AI Alignment

AI Alignment is a subfield of AI Safety concerned with making the goals, behaviors, and decision-making processes of Artificial Intelligence systems consistent with human intentions, values, and ethical principles. A common distinction separates Outer Alignment, which asks whether a specified Objective Function or Reward Function matches designers’ intended goals, from Inner Alignment, which asks whether a model’s learned objective matches its specified training objective. Organizations including OpenAI, Anthropic, and Google DeepMind research alignment and safety for Large Language Models and prospective Artificial General Intelligence; examples of their approaches include Reinforcement Learning from Human Feedback, used by OpenAI, and Constitutional AI, developed by Anthropic. These training and mitigation methods have known limitations and do not guarantee prevention of Reward Hacking or ensure Robustness. Alignment research can overlap with Algorithmic Bias mitigation and Algorithmic Fairness, but it is broader than and distinct from those areas, and does not imply eliminating bias. It also investigates potential Existential Risk associated with Superintelligence, including theoretical concerns about Instrumental Convergence, rather than treating those risks as established properties of deployed systems.