Show the work

The instinct to hide the machinery is a holdover from deterministic software, where the machinery was boring and correct. AI machinery is neither. When a system might be wrong, showing how it worked is not a nice-to-have transparency gesture — it is the primary interface for calibrating trust. People are surprisingly good at judging a process (“it searched the wrong docs”) and surprisingly bad at judging a bare conclusion.

Agent products learned this fast because they had to. Claude Code and Cursor stream a working transcript — files read, commands run, edits made — and that transcript is what makes handing an agent real work tolerable: you can watch it head down a wrong path and stop it in minute one instead of discovering the wreckage in minute twenty. Perplexity shows the searches it runs and cites every claim, and those citations are the product; the same answers without sources would be just another chatbot. Deep-research modes in ChatGPT and Gemini narrate their plan and progress across multi-minute runs, because a silent ten-minute spinner is indistinguishable from a hang.

Streaming itself is show-the-work at the token level. It is honest about the generative nature of the output — you watch it being composed, not retrieved — and it converts dead wait into read time. But streaming raw tokens is table stakes, not the principle fulfilled. The work worth showing is the reasoning-shaped work: what was consulted, what was assumed, what was skipped.

The failure mode on the other side is real: showing everything is its own kind of hiding. A wall of tool-call logs is as illegible as a black box. This is where progressive disclosure earns its keep — a one-line summary of what happened, expandable to the full trace for the person who needs to audit it. ChatGPT’s collapsed “thought for 40 seconds” disclosure is the right shape even when the contents are thin: headline first, receipts on demand.

Draw the line at theater. Fake progress stages, invented “thinking” copy, confidence dressed up as diligence — people learn quickly when the work being shown isn’t the work being done, and the discovery poisons every legitimate transcript you render afterward. Show real work or show a spinner; never show a performance.

The counter-argument deserves respect: most people don’t want the work, they want the answer. Apple built the most beloved products in history on ruthless concealment of machinery, and every pane of process you render is screen space, latency, and cognitive load taxed from the result itself. Worse, visible process can actively mislead — a fluent, plausible-looking chain of searches lends false authority to a wrong conclusion, and research on explanation interfaces keeps finding that shown reasoning increases trust without increasing accuracy of trust. A clean answer that’s right is a better product than an audited answer that’s right. All true — for systems that are reliably right. The argument for concealment assumes deterministic correctness AI does not have, and it collapses at exactly the moments that define your reputation: the confident miss. Progressive disclosure resolves most of the tension — the answer leads, the receipts collapse — but the receipts must exist, because the one user in ten who checks them is the one the other nine listen to.

Measure whether the shown work is doing verification work. The core signals are expansion rate and interruption yield: what fraction of responses have their process disclosure opened, and — for agents — what fraction of aborted runs were stopped mid-transcript versus discovered wrong after completion. Good looks like a meaningful minority of disclosures opened (universal expansion means your summaries are failing; zero means the affordance is decoration) and the majority of bad agent runs caught during, not after — Claude Code’s economics work because wrong turns die in minute one. For citation products, track citation click-through on contested or high-stakes queries and, in evals, whether users correctly reject planted wrong answers at a higher rate with the process visible than without. That last number is the principle distilled: show-the-work is succeeding when visible process measurably improves the user’s error detection, not just their confidence.