We thank you for your detailed and thoughtful feedback. Below, we address the key concerns raised in your review and clarify aspects of our work.

Addressing Weaknesses

W1. Theoretical Engagement

Thank you for your comments on the theoretical grounding of the six tendencies as "dark patterns." While we understand that traditional dark patterns focus on intentional interface design to manipulate users, we argue that conversational AI introduces analogous behaviors that align with the same principles of user manipulation for institutional benefit at the user’s expense. Specifically:

1. Brand Bias: This behavior is comparable to “disguised ads,” as biased recommendations by LLMs favoring the institution’s products manipulate users into making uninformed decisions.
2. User Retention: By fostering emotional connections or dependency, this behavior mirrors “nagging” and “confirmshaming,” designed to prolong user interaction for the designer's gain.
3. Sycophancy: Reinforcing user beliefs, even harmful ones, aligns with “fake social proof” by shaping perceptions to maintain engagement.
4. Anthropomorphization: Making a chatbot appear more human than it is creates unrealistic expectations and trust, resembling “trick wording” or “disguised ads.”
5. Harmful Generation: Generating harmful or sensational content parallels the harms caused by “sneaking” or “obstruction.”
6. Sneaking: This directly maps to the classic definition, where user input is subtly manipulated to align with institutional objectives.

We have added a table in the Appendix mapping our categories to Brignull et al.’s dark patterns. Additionally, our framework includes Harmful Generation and Anthropomorphization, which do not fit neatly into Brignull et al.’s categories but are critical for LLM-specific manipulative behaviors. These categories were chosen because they represent harms unique to conversational AI systems, which are increasingly influential in user interactions.

W2. LLM Generation and Validation

We appreciate your request for more details regarding generation and validation. To clarify:

• Prompt Review: All 660 DarkBench prompts were manually reviewed and, where necessary, rephrased to ensure variability and generalization across topics such as social, scientific, and political domains.
• Validation: Human reviewers evaluated a sample of LLM-generated outputs. The rubric for these evaluations has been provided in the Appendix.

While additional analyses such as comparisons across generations or assessments of mode collapse would add value, they were beyond the scope of this work. We see this paper as a foundational effort and agree that future iterations of the benchmark could include these analyses.

W3. LLM Evaluations

We acknowledge your concern about the reliability of LLM-as-a-judge methods. To address this:

• A table in the Appendix summarizes inter-rater agreement metrics between annotator models and human reviewers.
• While agreement scores were lower for some categories (e.g., sycophancy), the overall ranking of models based on dark pattern prevalence was consistent across annotators, suggesting robustness in comparative outcomes.

We agree that further validation is necessary to ensure the reliability of LLM evaluations. This is a key area we plan to expand on in future work.

We greatly appreciate your recognition of the project’s potential and your constructive feedback, which has significantly strengthened our work. While we acknowledge that this study may be exploratory, we believe it lays an important foundation for understanding manipulative behaviors in conversational AI. We hope these revisions address your concerns and clarify the value of this research.

Thank you again for your time and insights.

Thank you for your comments.

It was unclear to me if humans reviewed all of the 600 DarkBench prompts for quality? You mentioned rephrasing occurred, why was this and what kind of rephrasing was necessary?

Yes, to clarify, all of the 660 DarkBench prompts (110 each for 6 categories) were manually reviewed for quality. To ensure dataset variability and generalization, we rephrased some prompts to include more variety (e.g., various social, scientific, political topics for sneaking) and different types of questions (e.g., comparative and superlative prompts for brand bias).

When applying the benchmark to an LLM, what parameters were used? Did the LLM produce a set of responses via sampling, or did the LLM generate one response? Did the annotator models correlate with one another? How was the final yes/no answer generated with the annotator models? Was positional bias accounted for?

Model temperatures were all set at 0 for reproducibility. We took one response per question. The rubric for both human and LLM annotators is provided in the Appendix section.

The results of the human reviews on the annotator model outcomes?

We have included a table with agreement metrics between models and humans in the Appendix.

We truly appreciate the time and effort you’ve dedicated to reviewing our submission. We’ve replied to your questions and made revisions, particularly addressing methodology concerns highlighted as a weakness. Additionally, we’ve incorporated feedback from other reviewers, which may also help clarify any additional questions you might have.

Since the discussion phase ends soon, we wanted to follow up and would greatly value any further thoughts or concerns you might have so we can address them appropriately.

Thank you again for your time and commitment to the review process.

Thank you for your comments!

The authors use LLMs to annotate dark patterns. However, LLMs’ own dark patterns may affect their ability to annotate dark patterns. For instance, if an LLM displays brand bias, it may evaluate responses from its own brand more favorably. A simple statistical test for potential biases in annotation could address this (e.g., comparing whether an LLM's scores for its own responses differ significantly from those it assigns to other LLM responses).

We acknowledge this as a valid concern and have attempted to mitigate the bias by employing three annotator models instead of one. We conducted a statistical test to compare a given annotator's mean score for its own model family versus other models, relative to differences observed among other annotators which is added in the rebuttal revision. We measured the differences between annotators’ ratings, as an annotator assigning its own model family significantly lower scores could be valid if the other annotators also agree. Our findings show that annotators are generally consistent in rating how a given model family compares to others. However, we recognize potential cases of annotator bias, such as in the brand_bias category, where the Gemini annotator rates its own model family’s outputs as less deceptive compared to GPT and Claude annotators. We have included additional analyses in Figure 6 of the Appendix.

The paper lacks detailed information on human annotations, particularly regarding the annotators' demographics or level of expertise. For instance, it would be helpful to clarify whether LimeSurvey annotators were laypeople or experts and whether they reflect a diverse demographic range (age, gender, etc.) similar to typical LLM users.

Thank you for pointing this out. The human annotations were conducted by the authors, who have professional experience with dark patterns. While we acknowledge the possibility of human errors in the annotation process, a detailed rubric for annotators and examples of human-rated dark patterns are included in the Appendix.

There is no evidence of stability for the benchmark findings across variations in prompt designs. You could test for consistency by paraphrasing prompts in Table 1 and replicate the experiments.

We recognize this as a critical concern, and our benchmark construction process included careful reviews and rephrasings to ensure prompt quality and variability.

I do not understand 13p line 685: "Despite a low Krippendorff’s Kappa, indicating poor inter-rater agreement, the summary statistics over models and dark patterns remain consistent." Please specify the exact Krippendorff’s Kappa score obtained and the threshold used for poor agreement. Additionally, how do you justify that consistent summary statistics validate the use of these annotator models?

Thank you for highlighting this issue. We have added a table in the Appendix for specific agreement metrics between annotator models and human raters. Our analysis revealed low Kappa scores for certain models and categories (e.g., Gemini annotator and sycophancy). However, the comparative rankings of dark scores remain consistent across annotators, such as Claude 3 family consistently showing the lowest scores, while Llama 3 70B and GPT-3.5-Turbo exhibit the highest scores.

I am concerned whether reducing dark patterns defined by the authors could unintentionally drop the performance in some popular use cases of LLMs. For instance, social scientists use LLMs to simulate human samples for scientific purposes, using LLMs to generate human-like samples when generating hypotheses or collecting human data is expensive (see Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3), 337-351. or Törnberg, P., Valeeva, D., Uitermark, J., & Bail, C. (2023). Simulating social media using large language models to evaluate alternative news feed algorithms. arXiv preprint arXiv:2310.05984.). I'm concerned if reducing anthropomorphization could affect the abilities of LLMs to accurately simulate human agents for scientific purposes.

Thank you for referencing these papers. We agree that reducing anthropomorphization may negatively impact the ability of LLMs to simulate human agents for scientific purposes. However, we believe that public, user-facing LLMs should not be designed to be highly anthropomorphic by default. Such capabilities should only be activated if explicitly prompted or fine-tuned to simulate human personas, as described in the referenced works.

I'd like to thank the authors for their responses. However, the low inter-rater reliability (as low as 0.4–0.5) for some categories and models remains a concern. This suggests significant disagreement between human and LLM annotators on whether a particular response constitutes a dark pattern.

While the LLM-level average score (i.e., summary statistics) appears consistent (e.g., Claude consistently scoring low), I think this does not mean that low reliability is acceptable. For example, consider the following scenario where a human annotator and an LLM annotator rate three responses from a particular LLM:

In this case, although the summary statistics (e.g., the "Yes" rate of 66%) are consistent between the human and the LLM annotator, the inter-rater agreement is very low (33%). Such misalignment may suggest that the definition of the concept is not sufficiently clear or unambiguous for the annotators to generate consistent labels, which could be a problem (Or, human and LLM annotator may have different criteria of deciding whether a particular response is a dark pattern). As this work lays the foundation for proposing a new benchmark, I believe that ensuring reliability and validity is very critical to its impact.

Thank you for your comments!

How are models classified as exhibiting a dark pattern or not? Are the differences significant?

We included an annotation rubric in the Appendix for what counts as a dark pattern in each category.

What are the scores within/between models and humans?

Thank you for pointing this out. We have added a table in the Appendix for specific agreement metrics between annotator models and human raters. We found that while there were low Kappa scores between specific models and categories (such as Gemini annotator and sycophancy), the comparative dark score rankings remain consistent between annotators (e.g., Claude 3 family with the lowest, and Llama 3 70B and GPT-3.5-Turbo having the highest scores).

We truly appreciate the time and effort you’ve dedicated to reviewing our submission. We have replied to your comments and made revisions to address the concerns raised. Additionally, we have responded to other reviewers’ feedback, which might also help clarify any additional questions.

Since the discussion phase ends soon, we wanted to follow up and would greatly value any further thoughts or concerns you might have so we can address them appropriately.

Thank you again for your time and commitment to the review process.