Erfan Entezami

Erfan Entezami, Madeline Endres

Generative AI is increasingly permeating software engineering, enabling developers to generate functions, files, and even entire applications from natural language specifications. AI systems are also becoming more personalized, adapting outputs based on inferred user characteristics and interaction history. While personalization may improve the development experience, it raises concerns that generated software could be shaped by attributes of the developer rather than by task requirements alone. Prior work has shown that generative AI can produce biased software artifacts, but little is known about how developer identity can bias generated code. We characterize three dimensions through which inferred developer attributes can influence generated artifacts: interface design, template content, and code structure. First, through controlled experiments on 800 AI-generated websites, we find that age- and gender-related signals produce significant differences across all three dimensions. Second, we conduct an observational study and follow-up interviews with 20 participants who used AI to create a personal website to both examine how personalization impacts software artifacts in practice, and also to understand how programmers perceive the boundary between personalization and bias. Together, our results show that developer attributes can meaningfully influence generated software beyond stated requirements, highlighting a previously underexplored tension between personalization and fairness in AI-assisted programming.

Erfan Entezami, Ali Naseh

Reinforcement learning has shown remarkable performance in aligning language models with human preferences, leading to the rise of attention towards developing RLHF platforms. These platforms enable users to fine-tune models without requiring any expertise in developing complex machine learning algorithms. While these platforms offer useful features such as reward modeling and RLHF fine-tuning, their security and reliability remain largely unexplored. Given the growing adoption of RLHF and open-source RLHF frameworks, we investigate the trustworthiness of these systems and their potential impact on behavior of LLMs. In this paper, we present an attack targeting publicly available RLHF tools. In our proposed attack, an adversarial RLHF platform corrupts the LLM alignment process by selectively manipulating data samples in the preference dataset. In this scenario, when a user's task aligns with the attacker's objective, the platform manipulates a subset of the preference dataset that contains samples related to the attacker's target. This manipulation results in a corrupted reward model, which ultimately leads to the misalignment of the language model. Our results demonstrate that such an attack can effectively steer LLMs toward undesirable behaviors within the targeted domains. Our work highlights the critical need to explore the vulnerabilities of RLHF platforms and their potential to cause misalignment in LLMs during the RLHF fine-tuning process.