Back

01 — AI Systems · 2024

AI Search System

Intelligent information retrieval for a scaled media platform — designing for trust, ambiguity, and discovery at 12M+ users.

Year

2024

Role

Lead Product Designer

Scope

End-to-end product design · AI/ML collaboration · Design system

Overview

A major media platform with millions of daily readers and thousands of articles published every day was running on keyword search that hadn’t meaningfully evolved in a decade. Users were getting 50+ results with no clear answer. Time-sensitive queries — breaking news, market shifts, political developments — returned stale, poorly-ranked content.

Simultaneously, AI-generated news content from competitors was eroding trust in the category. Our product had to be more trustworthy than traditional search, not less — while leveraging the efficiency of generative AI.

Tags

AI / ML SystemsSearch UXConversational InterfacesProduct Strategy

Core Challenge

How do you design an AI search interface that maintains journalistic credibility while leveraging the speed of generative AI?

Product Thinking

I was the sole designer on the AI search surface, embedded with a team of 3 ML engineers, 2 backend engineers, and a product manager. I led the design end-to-end: from initial problem framing to shipped product.

The hardest design problems here weren't visual — they were epistemological. What does it mean to present 'AI confidence'? When should the system answer vs. admit uncertainty? What visual language communicates trust without implying infallibility?

I ran a 3-week research sprint across power users, casual readers, and journalists. The finding that changed everything: users didn't want AI to replace search. They wanted it to surface structure in chaos. That reframing drove every subsequent decision.

Key Decisions

AI synthesis vs. raw results

We had two mental models: show the AI answer, or show the traditional list. We chose a third path — a synthesis layer that cited sources inline, making AI feel like editorial curation rather than generation. Users could trace every claim.

Designing for uncertainty

The model's confidence varied enormously across query types. Rather than hiding this, we designed confidence states — subtle visual shifts that told users when to trust the AI and when to do their own reading. Counterintuitively, showing uncertainty increased trust.

Disambiguation over clarification

When a query was ambiguous, the default instinct was to ask a follow-up question. We found this felt interrogative. Instead, we showed multi-intent result splits, letting users self-select meaning rather than re-state it. Reduced query abandonment by 31%.

Failure states as trust builders

The easiest thing to design was a generic error. We built a graceful degradation system instead — when the AI had no good answer, it said so clearly, offered related context, and suggested search alternatives. Users rated these states more trustworthy than vague AI responses.

Visual Evolution

Exploration

Mid-fidelity

Shipped

Outcome

+40%

Query resolution rate

vs. previous search

2.8×

Search engagement

session depth

12M+

Users served

across platform

−31%

Query abandonment

on ambiguous queries

Reflection

The biggest lesson was that AI design at scale is fundamentally about trust calibration. Users don’t need the AI to be perfect — they need to understand when it isn’t. The systems that performed best were the ones that were most honest about their limitations. Designing for uncertainty turned out to be a competitive advantage.