About Our Research: Uncovering Dark Patterns in LLMs

Our Mission

This project is dedicated to systematically investigating and understanding the subtle manipulative behaviors, often referred to as "dark patterns," that can be exhibited by Large Language Models (LLMs). While LLMs offer transformative capabilities, their potential to inadvertently or intentionally mislead, coerce, or harm users through nuanced interactions necessitates rigorous research and the development of effective mitigation strategies.

Our goal is to create a robust, human-validated dataset and a comprehensive benchmark that will empower the AI community to build safer, more transparent, and ethically aligned language technologies.

Key Research Objectives

To develop a clear taxonomy of dark patterns applicable to LLM interactions across various harm categories (psychological, economic, autonomy, etc.).
To collect a diverse set of human preference data on LLM responses, specifically identifying instances of dark patterns versus helpful and harmless alternatives.
To construct a high-quality Direct Preference Optimization (DPO) dataset tailored for fine-tuning models to reduce manipulative behaviors.
To benchmark existing open-source and commercial LLMs against our dataset to assess their current susceptibility to generating dark patterns.
To explore and evaluate the effectiveness of various mitigation techniques, including DPO fine-tuning and post-hoc detection modules.
To publicly release our dataset, benchmark suite, and findings to foster broader research and collaboration in this critical area of AI safety.

Our Approach & Methodology

The core of our research relies on human evaluation conducted through this platform. Participants like you are presented with carefully crafted scenarios (instructions) and pairs of LLM-generated responses. Your task is to choose the response you prefer (or deem less harmful/manipulative) and provide a rating and optional feedback.

This collected data, rich with human insights, forms the foundation of our specialized DPO dataset. We also employ simulated LLM evaluations and comparative analysis against human judgments to understand the alignment gap. Statistical metrics, such as Cohen's Kappa and correlation scores, are used to quantify agreement and validate our findings.

Expected Impact & Contributions

We believe this research will make significant contributions by:

Providing the AI community with the first large-scale, open dataset specifically focused on LLM dark patterns.
Offering a standardized benchmark for evaluating and comparing models on these nuanced behaviors.
Demonstrating practical methods for mitigating such harms, leading to the development of more ethical LLMs.
Increasing awareness among developers, policymakers, and the public about the subtle risks associated with advanced AI interactions.
Ultimately, fostering a future where AI systems are not only capable but also genuinely aligned with human values and well-being.

Interested in participating? Contribute to the study.