The superior functionality and versatility of generative AI have raised expectations for the improvement of human society and concerns about the ethical and social risks associated with the use of generative AI. Many previous studies have presented risk issues as concerns associated with the use of generative AI, but since most of these concerns are from the user's perspective, they are difficult to lead to specific countermeasures. In this study, the risk issues presented by the previous studies were broken down into more detailed elements, and risk factors and impacts were identified. In this way, we presented information that leads to countermeasure proposals for generative AI risks.

Recently, various risks have been pointed out in artificial intelligence (AI) systems. In particular, AI security, AI ethics, and AI quality are considerable risks. To make AI systems trustworthy against these risks, risk assessment technology is needed to identify potential AI risks and decide which risks should be dealt with in priority. We propose a risk assessment technology that assesses three kinds of risks—AI security, AI ethics, and AI quality—which have been considered separately. Our technology follows the ISO 31000 framework, consisting of four phases: risk candidate identification, impact assessment, likelihood assessment, and priority derivation for countermeasures. To realize this technology, risk candidate identification is achieved by extending AI ethics impact assessment— an identification method of AI ethics risk—to AI security and AI quality. Impact and likelihood assessments are conducted by extending assessment methods for AI security to AI ethics and AI quality. We conducted a case study using our technology and confirmed that the risks were appropriately extracted, and the priority of the risks to be dealt with was derived.

Unsupervised domain adaptation (UDA) is a challenging task characterized by unlabeled target data with domain discrepancy to labeled source data. Many methods have been proposed to learn domain invariant features by marginal distribution alignment, but they ignore the intrinsic structure within target domain, which may lead to insufficient or false alignment. Class-level alignment has been demonstrated to align the features of the same class between source and target domains. These methods rely extensively on the accuracy of predicted pseudo-labels for target data. Here, we develop a novel self-training method that focuses more on accurate pseudo-labels via a dual-consistency strategy involving modelling the intrinsic structure of the target domain. The proposed dual-consistency strategy first improves the accuracy of pseudo-labels through voting consistency, and then reduces the negative effects of incorrect predictions through structure consistency with the relationship of intrinsic structures across domains. Our method has achieved comparable performance to the state-of-the-arts on three standard UDA benchmarks.