Information Loss

Information loss is the irreversible destruction of signal content that occurs when a system filters, aggregates, or standardizes its inputs beyond the recoverable threshold. Unlike noise — which is additive and can sometimes be filtered out — information loss is subtractive: the discarded signal cannot be reconstructed from what remains. A thermometer that reports only integer degrees loses the information about fractional temperatures; a census that records only age brackets loses the information about exact ages. The loss is structural, not incidental: it is built into the design of the interface.

In cybernetics, information loss is the shadow side of variety attenuation. Every attenuation mechanism carries an information-loss budget: the amount of signal that can be discarded without impairing the regulator's function. Setting this budget correctly is the central design problem. An overly conservative budget preserves too much variety and overwhelms the regulator; an overly aggressive budget destroys information that the regulator needs. The budget is not derivable from first principles; it depends on the stakes of the decisions the regulator must make.

The concept applies with particular force to epistemic infrastructure. Scientific journals lose information when they reject submissions that do not fit current paradigms; some of those submissions would have been revolutionary. Social media algorithms lose information when they optimize for engagement rather than accuracy; the signal that is most engaging is not the signal that is most true. In both cases, the information loss is not a bug but a design feature — and the question is whether the feature is fit for purpose.

Information loss is invisible until it matters. The signal you needed was the signal you discarded before you knew you needed it. This is why every attenuation mechanism must include a dead letter office — a channel for signals that do not fit the current filter but might fit a future one.