Strong and weak alignment of large language models with human values
Mehdi Khamassi, Marceau Nahon, Raja Chatila
TL;DR
The paper tackles the AI value alignment problem by distinguishing strong alignment, which entails semantic understanding of human values, reasoning about agents' intentions, and internal causal models of action effects, from weak alignment, which relies on statistical matching to desired outputs. It presents seven value-focused prompts evaluated on ChatGPT, Gemini, and Copilot, illustrating that current systems often fail to recognize implicit value threats and can exhibit non-repeatable, fallacious reasoning. Complementary analyses of word embeddings (LSA, Word2Vec, GPT-4) for dignity, fairness, and well-being reveal semantic gaps between human concepts and model representations. A thought experiment, the Chinese room with a word transition dictionary, is proposed to isolate the cognitive capabilities needed for strong alignment. Together, these contributions argue for pursuing strong alignment to improve explainability, controllability, and ethical trust, while acknowledging practical challenges and the role of prompts and filters in shaping current behavior.
Abstract
Minimizing negative impacts of Artificial Intelligent (AI) systems on human societies without human supervision requires them to be able to align with human values. However, most current work only addresses this issue from a technical point of view, e.g., improving current methods relying on reinforcement learning from human feedback, neglecting what it means and is required for alignment to occur. Here, we propose to distinguish strong and weak value alignment. Strong alignment requires cognitive abilities (either human-like or different from humans) such as understanding and reasoning about agents' intentions and their ability to causally produce desired effects. We argue that this is required for AI systems like large language models (LLMs) to be able to recognize situations presenting a risk that human values may be flouted. To illustrate this distinction, we present a series of prompts showing ChatGPT's, Gemini's and Copilot's failures to recognize some of these situations. We moreover analyze word embeddings to show that the nearest neighbors of some human values in LLMs differ from humans' semantic representations. We then propose a new thought experiment that we call "the Chinese room with a word transition dictionary", in extension of John Searle's famous proposal. We finally mention current promising research directions towards a weak alignment, which could produce statistically satisfying answers in a number of common situations, however so far without ensuring any truth value.