Alignment Is Physical Alignment

Transfers

To “align” an AI system with human values is to bring two things into parallel, like aligning wheels on an axle or magnets in a field. The physical metaphor frames values as fixed directions and AI behavior as a vector that can be measured, adjusted, and verified against a reference. This is the governing metaphor of AI safety, and it shapes billions of dollars in research priorities.

Key structural parallels:

• Values as directions — alignment presupposes that human values have a definite orientation, like compass needles pointing north. The metaphor makes it natural to ask “are we aligned?” as if values were measurable quantities with a correct answer. This frames the safety problem as an engineering challenge: find the right direction, point the AI at it.
• Misalignment as angular deviation — when an AI system does not behave as intended, it is “misaligned,” like a wheel pulling to one side. The metaphor quantifies the problem: misalignment has a magnitude (how far off) and can be corrected incrementally. This is deeply optimistic — it suggests the problem is tractable and convergent.
• Adjustment as calibration — RLHF (reinforcement learning from human feedback) is framed as an alignment procedure: human preferences provide the reference direction, and the model is adjusted to match. The metaphor imports the precision and repeatability of mechanical calibration into a domain of contested, context-dependent values.
• The alignment target as singular — physical alignment assumes one correct orientation. The metaphor makes it natural to speak of “human values” as a unified target, obscuring the fact that humans disagree profoundly about what those values are. Whose values? Which culture? What era? The metaphor elides these questions by assuming a fixed reference direction.

Limits

• Human values are not vectors — this is the fundamental break. Physical alignment requires fixed, measurable directions. Human values are contested, contextual, contradictory, and evolving. A person can simultaneously value freedom and security, honesty and kindness, individual rights and collective welfare. These are not directions in a space; they are a shifting, multi-dimensional tangle that resists reduction to a single orientation.
• There is no reference frame — physical alignment requires a fixed reference: true north, a plumb line, a laser level. Whose values serve as the reference for AI alignment? The metaphor assumes a consensus that does not exist. Choosing a reference is itself a political act that the metaphor disguises as a technical one.
• Alignment is not a binary — physical alignment is either achieved or not (within tolerances). But the relationship between AI behavior and human values is not binary or even continuous in the same way. An AI system can be “aligned” with one stakeholder’s values while being “misaligned” with another’s. The metaphor has no vocabulary for this multiplicity.
• The metaphor implies static maintenance — once wheels are aligned, they stay aligned until something knocks them out. The metaphor suggests that alignment is a state to be achieved and maintained, not an ongoing negotiation. But values change, contexts shift, and what counts as aligned behavior in one situation may be harmful in another. Alignment is not a destination; it is a continuous process with no equilibrium.
• Physical alignment is value-neutral — aligning wheels is not a moral act. The metaphor strips the ethical weight from AI safety by framing it as mechanical adjustment. This makes it easier to treat alignment as a purely technical problem, solvable by engineers, rather than a political and philosophical problem requiring broad democratic participation.

Expressions

• “AI alignment” — the canonical term, now an entire research field named after the physical metaphor
• “Misaligned AI” — an AI system whose behavior deviates from intended values, like a crooked wheel
• “Aligning the model with human preferences” — the RLHF framing, mapping calibration onto value adjustment
• “Value alignment problem” — the research question, framed as a problem of bringing two things into parallel
• “Alignment tax” — the performance cost of constraining AI behavior, mapping the friction of mechanical alignment onto capability reduction
• “Superalignment” — OpenAI’s term for aligning superintelligent systems, intensifying the physical metaphor with a prefix that suggests extreme precision

Origin Story

The term “alignment” in AI safety traces to Eliezer Yudkowsky and the Machine Intelligence Research Institute (MIRI) in the early 2000s, where it described the challenge of ensuring that an artificial general intelligence pursues goals compatible with human welfare. The physical metaphor was likely inherited from the broader use of “alignment” in organizational theory (strategic alignment) and philosophy (moral alignment), both of which draw on the same physical image of bringing things into parallel. Stuart Russell’s Human Compatible (2019) popularized the concept for a broad audience, framing it explicitly as ensuring AI objectives are “aligned” with human preferences. The establishment of dedicated alignment research labs (Anthropic, OpenAI’s Superalignment team, DeepMind’s alignment team) institutionalized the metaphor as the name of the field itself. The term is now so entrenched that questioning the metaphor — asking whether “alignment” is the right frame for the problem — has become a distinct line of critique in AI ethics.

References

• Russell, S. Human Compatible: Artificial Intelligence and the Problem of Control (2019) — popularizes the alignment framing for general audiences
• Yudkowsky, E. “Artificial Intelligence as a Positive and Negative Factor in Global Risk” (2008) — early formalization of the alignment problem
• Gabriel, I. “Artificial Intelligence, Values, and Alignment” (2020) — Minds and Machines, analyzes whose values alignment targets
• Maas, M. “AI is Like…” (2023) — places alignment within broader AI metaphor landscape