A system earns trust by being predictable in how it behaves, and explicit about what it doesn’t know.
I spent about a year on the design of Creator Assistant at Meta, an AI assistant that reads a creator’s own performance data and gives guidance. The rule we held wasn’t a confidence score or a disclaimer banner. It was simpler: if it couldn’t find an explicit answer, it told the creator that. For a product whose job is telling someone how to grow their livelihood, a confident wrong answer is worse than no answer, because the creator acts on it.
The conviction predates AI. I spent eight years at Capital One making authentication invisible in low-risk sessions and present only when risk rose. Trust wasn’t a feeling we asked users to have; it was the byproduct of a system behaving the same way every time. AI just raises the stakes: the moment someone sniffs a hallucination, creator or lawyer, their trust in the whole product degrades at once.
A minimally viable product means one that people can trust.
That line is from the Creator AI Design Principles I wrote for the org, adopted by leadership before any of us had a scar to point at. That’s the point of writing principles down: the standard shouldn’t depend on you being in the room.
Data is not an insight yet. The insight is the why.
I inherited a team at Facebook called Insights and had them renamed Analytics, specifically so we could have the argument out loud. You’re not providing insights. You’re providing data. The insight is the reason behind the number, and pretending otherwise flatters the dashboard while failing the person reading it.
At hundreds of millions of accounts, most creators will never be analysts, and their needs generalize down to a handful of patterns. So the temptation is theater: make every creator feel personally understood while telling everyone roughly the same thing. Personalization that’s real is a feature. Personalization that’s theater is a trust liability that only costs you once.
The honest version of the work sounds humbler and lands harder. When we rebuilt the creator dashboard, the pitch to executives was never “make it prettier.” It was “focus people on what matters.” Analytics on a phone exist for the stock-check moment: the subway, the waiting room. Snackable, with depth on demand. It shipped +3% DAU with no new features. The features were already there; the why wasn’t.
When everybody can build, everybody shares accountability for coherence.
On Marketplace, we rounded the corners of listing cards by four pixels. It caused a 10% regression. Nothing about the change was wrong in isolation. It was wrong at scale. Consistency is following the rules. Coherence is seeing the blast radius before you ship.
This one matters more now than it did when I learned it, because building stopped being scarce. AI lets anyone make the thing. I run a one-person product org on it. What stays scarce is judgment: a clear, specific point of view about whether the thing belongs. At certain phases you should break outside the design system, in service of finding the right execution. The discipline is knowing which phase you’re in, and bringing the work back inside when it’s time.
Design systems made us consistent. The next decade needs them to make us coherent: to show a team what four pixels touch before the four pixels ship.