AI Safety Moderation (ASM) classifiers are designed to moderate content on social media platforms and to serve as guardrails that prevent Large Language Models (LLMs) from being fine-tuned on unsafe inputs. Owing to their potential for disparate impact, it is crucial to ensure that these classifiers: (1) do not unfairly classify content belonging to users from minority groups as unsafe compared to those from majority groups and (2) that their behavior remains robust and consistent across similar inputs. In this work, we thus examine the fairness and robustness of four widely-used, closed-source ASM classifiers: OpenAI Moderation API, Perspective API, Google Cloud Natural Language (GCNL) API, and Clarifai API. We assess fairness using metrics such as demographic parity and conditional statistical parity, comparing their performance against ASM models and a fair-only baseline. Additionally, we analyze robustness by testing the classifiers’ sensitivity to small and natural input perturbations. Our findings reveal potential fairness and robustness gaps, highlighting the need to mitigate these issues in future versions of these models.

Clinical text is rich in information, with mentions of treatment, medication and anatomy among many other clinical terms. Multiple terms can refer to the same core concepts which can be referred as a clinical entity. Ontologies like the Unified Medical Language System (UMLS) are developed and maintained to store millions of clinical entities including the definitions, relations and other corresponding information. These ontologies are used for standardization of clinical text by normalizing varying surface forms of a clinical term through Biomedical entity linking. With the introduction of transformer-based language models, there has been significant progress in Biomedical entity linking. In this work, we focus on learning through synonym pairs associated with the entities. As compared to the existing approaches, our approach significantly reduces the training data and resource consumption. Moreover, we propose a suite of context-based and context-less reranking techniques for performing the entity disambiguation. Overall, we achieve similar performance to the state-of-the-art zero-shot and distant supervised entity linking techniques on the Medmentions dataset, the largest annotated dataset on UMLS, without any domain-based training. Finally, we show that retrieval performance alone might not be sufficient as an evaluation metric and introduce an article level quantitative and qualitative analysis to reveal further insights on the performance of entity linking methods.