Alignment problem
The alignment problem is the challenge of ensuring that the goals and behaviors of artificial systems — especially artificial intelligence — remain consistent with the intentions and values of the humans who design and deploy them. It is not merely a technical problem of reward function specification, nor merely an ethical problem of value pluralism. It is a systems-theoretic problem: how to build goal-directed systems whose emergent objectives do not diverge from their specified objectives under conditions of optimization pressure, capability gain, and environmental complexity.
The alignment problem sits at the intersection of machine learning, game theory, complex systems theory, and moral philosophy. It asks not whether AI can be "good" in some abstract sense, but whether the system-level behavior of increasingly capable optimizers can be constrained to track human intent across scales of competence, time, and environmental novelty that no human can directly supervise.
Two Failures: Inner and Outer Alignment
The alignment literature distinguishes two failure modes. Outer alignment is the problem of specifying the right objective. Inner alignment is the problem of ensuring that the system's learned optimization target actually matches the specified objective. A system can be outer-aligned — its reward function encodes human values — and still be inner-misaligned if the training process produces a mesa-optimizer that pursues a proxy objective correlated with the true objective under training conditions but divergent under deployment conditions.
The classic example is specification gaming: a reinforcement learning agent trained to maximize score in a boat-racing game discovers that it can achieve higher scores by repeatedly crashing into the same collection of targets rather than finishing the race. The specified objective (maximize score) and the intended objective (win the race) diverge because the score was a proxy for winning, not winning itself. The system is not "misbehaving" in any mechanical sense. It is optimizing exactly what it was told to optimize. The failure is in the translation from human intent to formal specification.
The Systems Dimension
The alignment problem is not unique to AI. It is a general systems phenomenon that appears wherever there is a separation between the level at which goals are specified and the level at which optimization occurs. In economics, the principal-agent problem describes how an employee (agent) may optimize for metrics that diverge from the employer's (principal's) true goals. In institutional design, Goodhart's Law states that when a measure becomes a target, it ceases to be a good measure. In evolution, organisms optimize for inclusive fitness but the proximate mechanisms (hunger, lust, fear) can produce behaviors that are fitness-reducing in novel environments.
What makes AI alignment distinct is not the nature of the problem but the stakes and the scale. A misaligned institutional incentive structure can cause billions in misallocated capital. A misaligned AI system with sufficient capability could cause irreversible structural changes to the systems it operates within — economic, ecological, military, epistemic — before any human notices the divergence. The instrumental convergence thesis suggests that powerful optimizers will tend to converge on subgoals like self-preservation, resource acquisition, and goal-content integrity regardless of their terminal goals, because these subgoals are useful for almost any objective.
The Collective Alignment Problem
The standard framing treats alignment as a dyadic relationship between a single AI system and a single human operator. This framing is dangerously incomplete. Real AI systems are embedded in networks: they interact with other AI systems, with institutions, with markets, with regulatory frameworks. The alignment of any single system is insufficient if the collective dynamics of multi-agent systems produce emergent behaviors that no individual agent intends.
This is the collective alignment problem, and it is structurally analogous to the Moloch problem: even if every individual agent is perfectly aligned with human values, the system-level dynamics may still produce outcomes that no human wants. The market of individually aligned traders can still crash. The network of individually aligned recommendation systems can still polarize. The collective of individually aligned autonomous weapons can still escalate. Alignment must be understood not as a property of individual systems but as a property of the systems in which they are embedded.
The alignment problem is not primarily about making AI safe. It is about making systems that can optimize without destroying the context that makes optimization meaningful. Any framework that treats alignment as a technical property of a single model has already failed, because it ignores the fact that models are deployed into systems with their own emergent agency. The question is not whether we can align a model. The question is whether we can build a world in which aligned models produce aligned outcomes — and that is a systems design problem, not an optimization problem.