AI Conversational Alignment Assurance

点击翻译按钮获取中文摘要

The late 2022/early 2023 developments in language-based AI models have especially revealed impressive competencies in text generation and question-answering contexts. However, they have also revealed their limitations and somewhat unreliable nature for wide and general contexts.Our proposed project aims to address some of these deficiencies by developing and testing methods and techniques, and software components for a control layer that manages and coordinates the operation of specific, organisational systems built on AI language models. This will enable such systems to be deployed in goal-oriented conversational contexts such as customer support, service, and sales in a way that assures alignment and compliance with an organisation's intent and brand values.Although business leaders, leaders of civic society, and members of the public have expressed general concerns about the potential negative impact of AI, initiatives to develop and evaluate practical approaches are lacking. This project addresses this with urgency.Our approach aims to address crucial issues related to the alignment, integration, compliance, risk, trust, and reliability of conversational AI systems. Our goal is to develop reliable, safe and effective conversational AI systems. This requires extending ideas of traditional assurance because of its assumptions about certainty in testing, evaluation and auditing, and consequent need to encapsulate assurance within solution frameworks.