The margin risk screen at SoFi showed concentration requirements,
excess equity, and liquidation thresholds that were conditional on
volatility assumptions most users couldn’t see. The product instinct
in the redesign was to simplify. Users weren’t reading the complex
conditional disclosures, and cognitive load was the standard
diagnosis. Clean up the interface, reduce the noise, make the key
number legible.
Compliance pushed back in a review. The conditional structure in the
disclosure wasn’t decoration. The threshold was genuinely sensitive
to concentration and volatility in ways that mattered enormously
under adverse conditions and were invisible in normal ones. Flatten
the display and the number stops meaning what the number means.
We ran user sessions to resolve the tension. What we observed was
uncomfortable.
Users didn’t misinterpret the complex disclosures. They over-trusted
the simplified ones. When the interface felt clean, users read it as
stability. When it showed conditional thresholds with explicit
dependencies, they slowed down, asked questions, and made different
decisions. The complexity of the disclosure was representing the
complexity of the underlying risk, and users were treating that
complexity as information.
Edward Tufte’s analysis of the Challenger decision makes the inverse
point from a different angle. The O-ring temperature data existed in
the presentations given to NASA engineers the night before the
launch. But the design of the slides made the trend invisible. The
information was formally present in the deck and operationally
absent in the decision. The problem wasn’t that the engineers had
bad data. The problem was that the presentation structure
communicated confidence the data didn’t support.
Our situation was the symmetrical case. The original screen
communicated uncertainty the users needed to see. Simplifying
removed the uncertainty cue without changing the underlying risk.
The compliance concerns were real and the UX problems were real, and
we ended up with a modified simplification that preserved the most
load-bearing conditionality. But I came away from that review with a
conviction that has held up in every AI project I have worked on
since: legibility of uncertainty is a design responsibility.
An interface that makes a probabilistic, conditional, context-
sensitive situation feel stable and simple is a miscalibration
machine. It produces users whose confidence exceeds what the system
has earned. The gap closes later, suddenly, at someone’s expense.
Every time I see a smooth AI output that collapses model uncertainty
into a single confident answer, I am back in that compliance review
asking the question I came out of it asking. Is this interface
making the system feel more reliable than it is, and who pays when
the gap between feeling and reality resolves?
