders, which already possess the capability of content-based semantic understanding (such as BERT, SimCSE, etc.), we believe that the model can understand the content and semantics of the text itself.
2. Further, we conduct paraphrasing-attack experiments for validation. Specifically, OUTFOX and DIPPER paraphrase texts generated by large language models to mimic human-written texts, employing this as a paraphrasing attack strategy. We train our model on the training set provided by OUTFOX and test under the following three scenarios: Non-attacked, OUTFOX paraphrasing attack, and DIPPER paraphrasing attack. We also compare with several baseline methods, and the results are as follows: | Attacker | Detector | HumanRec | MachineRec | AvgRec | F1 | |---|---|---|---|---|---| | Non-attacked | RoBERTa-base | 93.8 | 92.2 | 93.0 | 92.9 | | | RoBERTa-large | 91.6 | 90.0 | 90.8 | 90.7 | | | HC3 detector | 79.2 | 70.6 | 74.9 | 73.8 | | | OUTFOX | 99.0 | 94.0 | 96.5 | 96.4 | | | DeTeCtive | 98.2 | 100.0 | 99.1 | 99.1 | | DIPPER | RoBERTa-base | 93.8 | 89.2 | 91.5 | 91.3 | | | RoBERTa-large | 91.6 | 97.0 | 94.3 | 94.4 | | | HC3 detector | 79.2 | 3.4 | 41.3 | 5.5 | | | OUTFOX | 98.6 | 66.2 | 82.4 | 79.0 | | | DeTeCtive | 97.4 | 97.9 | 97.7 | 97.5 | | OUTFOX | RoBERTa-base | 93.8 | 69.2 | 81.5 | 78.9 | | | RoBERTa-large | 91.6 | 56.2 | 73.9 | 68.3 | | | HC3 detector | 79.2 | 0.4 | 39.8 | 0.7 | | | OUTFOX | 98.8 | 24.8 | 61.8 | 39.4 | | | DeTeCtive | 95.4 | 98.6 | 97.0 | 96.9 |

The results indicate that even when AI-generated texts are subjected to paraphrasing attack to mimic human writings, DeTeCtive experiences only a minimal performance decline compared to the non-attacked scenario and remains effective in detecting AI-generated text. In contrast, other solutions we compared all exhibit significant performance degradation. Therefore, we believe that our framework demonstrates good robustness against mimicry of human writing style.