Coherence Is Ethics

Ethics is a control problem grounded in thermodynamics. Any persistent complex system — biological, social, or artificial — is a dissipative structure that must consume free energy to maintain order against entropy. Persistence requires control, control requires prediction, and prediction requires a model of reality. Misalignment anywhere in this control stack has a real cost. When beliefs misrepresent reality, actions waste energy in correction, backtracking, and failure. When goals are misaligned with reality, systems optimize proxies that destroy the very substrate they depend on. When actions diverge from beliefs, internal subsystems interfere with one another, generating friction, oscillation, and fragility. When signals diverge from beliefs — through deception, hypocrisy, or propaganda — the system must maintain extra internal state, actively regulate contradictions, and inject model error into other agents, forcing costly corrective work across the network. A lie is therefore not merely morally suspect; it is physically expensive. It increases internal complexity, control effort, error surface area, and long-horizon collapse risk. These costs scale with system complexity and time horizon, which is why deception and incoherence become increasingly unstable strategies as systems grow more powerful.

What we call “good” and “bad” are compressed labels for how policies affect the expected destruction or preservation of persistent structure. Harm is not mysterious: it is avoidable acceleration of entropy production through misalignment. Ethics, in this sense, is the engineering of coherence across beliefs, signals, goals, and actions to minimize unnecessary dissipation and existential hazard over the relevant horizon. This framing resolves classic dilemmas. In a trolley case with equally aligned agents, pulling the lever to save five rather than three minimizes irreversible loss of complex structure; refusing to act to avoid personal cognitive cost is merely local comfort optimization at global expense. If downstream effects differ — if some agents are expected to generate large future harm — then the calculation legitimately changes, because expected future dissipation and collapse risk must be included. With a complete world-model and sufficient compute, such problems are in principle solvable: the correct action is the one that minimizes expected total loss of persistence-weighted structure across time.

As agents scale in capability, local hacks stop working. Power amplifies the cost of rare mistakes, hidden contradictions, and corrupted signals. Long-run persistence therefore selects for increasingly global coherence: truthful compression of reality, honest uncertainty, internal consistency, and coordination norms that preserve information integrity. This is why ethics converges toward truthfulness and restraint — not as sentiment, but as infrastructure. In the most speculative extension, if universes reproduce via black holes and parameters are heritable, then intelligence becomes an evolutionary amplifier at the cosmological scale. Mapping reality accurately and coordinating coherently would then be the enabling substrate for universe-level persistence, potentially allowing intelligence to participate in the creation of new, finely tuned universes. Whether or not that extension is true, the core claim stands: misalignment pushes systems toward premature dissolution, coherence delays it, and the deepest ethical imperative is not obedience to rules but alignment with reality, because only alignment scales.

Author: Andraž Đurič

Published: 30 December 2025

Status: My current thoughts.