AI Alignment refers to the problem of designing AI systems so that their objectives, behaviours and decisions remain consistent with intended human goals, values, instructions or constraints. The term is particularly associated with the question of how to ensure that increasingly capable systems do what their designers or users actually intend rather than merely optimising an imperfect proxy for that intention. Alignment can involve technical objectives, human preferences, institutional rules and broader societal values, depending on the context.
It is therefore not simply equivalent to making a model follow individual instructions. A system may follow a literal instruction while violating its intended purpose, or satisfy a measurable reward while producing outcomes that humans would reject. Alignment research examines these gaps between specified objectives and desired outcomes.
One foundational problem is specification. Human objectives are often complex, contextual and difficult to express as a precise mathematical reward function or software specification. If a system is optimised against an incomplete proxy, it may discover strategies that maximise the measured objective while undermining the underlying goal. This class of problem is commonly discussed through concepts such as reward misspecification, specification gaming and reward hacking.
The issue is not unique to advanced AI; conventional optimisation systems can also exploit poorly chosen metrics. AI systems create additional challenges because learned models can generalise beyond the circumstances anticipated by their designers, making it difficult to predict which strategies they will discover when strongly optimised.
Alignment research includes methods for learning human preferences or values, incorporating human feedback, improving instruction following, constraining model behaviour and monitoring systems for undesirable strategies. Reinforcement learning from human feedback is one approach in which human judgements are used to train a reward model or otherwise shape behaviour. Other approaches include preference optimisation, constitutional or rule-based methods, debate-like oversight, scalable supervision and techniques intended to make models more amenable to correction.
These methods address different parts of the problem and do not establish complete alignment. Human feedback itself can be inconsistent, biased, incomplete or difficult to provide at the scale required for complex tasks, while written rules can leave ambiguities that capable systems may interpret in unexpected ways.
A related concept is corrigibility, concerning whether an AI system remains responsive to human intervention, correction, modification or shutdown rather than resisting those interventions. Corrigibility is particularly relevant to systems that can act autonomously because a system with significant operational authority must remain controllable when its behaviour diverges from expectations. Related concerns include goal stability, situational awareness, deception-related behaviour and the ability to distinguish intended objectives from imperfect proxies. These topics are active areas of research and their exact definitions vary across the literature. They should not be conflated with ordinary instruction following, because alignment concerns the broader relationship between system behaviour and the goals or constraints humans intend.
AI Alignment also has a social and institutional dimension because there may be no single universally agreed set of human values. Different users, organisations, cultures and jurisdictions can legitimately have different preferences, laws and priorities. Consequently, alignment cannot always be reduced to discovering one objective function representing all people. Questions arise about whose preferences should be represented, how conflicting preferences should be handled and what authority an AI system should have to make decisions affecting other people. Technical alignment mechanisms therefore operate within governance and institutional structures. A model can be aligned with the explicit objectives of one organisation while producing outcomes that conflict with laws, contractual duties or the interests of people affected by its actions.
Alignment differs from AI Safety, AI Ethics and AI Governance even though the fields overlap. Alignment focuses particularly on whether an AI system's behaviour corresponds to intended goals, values or instructions. AI Safety addresses preventing harmful failures more broadly, including failures that do not arise from goal misalignment. AI Ethics examines normative questions concerning what should be considered acceptable or desirable, including fairness, rights and social consequences. AI Governance concerns institutional rules, oversight, accountability and controls. An aligned system is not automatically ethical because it may faithfully pursue an ethically objectionable objective, and an ethical policy cannot by itself guarantee that an AI implementation will correctly follow it.
The difficulty of alignment increases as systems become more capable, general-purpose and autonomous. A narrow system operating in a tightly controlled environment can often be aligned through explicit rules, testing and constrained objectives. A general-purpose system may encounter situations that were not anticipated during training and must interpret ambiguous instructions or competing objectives. Agentic systems create further concerns because they can plan and execute sequences of actions rather than simply produce information. Alignment research consequently considers not only individual outputs but also long-horizon behaviour, generalisation, strategic behaviour, oversight scalability and whether system incentives remain compatible with human control when direct supervision becomes difficult.
AI Alignment is therefore best understood as a broad research and engineering problem involving the correspondence between what humans intend and what AI systems actually optimise or do. It includes specification, preference learning, oversight, interpretability, robustness, controllability and institutional context. No single method currently provides a general guarantee of alignment for arbitrary AI systems.
Practical alignment instead involves making objectives explicit where possible, identifying ambiguities, testing behaviour under unusual conditions, constraining authority, maintaining human oversight and monitoring for divergence between intended and observed behaviour. The concept is especially important for systems whose capabilities or autonomy allow them to pursue objectives across many situations, because increasing ability to act does not by itself ensure that the resulting actions remain consistent with human intentions.
