The Interface Has to Show the Work
Trustworthy AI products have to reveal what they know, why they acted, and where human judgment still belongs.
AI products have a trust problem.
Not because people don’t understand that AI can be wrong. Most people know that by now. The problem is that the interface still behaves as if the answer is settled.
A model guesses. The product presents. The user decides.
That gap is where design has to do more work. The interface has to show what the system knows, how certain it is, why it acted, and where the human still gets to intervene.
The problem gets sharper when the system is probabilistic. Pratik Joglekar, writing in Smashing Magazine, warns about “probabilistic systems wrapped in deterministic interfaces”. The model offers a guess. The interface presents it like an answer. Then the user, the team, or the organization acts on it. In medical diagnostics, financial forecasting, commerce, hiring, or even a generated product summary, that confidence can be dangerous when the UI hides the uncertainty behind it.
This is where a color catalog becomes a surprisingly useful product lesson. Storied Colors makes provenance part of the interface. Each entry carries source labels, status labels, chemistry, dates, alternate labels, citations, and a correction process. Its methodology page answers the question every AI product should be asking itself: how do you know? The site is beautiful. The trust comes from showing its evidence.
AI interfaces need more of that. Elizabeth Pizzuti’s piece on agentic commerce separates two very different problems: AI shoppers buying from stores, and AI workers operating for merchants. On the shopper side, the system needs structured data that an agent can read. On the merchant side, the interface has to expose the agent’s homework: the observation, impact, reasoning, margin, risk, and undo path behind a recommendation. A merchant will not approve a pricing change because the system sounds confident. They need to see the work.
That changes what “personalization” means in enterprise software. Tiina Golub points at software that knows the user’s real context: role, permissions, workflow state, and the kind of decision the person is trying to make. That is much more useful than a dashboard with someone’s name on it. But it also raises the bar for explanation. The more the product adapts to me, the more it has to explain which version of me it thinks it is serving.
Jenny Xie’s collection of future-state sketches for Figma pushes the same idea into the interface itself. Controls appear only when they are needed. Pacing changes when the user is overwhelmed. Sound, motion, and visual cues help people understand where they are in the experience. Those details become signals that orient the user, especially when the system is doing more on its own.
And then there is the question of care. Nolen Royalty argues that AI has broken the old shortcut where polish implied effort. A warm-cream Claude website can look finished without feeling considered. That does not mean we should start adding fake typos to prove a person was involved. It means polish no longer proves care. The work has to carry more specific evidence of judgment: why this hierarchy, why this recommendation, why this amount of friction, why this handoff back to the user.
So the interface has to show the work. It is the place where a person clicks, approves, edits, or buys, and the place where the system accounts for itself. That is a design problem, and it is exactly where design gains power as AI makes output cheaper. Someone still has to decide what the product should reveal, when confidence should be softened, when automation should stop, and what trail a user needs before they can trust the next action.
A good AI product earns belief by showing its work.
What I’m Consuming
Using Opus 4.8 to get a second opinion on an MRI and where it leaves me. Antoine uses Opus 4.8 inside Claude Code to review a shoulder MRI after a clinic diagnosed a Grade III partial-thickness tendon tear. The unsettling part is not that the AI disagreed; it is that the disagreement left him between two systems he could not fully trust. It is a very human version of the same AI problem: more analysis does not automatically create more certainty. (Antoine / antoine.fi)
The Web We Know Is Going to Disappear. The author of Minid.net walks through BBSs, the Web, Flash, mobile apps, and AI chat as successive containers for the same human behaviors: publishing, searching, learning, arguing, and finding one another. The sobering part is the claim that websites may become infrastructure for machines more than destinations for people. The Web survives, but maybe as a smaller place for the people who still want an online home they control. (Minid.net)
Config 2026: New Materials, New Tools and a More Expressive Canvas. Figma’s recap is a useful snapshot of where the canvas is headed: code layers, native motion, shaders, generative plugins, Weave, and an agent with skills and connectors. The code-layer idea is the one I’ll be watching, because it lets designers manipulate code on the canvas instead of treating it as a separate handoff artifact. (Figma)
Daring Fireball: Om. John Gruber writes a moving remembrance of Om Malik as a friend, critic, journalist, investor, and fixture of Apple events. The best part is how Gruber tracks Malik’s transformation after his 2008 heart attack, from relentless breaking-news blogger to essayist interested in why things happen. It is also a reminder to tell people when their work matters while they can still hear it. (John Gruber / Daring Fireball)
Your brain was never designed for this much bad news. This ScienceDaily piece explains news fatigue as a mismatch between our threat-sensitive brains and a global stream of bad news. The numbers are stark: the Reuters Institute’s 2025 Digital News Report found that 40 percent of people globally at least sometimes or often avoid the news, and another study found severe problematic news consumption in 17 percent of American adults. The useful advice is boundaries rather than withdrawal: defined news windows, depth over volume, and a clear link between information and action. (ScienceDaily)