>
Home
How might we create a scalable yet efficient livestream moderation chat experience?
By working with AI-powered moderation, channeling system to support fast bulk actions without losing user empathy.
Scalability
Empathy
Efficiency
Engagement
Defining Success
Given YouTube’s broad and diverse user base with varying viewing habits, I defined success metrics alongside counter-metrics to ensure balanced evaluation and avoid optimizing for a single user group or behavior.
Success Metrics
We want to achieve...
Counter Metrics
but not at the cost of ...
Hover-over tags to explain AI decisions
Research Inisghts
I interviewed six regular livestream viewers and two YouTube staff members, and observed more than ten livestreams for this project.
Top Themes
Lack of Signals
Livestream chat moderation is difficult because it lacks sufficient signals.
Low sense of engagement
Audience feels unseen and lack of agency.
Build trust by problem detection
Aligned with YouTube’s trust principles, chat moderation must identify trusted and untrusted users in real time.
Low explainability is what we need
Low explainability is sometimes necessary to prevent malicious learning.
Supporting Quotes
9:41
Is there anything else you'd like to share with us to help improve your experience?
Submit
Back
1
2
3
4
5
6
7
8
9:41
Next
Back
How would you describe the condition of the café?
Drink & food prep area
Dine-in area (if used)
Bathroom (if used)
Register area
Area
Rating
Poor
Excellent
1
2
3
4
5
6
7
8
9:41
Next
Back
Were you given enough information about:
(select all that apply)
Estimated wait time
Peak times (longer wait)
Help from staff when needed
Confirmation that my needs (e.g., allergy) were taken care of
None of the above
1
2
3
4
5
6
7
Boost Retention Rate in a B2C User Experience
I conducted user research and redesigned the survey to build a pipeline for feedback from customers.
Petunia’s
Favorite activities
Fun Facts about
Petunia
Pig / Göttingen minipig
Sunbathing
Rooting
Rolling in the mud
Nice to meet you!
Snuggling with her friends
Species/ Breed
Bella, Tofu, Peppa, Carrie,
Saigo, and Rober the pigs
Best friends
Pig food, most fruits and
veggies, grass, hay
Favorite Food
Onions, Brussels sprouts
Least Favorite Food
Rescued from an animal
testing lab by the Beagle
Freedom Project
Background
Date of birth
Arrival at Charlie Acres
December 10, 2015
December 10, 2016
Accessible Design System for Humane Education
I created a custom icon set, maintained and expanded the design system to support the development of digital and print materials to create a memorable education experience for visitors, children, and teachers.
Back to top
Brand Value
Trust & Safety
Freedom of Speech
Explainability
Abuse Prevention
Scalability
Empathy
Efficiency
Problems
Subtle or missing signals make user intent difficult to detect, leading to errors that undermine system reliability
Trust and Safety
2K views
• 2 months ago
Automation often misses tone and emotion, which can reduce user trust and hurt long-term community engagement.
Community Energy
2M views
• 3 months ago
Moderators share the same interface as users, limiting their control and making platform safety difficult to scale
Efficiency
1K views
• 1 months ago
Solutions
Trust score =
Tone weight +Certainty weight + Contextual Risk
AI-supported trust score system leveraging real-time analysis to differentiate trusted and untrusted users
Trust and Safety
2K views
• 1 months ago
toxic
concern
spam
misinfo
question
super user
NEW
AI-powered taxonomy system to classify messages based on intent, language, and risk level
Community Energy
2M views
• 3 months ago
Dual-panel moderation interface with a dashboard for greater efficiency
Efficiency
1K views
• 1 months ago
Three Key Pillars in Livestream Chat Moderation
After synthesizing research insights, I defined three key pillars to guide this project.
Moderation priorities across different livestream genres
Understanding how to integrate them effectively.
What
01
Who
User engagement patterns
Studying behaviors and identifying scalable detection methods.
02
How
Accessible moderation experiences
Designing chat interfaces that work across livestreams of varying scales and genres.
03
Hover us to see the solutions
Dual-panel Moderation
supports streamlined workflow
Moderator Panel
(Dashboard)
+
Engagement Panel
(Chat)
before
after
Dual-Panel Moderation
Here are three primary actions in the new moderator dashboard:
Z and C
Z
I think the guys blind hi everyone
trapstar4082
I just subed
Amy Rachel
A
A
I think ts trash it's not even cool plus he is not moving
A
Ainsley14
I thought this was real life
Arlena Walcott
🫣🫣🫣🫣🫣🫣🫣
gingerhello16
Totally agree! The ray tracing implementation is such a chef kiss
Engagement Panel
Nezuko
The lighting in this new map looks incredible
2
1
super user
noobHunter88
The lighting in this new map looks incredible. Major upgrade from the last update.
mochiball
NEW
hi everyone! this is my first time. great content so far
send welcome message
trapstar
question
honestly, this soundtrack is s tier. anyone know the composer?
4
3
Remy
Say something...
0/200
Moderator Panel
AI Assistant
Current viewers
1,284
Audience sentiments
Peak today
Chat summary
Recent activity
Viewers were engaged and excited, though some criticized the streamer’s gameplay and unclear guidance on accessing game features.
1,529
+2.1% vs last week
Neutral
Positive
Last updated 2 mins ago
25%
27%
Negative
Quick actions
Review held comments
12
Manage audience reports
9
View audience question
20
Add blocked-word
Blocked words
Enter comma-separated values
Settings
8 users flagged comments as spam in the last minute.
30s
Moderator_b blocked user choco_muffin.
30s
UrbanExplorer removed a message from TriumphEdits.
30s
AI summary updated: new concern raised by audience.
30s
Dashboard
Held for review
12
From audience
9
62%
Neutral
Positive
Negative
View high-level metrics, such as number of viewers, audience sentiment, and chat summaries.
Review comments flagged by AI and audience reports in real time, and approve or block them.
Manage moderation settings, such as adding blocked words and turning off negotiable pre-filtered tags.
Limited controls:
warn, timeout, ban
Same view as audience
Integrating AI into Workflow: The Third Panel
Activated on click to help moderators:
Adjust moderation sensitivity (for example, how strict the AI assistant is with comments).
Understand AI decisions and disagree by reporting or overriding them.
Surface comments labeled as questions to support audience engagement.
Hidden Comments
Highlighted Comments
Demonstrates scalability for complex control scenarios
While offering simplicity by showing only one panel at a time for users who prefer focused interaction
Project Type
Graduate Capstone
Tools
Figma, Lovable, FigJam
Role
Sole Product Designer
Duration
3 months
Impact
Through user segmentation and AI-powered semantic analysis, the system differentiates trusted and untrusted users, enabling a safer, more trustworthy, and scalable livestream chat moderation experience with greater efficiency.