Designing AI Products and Services: An Annotated Syllabus
Carnegie Mellon University Human-Computer Interaction Institute 05617/317 (2019 Spring)
We live in a world where the coolest things that can be made with AI are most often envisioned by someone with a Ph.D. in machine learning. Very few of them, however, have expertise in giving form to new products or services that people experience as valuable or meaningful in day-to-day lives, nor it is the focus of their work. Designers have such expertise, yet bringing such expertise to bear in the use of AI has proven to be difficult. Further, in the midst of the AI hype, novice designers often think of AI as magic; They envision things that cannot be made in the near future.
In 2018, I had the great pleasure to join Prof.John Zimmerman and Prof. Paul Pangaro in developing this new course <Design of Artificial Intelligence (AI) Products and Services>. Through this course, we wanted to help design/HCI students to put a stake in the ground on applying AI in today’s commercial products; We wanted Computer Science students to join us contemplating what is a preferred AI-infused future.
As with so many of our students, we are incredibly excited about sharing with you this exceptional curriculum here. New courses represent new challenges. This Medium publication will be a living document where we continue to share pedagogical reflections on designing AI. Your comments, critiques, and questions will be much appreciated.
Below, you’ll see our annotated syllabus as of 2019 Spring, the first time this course was taught.
Learning Objectives
Through this course, students will gain a broad overview of what AI can do and, more importantly, a felt understanding of working with AI as a design material. Students will be able to:
- Develop a basic understanding of different kinds of AI/machine learning systems and how they function to enable a product or service.
- Understand how designers engage in a conversation with design materials through reflection in- and on- action.
- Gain hands-on experience in engaging AI as a design material in order to effectively envision new products and services that 1) people desire and 2) that can be technically achieved.
- Present AI product/service designs, describing both the human need and technical capability.
This course requires no software development.
Course Activities
This course will involve lectures, in-class discussions, in-class design sprints, and four design projects (including in-class presentations and critiques).
- Intro
- Project 1: Matchmaking, AI in search for users
- Project 2: Designing Crowd as AI Proxy
- Project 3: Designing Adaptive User Interfaces
- Project 4: Making NLP Useful
- AI Ethics, AI Discrimination, and Data Privacy
Instructors
- John Zimmerman, Professor, HCI Institute
- Qian Yang, Ph.D. Candidate, HCI Institute
Special thanks to Paul Pangaro (Professor of Practice at HCII), who provided tremendous help with designing this course and gave a wonderful guest lecture on conversational designs.
Part 0. To Orient and Calibrate
Both “AI” and “design” are hard to define. In the first lecture, students will read about different definitions. We will then calibrate them with some working definitions:
- What differentiates a product from a service?
- What do we mean by design (verb.)?
- What do we mean by Artificial Intelligence?
In doing so, students will begin to discern the gap between an algorithm’s and a UX/service designer’s view on users, on their behaviors and contexts, on preferred futures and AI breakdowns.
Readings
- My Favorite UX Design Process Diagrams
by Joe Steinkamp
Link to article - On Modeling the Analysis-Synthesis Bridge (Interactions, 2008)
by Hugh Dubberly and Shelley Evenson
Link to article - Machine Learning: The Power And Promise Of Computers That Learn By Example (royalsociety.com, 2017)
by The Royal Society.
Link to article - This Is How Netflix’s Top-Secret Recommendation System Works (Wired, 2017)
By Libby Plummer
Link to article
Part 1. Design through Matchmaking
Designers typically bridge available technologies and user needs via a user-centered design (UCD) process. The task of “designing AI products”, however, does not easily fit the mold. If designers embark on a UCD process and select a group of target users, there is no guarantee that the user research will reveal a need for AI. Even if designers do find a user need for AI, there is no guarantee that there will already be (or ever be) data to train the system.
In week 1–3, students will learn many different ways to meaningfully connecting what do users want and what can AI realistically offer. Students will begin by reading about cybernetics and “designing AI”; about five machine learning tribes and machine teaching. In Project 1, students will delve more deeply into “matchmaking”, a four-step design process that helps designers start from existing technology and systematically look for users who are most likely to benefit from it.
The reading, discussion, and hands-on practice will bring about two central questions that undergird this course: How to bring a human-centered ethos to AI, when designers have already preselected a class of technological solutions? How can designers envision new, buildable AI products and services that users want?
Readings
- Pedro Domingos’ on “Five Machine Learning Tribes” (2015)
by Shannon Kempe
Link to article
- Design through matchmaking: technology in search of users
(Interactions, 1999)
by Sara Bly and Elizabeth F. Churchill
Link to article - Data Fail! How Google Flu Trends Fell Way Short
(LiveScience.com, 2014)
by Stephanie Pappas - Cybernetics and the Design of the User Experience of AI Systems
(Interactions, 2018)
by Nikolas Martelaro and Wendy Ju - From machine learning to machine teaching: the importance of UX
(Interactions, 2018)
by Martin Lindvall, Jesper Molin, and Jonas Löwgren
Interactions 25, no. 6 ( 2018): 52–57
Link to article - Designing AI (Interactions, 2018)
by Elizabeth F. Churchill, Philip van Allen, and Mike Kuniavsky
Link to article
Project 1: Design through Matchmaking
Students will work individually to ideate a broad set of possible product/service concepts. We provided three “seeds” to drive their envisionment. For each of the seeds, ideate at least 20 product/service concepts. Then curate a selected set of 4 based on the following criteria: 1) The need of the target customers; 2) Ease of developing the system; 3) Risk of errors; 4) Size of the potential market.
- Seed 1: A dataset. Uber has hired you to discover new value that resides in their datasets. Their datasets include things like data on ride requests, food requests, and driver movements during and between rides. How can you generate new value for Uber by making inferences that benefit their stakeholders or benefit organizations they do not currently interact with?
- Seed 2: A technical capability (a patent). You have been hired by a company that makes and holds a strategic patent on a depth camera that can recognize gestures; intentional and unintentional movements and poses people make. How can this company discover some first, best customers for this technical advance? How can they squeeze the most value possible from the investment they made?
- Seed 3: A technical capability bounded by product form (e.g., tablet). Your client makes apps for iOS devices. The new iPhone and iPad have “Face ID”, a face recognition capability. It appears that Apple will make this feature available to 3rd party app developers. What new apps can you envision that take advantage of this capability to deliver new value to users?
Part 2: Crowd as AI Proxy
In week 4–6, students will begin to consider how their data-driven designs can motivate users to provide the desired data in order to feed the underlying AI system.
By studying crowdsourcing systems, students will learn to investigate and envision the when, where, and why of user producing data. AI and crowdsourcing are a perfect match: AI empowers crowds and enhances the value of their production, while corporates and organizations access the crowd not only for co-creation of products or their ingenuity but rather as trainers for AI systems.
Students will read about two typical paths towards using the crowd as a proxy for AI. They will engage both paths as starting places for their project:
- Designers want to deploy an AI system to address a particular user need; however, there is no data set available to train the system. Or, the problem could be that statistical approaches are not very good at fulfilling the specific need. (VizWiz, a system we will read about, offers one example of this innovation path.)
- Designers recognize that people are engaged in an activity that could be recaptured to build data for training an AI system. A crowd plays the role of an AI system through the work they are already doing, and this produces valuable training data to develop such a system. (Luis von Ahn’s ESP Game and reCAPTCHA both provide an example of this innovation path.)
Readings
Students will begin by learning about crowd-sourcing and human-computation.
- VizWiz: nearly real-time answers to visual questions (UIST’10)
by Jeffrey P. Bigham et al.
Link to article - Human computation: a survey and taxonomy of a growing field
(CHI’11)
by Alexander J. Quinn and Benjamin B. Bederson
Link to article
Next, about motivating crowd in the context of designing social network platforms:
- Chapter 2: Encouraging contribution to online communities
Building successful online communities: Evidence-based social design (2011): page 21–76.
by Robert E. Kraut and Paul Resnick
This is a very long chapter. Please skim and focus on where they discuss interaction design techniques to increase people’s willingness to make contributions.
Link to article - Online forums supporting grassroots participation in emergency preparedness and response (Communications of the ACM, 2007)
by Leysia Palen, Starr Roxanne Hiltz, and Sophia B. Liu
Link to article
Finally, students will read about human-centered machine learning, where motivating crowds converged with enabling machine learning systems.
- Power to the people: The role of humans in interactive machine learning (AI Magazine, 2014)
by Saleema Amershi et al.
Link to article
Project 2: Crowd as AI Proxy
Through this project, students will learn:
- To investigate user motivation using value flow models
- To gain a felt understanding of what people can do, what AI can do, and what AI needs data to do
- To use the crowd as AI proxy, designing the when, where, and why of user producing data
Students will work on teams to design a system that uses a crowd of people as a proxy for an AI computing system. The project has three distinct stages:
- Explore design space: Teams will consider many possible opportunities for services that could benefit from a crowd as an AI proxy.
- Model preferred future: Teams develop and refine a value flow model that describes how all of the stakeholders within their ecology gain value. At this stage, it is critical to understand how the crowd is motivated to produce the required data in a time-sensitive way.
- Refine interaction: Teams will develop a set of wireframes that show the transactional flows for crowd participants who generate the data. This should include scenarios of use, that describe a typical interaction with the system.
Design Sprint: Modeling Value Flow
Design Sprint: Framing a Good AI Problem
Unlike the popular portrayals of AI in the media, what AI can do even provided a sufficient amount of data can often be frustratingly limited and brittle. We will provide some rules of thumb to help students check the technical feasibility of their designs.
Part 3: Adaptive User Interface (UI)
In week 7–8, students will learn to integrate the aforementioned perspectives of AI product design (e.g., user needs, AI capabilities and limits, data flow, risks of error, etc.) into one simple product form: adaptive UIs.
Through this project, students will put a stake in the ground on how to design interactions that 1) adapt to a large number of different personas and use contexts; and 2) evolve over time.
Readings
- Planning Adaptive Mobile Experiences When Wireframing (DIS’16)
By Qian Yang, Aaron Steinfeld, John Zimmerman, and Anthony Tomasic
Link to article - Principles of mixed-initiative user interfaces (CHI’99)
By Eric Horvitz
Link to article - Being accurate is not enough: how accuracy metrics have hurt recommender systems (CHI EA’06)
by Sean M. McNee, John Riedl, and Joseph A. Konstan.
Link to article
Group Project: Adaptive Mobile UI
Through this project, students will learn:
- To recognize predictable and consistent user behavior patterns within an interaction flow, in order to identify collapsible or personalizable interactions;
- To design adaptive interaction flows that 1) can effectively improve user experience (with risks of inference errors taken into consideration), and 2) can create high quality labeled data needed to make the planned inferences;
- To design graceful recovery from AI inference errors;
Students will work on teams to improve the interaction flow of a well-known mobile app by making it adaptive. An adaptive mobile app reduces the time and interaction effort needed by collapsing navigation and selection based on interaction history. It improves the user experience by making the app more personal. In addition, an adaptive app design should be robust to error; users should be able to easily recover from an inference error without a significant increase in time or effort.
Part 4: Natural Language Intelligence
In the last 5 weeks of the semester, students will learn to design NLP-powered products. The fundamental challenges of this task are not too different from those of designing mundane adaptive UI (e.g., user needs, tech capabilities and limits, risks of error, etc.). In addition, students will take on a number of new challenges:
- Designing language-based interactions: Most design techniques and tools, such as wireframes, have a strong graphical underpinning and can be ill-fitted for designing language-based interactions;
- Error recovery: NLP-powered interactions often do not have a non-intelligent, baseline interaction to fall back on.
- Embodiment and agency: Computational programs that utter human languages inherently, more or less, embody human, complicating user expectations and interactions.
Readings
Students will begin by gaining a basic understanding of NLP, including its mechanisms, canonical applications, and aesthetic/experiential potentials.
- Advances in Natural Language Processing
(Science, 2015)
by Julia Hirschberg and Christopher D. Manning
Link to article - Experimental Creative Writing with the Vectorized Word
Talk by Allison Parrish (New York University)
Next, some of the most common applications of NLP in today’s commercial products, such as social media mining and recommenders.
- Computational approaches toward integrating quantified self-sensing and social media (CSCW’17)
by Munmun De Choudhury, Mrinal Kumar, and Ingmar Weber
Link to article - Clippy Didn’t Just Annoy You — He Changed the World
New York Magazine
By Brian Feldman
Link to article
The guest lecture by Prof. Paul Pangaro introduces the history of Conversational Theory and designing for conversations. “Conversations”, though in recent years often used to refer to language-based interactions (as in “conversational UI”), have a long idiosyncratic history in software design. All interactions are conversations, as computers and users understand, agree, and collaborate on effective action.
- What is conversation? How can we design for effective conversation?
(Interactions Magazine, 2009)
by Hugh Dubberly and Paul Pangaro
Link to article
This prepares students for the discussion about chatbots, their usefulness, and even implications for humanity.
- Intelligent Machines: The technology behind Open AI’s fiction-writing, fake-news-spewing AI, explained
by Karen Hao
MIT Technology Review (February 16, 2019).
Link to article - May A.I. Help You?
New York Times (2019) by Clive Thomson
Link to article - Google Duplex: A.I. Assistant Calls Local Businesses To Make Appointments (Demo)
Link to video - Google’s Duplex AI Scares Some People, but I Can’t Wait for It to Become a Thing
by David Pogue
Scientific America, August 1, 2018.
Link to article
- Should A Bot Have to Tell You It’s a Bot? by Kate O’Neill
Medium blog <Artificial Intelligence>, Mar 21, 2018
Link to article
Project: Making NLP Useful
Students will work on teams to envision a novel service that employs natural language processing (NLP) technology. The challenge for this assignment is to discover where the often very limited ability of NLP technology to make sense of people’s messages, conversations, and documents might still deliver value to a target set of users and to a service provider. This project has four distinctive stages:
- Matchmaking and focus setting (Similar to Project 1)
- Designing product form and context
- Prototyping to the biggest risk: Teams will identify the biggest risk to their design, construct a prototype to overcome this risk, and demonstrate it will not derail the design. To prototype, teams should construct a plan for simulating the likely behaviors and errors of the NLP system and identifying the quality of NLP performance needed to gain user acceptance.
- Refining interaction design (Similar to Project 2 & 3)
Part 5: Discrimination, Ethics, and Privacy
Issues of algorithmic discrimination, AI ethics, and data privacy will undergird all parts of the course, throughout the semester. In every project, students will grapple with these boundaries when trying to milk most data out of their selected algorithm capabilities or user populations. Towards the end of the semester, students will have a more focused discussion on these issues.
Readings
- The code of ethics for AI and chatbots that every brand should follow
by Trips Reddy
IBM Watson Blog (October 15, 2017)
Link to article - How Automated Tools Discriminate Against Black Language Platforms can have the power to moderate not just content, but language itself
by Anna Chung
Medium (Feb 28, 2019)
Link to article - What Your Boss Could Learn by Reading the Whole Company’s Emails
by Frank Partnoy
The Atlantic (September 2018)
Link to article
