Redundancy is usually treated as a flaw in communication — something to trim, compress, optimize away. But Claude Shannon, designing the mathematical foundations of information theory in 1948, showed that redundancy is not noise in a channel; it is the very mechanism by which meaning survives noise. English, he calculated, is roughly 50% redundant — and that redundancy is why you can read a typo-ridden text or hear someone through a crackling phone. The channel fails; the message lands anyway. What Shannon revealed about signal transmission turns out to describe something deeper about how we communicate with AI systems: when you strip your prompts down to maximum efficiency — no context, no repetition, no apparent 'waste' — you are removing the redundancy that lets the system correctly resolve ambiguity. The psychologist George Miller showed in the 1950s that human working memory caps out around seven chunks of information, which is why we compensate in conversation through restatement, examples, and elaboration. AI language models have different constraints but the same vulnerability: sparse input produces high-variance output. A little deliberate redundancy — restating your intent in two different ways, adding a concrete example alongside your abstract request — is not padding. It is engineering. Today, before you send a tightly compressed prompt or message, add one sentence that says the same thing differently.

Think of a recent AI output that missed what you meant — what context did you assume was obvious but never actually included?

Drawing from Information Theory / Cognitive Psychology — Claude Shannon & George Miller