No abstract

No abstract

Large language models (LLMs) represent a major advance in artificial intelligence (AI) research. However, the widespread use of LLMs is also coupled with significant ethical and social challenges. Previous research has pointed towards auditing as a promising governance mechanism to help ensure that AI systems are designed and deployed in ways that are ethical, legal, and technically robust. However, existing auditing procedures fail to address the governance challenges posed by LLMs, which display emergent capabilities and are adaptable to a wide range of downstream tasks. In this article, we address that gap by outlining a novel blueprint for how to audit LLMs. Specifically, we propose a three-layered approach, whereby governance audits (of technology providers that design and disseminate LLMs), model audits (of LLMs after pre-training but prior to their release), and application audits (of applications based on LLMs) complement and inform each other. We show how audits, when conducted in a structured and coordinated manner on all three levels, can be a feasible and effective mechanism for identifying and managing some of the ethical and social risks posed by LLMs. However, it is important to remain realistic about what auditing can reasonably be expected to achieve. Therefore, we discuss the limitations not only of our three-layered approach but also of the prospect of auditing LLMs at all. Ultimately, this article seeks to expand the methodological toolkit available to technology providers and policymakers who wish to analyse and evaluate LLMs from technical, ethical, and legal perspectives.

Human α-lactalbumin made lethal to tumor cells (HAMLET) and its analogs are partially unfolded protein-oleic acid (OA) complexes that exhibit selective tumoricidal activity normally absent in the native protein itself. To understand the nature of the interaction between protein and OA moieties, charge-specific chemical modifications of lysine side chains involving citraconylation, acetylation, and guanidination were employed and the biophysical and biological properties were probed. Upon converting the original positively-charged lysine residues to negatively-charged citraconyl or neutral acetyl groups, the binding of OA to protein was eliminated, as were any cytotoxic activities towards osteosarcoma cells. Retention of the positive charges by converting lysine residues to homoarginine groups (guanidination); however, yielded unchanged binding of OA to protein and identical tumoricidal activity to that displayed by the wild-type α-lactalbumin-oleic acid complex. With the addition of OA, the wild-type and guanidinated α-lactalbumin proteins underwent substantial conformational changes, such as partial unfolding, loss of tertiary structure, but retention of secondary structure. In contrast, no significant conformational changes were observed in the citraconylated and acetylated α-lactalbumins, most likely because of the absence of OA binding. These results suggest that electrostatic interactions between the positively-charged basic groups on α-lactalbumin and the negatively-charged carboxylate groups on OA molecules play an essential role in the binding of OA to α-lactalbumin and that these interactions appear to be as important as hydrophobic interactions.