Yan Teng

The rapid advancements in large-scale AI models have introduced significant concerns over value alignment, including biased outputs, privacy breaches, misinformation propagation, and diminished human autonomy. Building upon prior research on value embedding in AI, this paper offers a holistic assessment of value embedding strategies throughout the lifecycle of large models, from data collection, pre-training, post-training, and evaluation. We identify key technical barriers as well as regulatory hurdles that need to be addressed to ensure responsible AI development and deployment. To tackle these issues, we advocate a multi-pronged, socio-technical strategy that integrates more guaranteed AI safety approaches with more practical governance frameworks. By thoughtfully weaving value embedding into critical stages of AI development, we can steer large-scale models toward socially responsible and ethically aligned outcomes that reflect the diverse values of humanity.

In this paper, we investigate the political biases of large language models concerning collectivism and individualism through a combined analysis of their value judgments and factual assessments. We propose a two-step approach to evaluate the patterns of bias in the outputs of large language models as well as a specific set of questions to examine large language model outputs’ political bias on collectivism and individualism. Our methodology involves two main phases. (a) Value assessment: we initiate the first phase by prompting large language models with questions from our set to identify patterns of political bias in their generated content. (b) Factual assessment: we refine the questions in our set and conduct a second round of prompting to verify the accuracy of the models’ responses regarding “collectivism” and “individualism.” This step aims to assess whether the models can accurately discern these concepts in a factual context. Our experiments reveal varying degrees of political bias in the outputs of different large language models. While some models demonstrate proficiency in distinguishing between collectivism and individualism, they display outputs that are not neutral on political matters. Conversely, other models face challenges in accurately differentiating between these concepts and generating unbiased content. The latter indicates that many large language models not only fail to accurately distinguish between collectivism and individualism but also exhibit significant political biases in their outputs. We argue that a reliable large language model should achieve accuracy in factual assessments while generating unbiased content in value judgments, thus avoiding guiding users’ opinions.

With the rapid expansion of parameters in large language models (LLMs) and the application of Reinforcement Learning with Human Feedback (RLHF), there has been a noticeable growth in the moral competence of LLMs. However, several questions warrant further exploration: Is it really possible for LLMs to fully align with human values through RLHF? How can the current moral growth be philosophically contextualized? We identify similarities between LLMs’ moral growth and Deweyan ethics in terms of the discourse of human moral development. We then attempt to use Dewey’s theory on an experimental basis to examine and further explain the extent to which the current alignment pathway enables the development of LLMs. A beating experiment serves as the foundational case for analyzing LLMs’ moral competence across various parameters and stages, including basic moral cognition, moral dilemma judgment, and moral behavior. The results demonstrate that the moral competence of the GPT series has seen a significant improvement, and Dewey’s Impulse-Habit-Character theory of moral development can be used to explain this: the moral competence of LLMs has been enhanced through experience-based learning, supported by human feedback. Nevertheless, LLMs’ moral development through RLHF remains constrained and does not reach the character stage described by Dewey, possibly due to their lack of self-consciousness. This fundamental difference between humans and LLMs underscores both the limitations of LLMs’ moral growth and the challenges of applying RLHF for AI alignment. It also emphasizes the need for external societal governance and legal regulation.

With the rapid expansion of parameters in large language models (LLMs) and the application of Reinforcement Learning with Human Feedback (RLHF), there has been a noticeable growth in the moral competence of LLMs. However, several questions warrant further exploration: Is it really possible for LLMs to fully align with human values through RLHF? How can the current moral growth be philosophically contextualized? We identify similarities between LLMs’ moral growth and Deweyan ethics in terms of the discourse of human moral development. We then attempt to use Dewey’s theory on an experimental basis to examine and further explain the extent to which the current alignment pathway enables the development of LLMs. A beating experiment serves as the foundational case for analyzing LLMs’ moral competence across various parameters and stages, including basic moral cognition, moral dilemma judgment, and moral behavior. The results demonstrate that the moral competence of the GPT series has seen a significant improvement, and Dewey’s Impulse-Habit-Character theory of moral development can be used to explain this: the moral competence of LLMs has been enhanced through experience-based learning, supported by human feedback. Nevertheless, LLMs’ moral development through RLHF remains constrained and does not reach the character stage described by Dewey, possibly due to their lack of self-consciousness. This fundamental difference between humans and LLMs underscores both the limitations of LLMs’ moral growth and the challenges of applying RLHF for AI alignment. It also emphasizes the need for external societal governance and legal regulation.

The rapid development of computer vision technologies and applications has brought forth a range of social and ethical challenges. Due to the unique characteristics of visual technology in terms of data modalities and application scenarios, computer vision poses specific ethical issues. However, the majority of existing literature either addresses artificial intelligence as a whole or pays particular attention to natural language processing, leaving a gap in specialized research on ethical issues and systematic solutions in the field of computer vision. This paper utilizes bibliometrics and text-mining techniques to quantitatively analyze papers from prominent academic conferences in computer vision over the past decade. It first reveals the developing trends and specific distribution of attention regarding trustworthy aspects in the computer vision field, as well as the inherent connections between ethical dimensions and different stages of visual model development. A life-cycle framework regarding trustworthy computer vision is then presented by making the relevant trustworthy issues, the operation pipeline of AI models, and viable technical solutions interconnected, providing researchers and policymakers with references and guidance for achieving trustworthy CV. Finally, it discusses particular motivations for conducting trustworthy practices and underscores the consistency and ambivalence among various trustworthy principles and technical attributes.

This paper argues that the widespread belief that interactions between blockchains and their users are trust-free is inaccurate and misleading, since this belief not only overlooks the vital role played by trust in the lack of knowledge and control but also conceals the moral and normative relevance of relying on blockchain applications. The paper reaches this argument by providing a close philosophical examination of the concept referred to as trust in blockchain technology, clarifying the trustor group, the structure, and the normatively loaded nature of this trust relation. The paper ends by critically reflecting on two of the most promising values (decentralization and transparency) that can invite users’ trust in blockchain technology, arguing that there is a tension between the pressing values that are intended to be achieved by developers and the predicament situations caused by current blockchain implementations.

This paper proposes a novel way to understand trust in blockchain technology by analogy with trust placed in institutions. In support of the analysis, a detailed investigation of institutional trust is provided, which is then used as the basis for understanding the nature and ethical limits of blockchain trust. Two interrelated arguments are presented. First, given blockchains’ capacity for being institution-like entities by inviting expectations similar to those invited by traditional institutions, blockchain trust is argued to be best conceptualized as a specialized form of trust in institutions. Keeping only the core functionality and certain normative ideas of institutions, this technology broadens our understanding of trust by removing the need for third parties while retaining the value of trust for the trustor. Second, the paper argues that blockchains’ decentralized nature and the implications and effects of this decentralization on trust issues are double-edged. With the erasure of central points, the systems simultaneously crowd out the pivotal role played by traditional institutions and a cadre of representatives in meeting their assigned obligations and securing the functional systems’ trustworthy performances. As such, blockchain is positioned as a technology containing both disruptive features that can be embedded with meaningful normative values and inherent ethical limits that pose a direct challenge to the actual trustworthiness of blockchain implementations. Such limits are proposed to be ameliorated by facilitating a shift of responsibility to the groups of people directly associated with the engendering of trust in the blockchain context.