Me.Ai

Jun 2017 - Aug 2017

Role

UI Designer

Front End-Engineer

AI Experimenter

Tools

React

Collaborative Filtering

Sketch

Team

Sisi

Kenny

Fun Fact

None of us went into this project as AI experts

How might we use cross-domain information about a person to make better product recommendations?

Warning: This case study is more about design experimentations with AI and Digital Identity than finished visual designs.

First Thoughts

Unlike a "typical" IDEO process, we began with interviews with experts in the field of consumer recommendations, advertisement, and data collection rather than the end users. We aimed to uncover patterns and opportunities in AI that fit frictionlessly into our daily lives. Here are some learnings:

Changing From Demographics to Personas (AI Insight #1)

Very early on we found a correlation between people with similar interests, hobbies, or intangible "design appreciation" and their product preferences. Rather than recommending products from people in similar demographics, we believe a deeper understanding of unusual descriptors of a user can lead to better or at least more diverse recomendations.

Here you can see how we believe a shift from demographic buckets to more abstracted categories may be the key to actually "understanding" people's preferences online. What if my Bowers and Wilkins headphone purchase was more related to the fact that I'm an industrial designer and hip hop lover than the fact that I'm a techie aged 20.

Utilizing Amazon Mechanical Turk (Experiment #1)

The framework from above led us to test if people's perception of the personas they were a part of actually influenced product preferences.

Original Hypothesis: do the "personas" or groups that people label themselves as correspond to the products they buy? More interestingly, do two outdoorsy people share a love of not only similar hiking boots (obvious) but also perhaps similar headphones?

Using Mechanical Turk, we collected >10,000 data points on products with inconclusive but interesting results from collaborative filtering algorithms. To the left are the results from running the model with grouped individuals versus every "user." There seemed to be a trend in which narrowing categories of users down to "Deal Finders" and "Fashionistas" yielded much higher predictions of preference for a specific product.

Let's understand trust with ai (Pivot)

Based on the skillsets of the team and inconclusive data analysis from the previous experiment, we decided to push on the possible interaction principles we could put forward to address a main issue with data collection for AIs: trust.

HMW evoke trust between a person and an online AI "agent." Is there a way this can live in a personalized and curated online shop?

An updated, personalized shop (Prototype #1)

Hypothesis: These connections we found in personas can create a fully personal store with products uniquely recommended for one person.

Bring AI "Training" Into the Browser (Experiment #2)

It's weird that advertisements are areas that consumers feel unsafe clicking because of their direct links away from what we're doing! They are often irrelevant and creep us out when they know too much that we didn't explicitly tell them.

Hypothesis: people will feel compelled to create a more robust AI model if their data is transparently given to a company rather than being taken from them without consent.

I built a Chrome Extension that allows you to block ads but instead fill them with a "sandbox" to look through products that your personal store recommends. Any choices here are reflected when you next visit your curated site.

Expressing preference through Emojis (Experiment #3)

How can we make training an AI more interesting and aligned with the normal behaviors of users?

Hypotheses: People are good at eliminating tasks, rating on a non-binary scale, and choosing visually.

Fully reactive and personal "zine" shop (prototype #2)

All the insights converged in on the form of a personalized site or "zeen" for each person that takes cross-platform information that the user consents to share in order to create a safe and personal space to shop on ME.AI.

Can sites be aesthetically evolutionarY? (Provocation #1)

All the insights converged in on the form of a personalized site or "zeen" for each person that takes cross-platform information that the user consents to share in order to create a safe and personal space to shop on ME.AI.I fell in love with the idea that a site could grow over time with its user in everything from the styling (CSS) to the content.This prototype includes a subtle display of this in the top right corner. We imagine a system in which your AI agent is visualized and physical grows as you interact with it.

Sites as training Sandboxes (provocation #2)

There is a very interesting mindset shift from sites being static deliverers of content to users and instead, dynamic beings we can change and interact with. To test one aspect of this, we allow one of the large image cells to be filled with whatever the user is feeling in the moment. Not only does this give the user a sense of ownership over the space, it helps us collect data on what they feel when shopping.