Behind the AI Curtain - Designing for Machine Learning Products

When startups first launch, they can make the news with application of cutting edge AI - but convincing users to trust the AI is often another story. There's often also no process for integrating future AI development into product roadmaps.

This session covers three key principles for how design and data science teams can work together better to build greater trust among users. Additionally, a case study on how a design and data science team partnered to redesign predictive analytics scores powered by machine learning will illustrate those principles in practice.

Transcript

1. Behind the AI Curtain:
   Designing for Machine Learning Products
   #aicurtain
   Crystal Yan
   

   View Slide

2. I’m Crystal Yan
   I’m a designer and product manager who works
   with data scientists every day.
   You can find me at @crystalcy
   Hello! ✋
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

3. Terminology Cheatsheet
   Data Science
   Getting insights from
   data
   (everything from business
   analytics and statistics to
   machine learning)
   Artificial Intelligence
   Machines have
   intelligent behavior
   (goals and methods include
   machine learning, natural
   language processing,
   computer vision, facial
   recognition)
   Machine Learning
   Computers learn on
   their own without
   explicit programming,
   requires lots of data
   (ML is a method that can be
   applied to create models that
   will predict, aka predictive
   analytics)
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

4. Today’s Agenda
   Introduction
   Why this matters
   Principles
   1. Less is more
   2. Ask the right
   questions
   3. Writing well
   matters
   Case Study
   Redesigning
   predictive analytics
   scores
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

5. Artificial intelligence is
   changing the
   It’s everywhere, whether you see it or not. We interact with more
   systems powered by AI each day.
   I work with data scientists and often meet designers and clients who
   ask, “Are algorithms here to take my job?! ”
   1
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

6. The New Yorker
   What happens when
   machines out-diagnose
   doctors?
   From medicine to SaaS
   VC blogs
   Will your users trust your
   analysis / will they pay for it?
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

7. “
   “Artificial intelligence will have reached human
   levels by 2029. Follow that out further to say,
   2045, we will have multiplied the intelligence, the
   human biological machine intelligence of our
   civilization a billion-fold.”
   -Ray Kurzweil
   Inventor
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

8. “
   “The field of AI has traditionally been focused on
   computational intelligence, not on social or
   emotional intelligence. Yet being deficient in EQ
   can be a great disadvantage in society.”
   -Rana el Kaliouby
   Cofounder of Affectiva
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

9. “
   “We all have a responsibility to make sure
   everyone - including companies, governments
   and researchers - develop AI with diversity in
   mind.”
   -Fei-Fei Li
   Professor at Stanford University and Director of AI Lab
   @crystalcy | #aicurtain
   crystalcyan.github.io
   

   View Slide

10. Three Principles
    Here are three principles for establishing greater trust in the machine
    learning behind your design.
    2
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

11. Principle #1:
    Less is more.
    Sometimes...it pays to hide the numbers. @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

12. Principle #2:
    Ask the right questions
    Just right:
    How would you
    explain this?
    Too leading:
    Does this make
    sense?
    Do you like this?
    Too
    open-ended:
    What would you
    do? What would
    you call this?
    Most importantly: listen to their questions. @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

13. Principle #3:
    Writing well matters.
    ◉ Define your audience and purpose.
    ◉ Set tone/personality and match to brand.
    ◉ Be concise. Solution first, evidence after (for those who seek it).
    Good writing is concise, scannable, objective, and actionable.
    Resources: plainlanguage.gov, Letting Go of the Words (book)
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

14. Redesigning Predictive
    Analytics Scores
    Case Study: FiscalNote
    3
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

15. test
    Our design process
    dev &
    release
    iterate
    define
    problem
    ideate/
    prototype
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

16. We promised the world.
    But people had trouble
    understanding this score,
    and the company strategy
    changed.
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

17. But...why?
    The to the
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

18. The redesign.
    We hid the numbers
    We gave an explanation
    We adopted a
    conversational tone
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

19. $MONEY
    a lot of revenue at the time attributed to predictive analytics
    140+
    number of training docs we created on our internal drive to try to
    explain the scores
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

20. 5/5
    everyone could concisely articulate how they would explain the new scores
    to a coworker
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

21. Recap
    Less is more
    In our case, it made sense to
    hide the numbers.
    Ask the right questions
    What you ask defines what
    you’ll get. Listen for insights from
    the questions users ask.
    Writing well matters
    Brush up on your writing skills,
    or risk getting left behind.
    Adapt
    Algorithms might not take your
    job, but you must adapt.
    Why matters more than what
    People wanted to know why we
    gave a particular score. In
    general, they preferred a less
    accurate human analyst over a
    more accurate black box.
    Copy > graphics
    Our redesign shifted focus from
    charts to copy.
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide

22. Any questions?
    You can find me at
    ◉ @crystalcy
    ◉ crystalcyan.github.io
    Thanks!
    @crystalcy | #aicurtain
    crystalcyan.github.io
    

    View Slide