Writing

Why "Accurate" AI assisted Products Fail

AI performance is improving fast. We measure accuracy, precision, and benchmarks. But in the real world, adoption doesn't track model quality linearly. Two systems can be similarly “accurate”... yet one is trusted while the other is ignored.

Because users don't experience metrics. They experience signals.

Symbolic vs. Analogic Communication

  • Symbolic Communication (The Content). What the AI says, ie the answer. The words, numbers, and citations. This is the “content channel.”

  • Analogic Communication (The Signal). How the AI feels as it works. The cues for confidence, control, and safety. This is the “trust channel.”

  • The product trap. Often the product experience over-indexes on symbolic (improving the answer) and under-invests in analogic (improving the user's confidence in using the answer).

Examples of analogic wins

  • ChatGPT's “Thinking” States. These signal competence. Before you even see the answer, the UI tells you: This isn't just an autocomplete — I'm doing work.

  • Gmail Smart Compose. The magic is the micro-control. Accept with a tab, ignore by typing. Low effort to try, zero cost to reject. This builds trust through repeated, safe wins.

  • Shopify's AI Product Description. It doesn't just “write copy.” It highlights exactly what it changed. It reduces the risk of the AI touching the user's brand voice.