Recursive degradation

Recursive degradation is the progressive loss of quality or information that occurs when a system's outputs are fed back into its inputs across multiple iterations. Unlike simple decay, which is uniform over time, recursive degradation is self-reinforcing: each iteration compounds the errors, biases, or simplifications of the previous one. The result is not a linear decline but a convergent collapse toward a degenerate fixed point.

The phenomenon is most visible in model collapse, where generative models are trained on synthetic data produced by earlier models. But recursive degradation is a general systems phenomenon. It occurs in any system where the signal path forms a closed loop and the loop lacks error correction. Audio systems with feedback produce squeals; information systems with feedback produce lies; cognitive systems with feedback produce delusions.

The Recursive Structure

The recursive structure is simple: Output(t) = f(Input(t)), where Input(t+1) = Output(t) + noise. If f preserves all information in Input(t), the system is stable. If f loses information — through approximation, bias, or compression — the information loss compounds. After n iterations, the information content is approximately I_0 * (1 - epsilon)^n, where epsilon is the per-iteration loss rate. For small epsilon, the decay appears gradual. But the system is approaching a threshold where the remaining information is insufficient to maintain function.

The critical insight is that recursive degradation does not require malicious intent or catastrophic failure. It requires only that the system's outputs are slightly less informative than its inputs. A 1% information loss per iteration produces a 50% loss in 69 iterations. A 5% loss produces a 50% loss in 14 iterations. The compounding is relentless, and it operates below the threshold of human perception until the collapse is sudden and irreversible.

Distinctions

Recursive degradation should be distinguished from autocatalytic degradation, in which the degradation itself accelerates the rate of degradation. In recursive degradation, the rate of loss is constant; in autocatalytic degradation, it increases. Model collapse exhibits both: the recursive structure produces compounding loss, and the loss of diversity accelerates the rate at which further diversity is lost.

It should also be distinguished from epistemic collapse, which is the systems-level consequence of recursive degradation in knowledge-producing systems. Recursive degradation is the mechanism; epistemic collapse is the outcome.

Prevention

Prevention requires breaking the recursion or injecting error correction. Breaking the recursion means ensuring that the system's inputs come from sources external to the system — human-generated data, independent measurements, or adversarial tests. Injecting error correction means designing feedback loops that detect and correct degradation before it compounds. Neither is easy. Both are necessary.