The UX of AI

Google Clips is an intelligent camera designed to capture candid moments of familiar people and pets. It uses completely on-device machine intelligence to learn to only focus on the people you spend time with, as well as to understand what makes for a beautiful and memorable photograph. Using Google Clips as a case study, we’ll walk through the core takeaways after three years of building the on-device models, industrial design, and user interface—including what it means in practice to take a human-centered approach to designing an AI-powered product.

Just getting more UXers assigned to projects that use ML won’t be enough. It’ll be essential that they understand certain core ML concepts, unpack preconceptions about AI and its capabilities, and align around best-practices for building and maintaining trust. Every stage in the ML lifecycle is ripe for innovation, from determining which models will be useful to build, to data collection, to annotation, to novel forms of prototyping and testing.

We developed the following truths as anchors for why it’s so important to take a human-centered approach to building products and systems powered by ML:

• Machine learning won’t figure out what problems to solve. If you aren’t aligned with a human need, you’re just going to build a very powerful system to address a very small—or perhaps nonexistent—problem.

• If the goals of an AI system are opaque, and the user’s understanding of their role in calibrating that system are unclear, they will develop a mental model that suits their folk theories about AI, and their trust will be affected.

• In order to thrive, machine learning must become multi-disciplinary. It's as much–if not more so—a social systems challenge as it's a technical one. Machine learning is the science of making predictions based on patterns and relationships that've been automatically discovered in data. The job of an ML model is to figure out just how wrong it can be about the importance of those patterns in order to be as right as possible as often as possible. But it doesn't perform this task alone. Every facet of ML is fueled and mediated by human judgement; from the idea to develop a model in the first place, to the sources of data chosen to train from, to the sample data itself and the methods and labels used to describe it, all the way to the success criteria for the aforementioned wrongness and rightness. Suffice to say, the UX axiom “you are not the user” is more important than ever.

Three ways human-centered design elevates AI

Addressing a real human need

This year, people will take about a trillion photos, and for many of us, that means a digital photo gallery filled with images that we won’t actually look at. This is especially true with new parents, whose day-to-day experience is full of firsts. During moments that can feel precious and fleeting, users are drawn to their smartphone cameras in hopes of capturing and preserving memories for their future selves. As a result, they often end up viewing the world through a tiny screen instead of interacting using all their senses.

What if we could build a product that helped us be more in-the-moment with the people we care about? What if we could actually be in the photos, instead of always behind the camera? What if we could go back in time and take the photographs we would have taken, without having had to stop, take out a phone, swipe open the camera, compose the shot, and disrupt the moment? And, what if we could have a photographer by our side to capture more of those authentic and genuine moments of life, such as my child’s real smile? Those moments which often feel impossible to capture even if one is always behind the camera. That’s what we set out to build.

Guiding the intelligence

When we started the process, the most pressing question was: if people take tons of photos but don’t actually want to go back and curate them, how will we label ground truth? This is where the foundational “HCML exercise” was born: Describe the way a theoretical human “expert” might perform the task today. The theory was twofold: First, if a human can’t perform the task, then neither can an AI; second, by diving deep into the methods of an expert, we can find signal-to-guide data collection, labeling, and component model architecture.

The closest approximation I could think of was a wedding photographer, so I set out to interview and hire contractors using a sufficiently ambiguous job posting (secret project! photography!). We ended up discovering—through trial and error and a healthy dose of luck—a treasure trove of expertise in the form of a documentary filmmaker, a photojournalist, and a fine arts photographer. Together, we began gathering footage from people on the team and trying to answer the question, “What makes a memorable moment?”

Building trust

The starting point for our work was an assumption that we could ‘show’ the model the stuff we thought was beautiful and interesting, and it would, y’know, learn how to find more. We had romanticized conversations about depth of field, rule of thirds, dramatic lighting, match cuts, storytelling … but what I learned was that we should never underestimate the profound human capability to wield common sense.

These early experiments exposed crucial technical and methodological gaps that helped us reassess our assumptions about what the product could realize, as well as take stock in the unprecedented nature of the work. We shifted our paradigm from putting ML on a pedestal to understanding that it can only learn effectively under quite reductionist framings. Basically, we were trying to teach English to a two-year-old by reading Shakespeare instead of Go, Dog. Go!. This was where the myth of the AI ‘monolith’ crashed hardest for me; the idea that there’s some singular ‘intelligence’ that understands all things and can generalize and transfer knowledge from context to context. Nope. Not even close.

Back to basics

Consistency is the name of the game when trying to teach anything. It’s why we wait as long as possible to unleash the madness of O-U-G-H (e.g. tough, through, thorough) on children when teaching them how to read and speak English. Spelling and pronouncing words like cat, bat, and sat, with their predictable “at” sounds, is so much more consistent!

With consistency comes confidence. Think about how quick—and eager—most students are to point out incongruity when a teacher provides two examples that don’t seem to line up. Algorithms provide no such feedback. As far as an algorithm is concerned, everything they’re shown is of equal value unless directed otherwise. For Clips, that meant we not only needed consistency between examples, but also within each example. Every individual frame needed to be representative of the specific prediction we’re trying to teach it to make. And often that can come in the form of teaching it what to ignore.

Capture

We needed to train models on what bad looked like: hands in front of the camera, quick and shaky movements, blurriness.

Composition

We needed to train models about stability, sharpness, and framing. Without careful attention, a face detection model will appreciate a face at the edge of the frame just as much as one in the center.

Social norms

Familiarity is such a cornerstone of photography. You point a camera at someone and they offer implicit consent by smiling or posing. Moreover, you’re the one looking through the viewfinder framing and composing the shot. With an autonomous camera, we had to be extremely clear on who is actually familiar to you based on social cues like the amount of time spent with them and how consistently they’ve been in the frame.

Editing

Diversity and redundancy is something we take for granted in the way we shoot photos; there’s a little voice in the back of our head saying, “You haven’t seen anything like this!” Or, “You’ve got enough shots of your kid for now, relax.” But our models needed a lot of help.

We approached diversity along three different vectors:

• Time: The simple value of time passing is an important signal to appreciate. Don’t go too long without capturing something.

• Visual: Subtle or dramatic changes in color can tell a lot about changes in environment and activity. Try to capture moments that have distinct aesthetic qualities.

• People: Are you in a big group or a small group or alone? Understanding how many different familiar faces you’re encountering is a crucial part of feeling like you haven’t missed important moments.

Trust and self-efficacy

One of the reasons we invested in Clips was because of how deeply important it was to demonstrate the importance of on-device and privacy-preserving machine learning to the world—not to mention its remarkable capabilities (e.g. it uses less power, which means devices don’t get as hot, and the processing can happen quickly and reliably without needing an internet connection). A camera is a very personal object, and we’ve worked hard to ensure it—the hardware, the intelligence, and the content—ultimately belongs to you and you alone. Which is why everything—and I mean everything—stays on the camera until the user says otherwise.

Concept budgeting

With an eye on trust and self-efficacy, we were also very intentional in the way we approached UI design. At the start of the project, that meant working through a few of our own funny assumptions about how “out-there” an AI-powered product needed to be.

When we reach into our brains for future-tech reference points, many designers will jump to the types of immersive experiences seen in movies like Minority Report and Blade Runner. But just imagine of how crazy it’d be to actually explain something like the UI in Minority Report to users: Here, just extend your arm out, wait two seconds, grasp at thin air, then fling wildly to the right while rotating your hand counter-clockwise. It’s easy! Almost every sci-fi faux UI is guilty of something similar; as if the complexity of an interaction model needs to keep pace with the complexity of the system it’s driving. But that’s sort of where we were for awhile during our early design phase, and we got away with it in large part for three reasons:

• We were showing people fake content in an obviously simulated environment, where they had no real connection to the imagery. Note that this issue isn’t unique to AI; it’s often one of the confounding factors when you bring people into the usability lab.

• We were surrounded by people every day who were all speaking the same language; thinking deep thoughts about AI-enabled futures. We were making the mistake of losing touch with the reference points that everyone else would bring to the table.

• We thought our new designs were super cool, so we gave ourselves a healthy amount of forgiveness when people didn’t immediately get it.

Over time, we snapped out of it. We began fiercely reducing complexity in the UI, and made control and familiarity cornerstones of our experiential framework. We added a software viewfinder and a hardware capture button to the camera. We made sure that the user had the final say in curation; from the best still frame within a clip to its ideal duration. And we showed users more moments than what we necessarily thought was just right, because by allowing them to look a bit below the ‘water line’ and delete stuff they didn’t want, they actually developed a better understanding of what the camera was looking for, as well as what they could confidently expect it to capture in the future.

Through this process we discovered another critically important finding for testing an AI-powered product: fake it till you make it. If forced to choose, it’s leaps-and-bounds more useful to prototype your UX with a user’s real content than it is to test with real ML models. The latter takes an incredibly long time to build and instrument (and is far less agile or adaptive than traditional software development, so it’s more costly to swing and miss), while the former affords you genuine insights into the way people will derive value and utility from your (theoretical) product.

In the context of subjectivity and personalization, perfection simply isn’t possible, and it really shouldn’t even be a goal. Unlike traditional software development, ML systems will never be “bug-free” because prediction is an innately fuzzy science. But it’s precisely this fuzziness that makes ML so useful! It’s what helps us craft dramatically more robust and dynamic ‘if’ statements, where we can design something to the effect of “when something looks sort of like x, do y.” And in that departure from rigid logic rules, we also needed to depart from traditional forms of measuring engagement. Success with Clips isn’t just about keeps, deletes, clicks, and edits (though those are important), it’s about authorship, co-learning, and adaptation over time. We really hope users go out and play with it.

Designing with purpose

By re-orienting the conventional AI  paradigm from finding ways to make the machine smarter, to exploring ways to augment human capability, we can unlock far greater potential in machine learning. It can become a tool for unprecedented exploration and innovation; a tool to help us seek out patterns in ourselves and the world around us. As human-centered practitioners, we have a tremendous opportunity to shape a more humanist and inclusive world in concert with AI, and it starts by remembering our roots: finding and addressing real human needs, upholding human values, and designing for augmentation, not automation.

The role of AI shouldn’t be to find the needle in the haystack for us, but to show us how much hay it can clear so we can better see the needle ourselves.

For more on Google’s approach to UX for AI, check out our full collection of articles.