Josh Clark | A.I. is your New Design Material | UI Special, CSS Day 2019

Summary

The speaker discusses how machine learning is becoming a new design material that designers must understand and thoughtfully apply. He explores strengths like detecting patterns and surfaces four opportunities: being smarter about existing questions, asking new questions, tapping new data sources, and revealing invisible insights. However, machine learning has a strange, probabilistic nature and biases towards reinforcing normal. Therefore, designers should approach it with humility, set proper expectations around its capabilities, anticipate failures, and clarify their intentions and values. If used intentionally as a design material, machine learning can free designers from repetitive tasks and amplify human creativity, insight, judgment and compassion.

Takeaways

• 😀 The promise of machine learning is detecting patterns in anything and then acting on them.
• 👩‍💻 Machine learning can enhance existing products through small interventions like predictive keyboards.
• 🤖 The machines have a strange, weird perspective that can reveal unexpected connections.
• 🌀 Machine learning is probabilistic - it deals in shades of gray and confidence scores.
• 😟 Historical data and notions of 'normal' can reinforce bias - but surfacing it lets us act.
• 💡 Machine learning allows us to ask smarter questions, new kinds of questions, and tap new data sources.
• 🙋‍♂️ The goal should be amplifying human potential by taking over repetitive, error-prone tasks.
• 🐶 Systems should fail endearingly, like puppies - not dangerously, like babies.
• 🔎 Present suggestions and signals, not definitive one-size-fits-all answers.
• 🤔 As designers of these systems, we must be intentional about the values and impacts.

Q & A

What are some everyday examples of how machine learning is currently being used?

Some everyday uses of machine learning include predictive keyboards on phones, recommendation engines like on Netflix, and semantic searches like on Slack to find experts within an organization.

How can machine learning help amplify human skills and capabilities?

Machine learning can take over repetitive, detail-oriented tasks from humans to let them focus on more creative, insightful work. It can also uncover insights and patterns that were previously invisible to humans.

What are some unique strengths and weaknesses of machine learning systems?

Strengths include the ability to detect subtle patterns in large volumes of data that humans miss. Weaknesses include the tendency to come up with strange, unrealistic results at times and being overly confident in incorrect predictions.

How should designers set expectations when working with unpredictable machine learning systems?

Designers should design for a range of fuzzy results rather than just success. Interfaces should match the system's actual capabilities and convey uncertainty using language like "this may be" instead of definitive statements.

How can machine learning reinforce societal bias?

Machine learning models can absorb implicit biases in training data sets. But surfacing those biases also presents an opportunity to address them.

What are some machine learning services designers can start using today?

Services like Microsoft Cognitive Services, IBM Watson, Amazon SageMaker, and Google Cloud AI provide ready-made ML capabilities like speech recognition and computer vision for designers to easily incorporate into prototypes and products.

How might machine learning change the focus and priorities of design work?

It could shift more production work like assembling known interface patterns to machines, freeing up designers to focus on solving new problems and areas requiring human creativity.

What are some ways designers can direct machine learning technology towards beneficial ends?

Designers have an opportunity to influence their companies' intentions, values and business models around machine learning rather than just optimizing for efficiency. They can shape products to amplify human potential.

How can interfaces convey the uncertainty in machine learning predictions?

Interfaces can order predictions by confidence, include percentages for each prediction, suggest multiple low-confidence possibilities, and use hedging language like "this may be" instead of definitive statements.

What is an example of machine learning revealing bias that presented an opportunity for positive change?

Amazon discovered their AI recruiting tool was biased against women. While troubling, it allowed them to reevaluate and address gender bias in their hiring practices.