Netflix AI Discovery
Using reinforcement learning to help users narrow down options using an adaptive flow and discover what to watch next
The Problem Area
AI is improving retrieval but not helping users make better decisions
Netflix introduced AI search to simplify content discovery through natural language prompts.
But despite being positioned as an AI-powered experience, the interaction model still mirrors traditional search systems:
User enters prompt → system returns static results
This assumes users know exactly what they want. But entertainment decisions are rarely that linear.
The Challenge
Streaming decisions are emotional, contextual and constantly evolving. Yet current AI search experiences treat discovery like a one-time query problem.
Traditional search UX works when users know exactly what they want.
AI creates an opportunity to handle ambiguity better—but most products still rely on old mental models.
Framing
How might AI help users discover what they want when they don’t know how to articulate it yet?
This meant solving for:
vague intent
evolving preferences
recommendation trust
user control
AI transparency
Design Opportunity
AI is uniquely positioned to adapt in real-time
Unlike traditional recommendation systems, AI can:
interpret vague prompts
learn through feedback
refine recommendations dynamically
explain reasoning
adapt to changing intent
This creates an opportunity to redesign content discovery as a conversation instead of a transaction.
The Solution
Adaptive Discovery: Instead of asking users to generate the perfect prompt upfront, the system progressively helps them discover what they want.
AI Orchestration
01
Explicit Intent
These are signals users intentionally provide.
Inputs:
mood
genre preferences
runtime preference
Familiarity
02
Behavioral Signals
The system learns from what users do—not just what they say.
Inputs:
watch history
completion rate
skipped titles
abandoned recommendations
rewatch behavior
03
Feedback Signals
The most important layer. The system continuously improves through directional feedback.
Users can respond with:
more like this
something different
shorter
surprise me
04
Recommendation Engine
The orchestration layer combines all signals to generate recommendations.
Instead of surfacing endless options, the system identifies:
best fit recommendation
confidence level
explanation layer
05
Confidence Layer
AI shouldn’t pretend certainty. When signals are weak, the system communicates lower confidence. This builds trust through transparency