There’s an old joke about a heli-tour pilot who gets lost in the fog around Seattle. By chance, the pilot sees someone at the top of a building and yells out, “where am I?” To which the person replies, “you are 300 feet in the air.” The pilot says thanks, turns the helicopter to the southwest, and flies straight to the airport.
Thoroughly amazed, the passengers ask how the pilot knew where the airport was just based on the answer that person gave him. The pilot replied, “their answer was factually correct, but otherwise completely useless, so that must have been the Microsoft Customer Support office, and I know exactly where the airport is from there.”
Google recently posted an informative article on The UX of AI in their library, which prompted this post.
My degree program included Artificial Intelligence, Machine Learning, and Neural Network classes that have helped me include AI design in several projects over the years. To a practiced UX designer with an AI perspective it is clear that we must consider the human element when designing AI and Machine Learning systems to make them truly effective.
A common AI design mistake focuses on utilizing the available data rather than determining what data the system and users actually need. For instance, current GPS systems provide very detailed maps with very precise measurements. Driving aids provide directions based on this data, such as “turn left in 650 feet.” While we have become tolerant of that norm, it is a rather ineffective design. Most of us do not accurately know how far 650 feet is while we are driving a car looking for the correct place to turn.
We don’t base our driving knowledge on distance, we base it on time. A more Human-Centered Approach might calculate the time it will take to reach 650 feet given the current speed of the car and suggest the driver should turn right in “3, 2, 1, turn right here.” Additionally, humans rely on visible waypoints. A well designed system should suggest that the driver turn right at the gas station.
A human-centered driving AI might also stop annoying me with precise driving directions from my house to the highway. I have driven it thousands of times and don’t need turn-by-turn directions, unless there is something novel to attend to, such as road construction or an accident. Once I reach the highway, a better driving aid might merely suggest that I take the highway south for about 15 minutes and then chime in as I approach the correct exit. I don’t need detailed directions telling me to drive 17.2 miles when I’m getting on the highway.
Good UX research not only accurately identifies user needs, but also the data required to address those needs. The GPS direction system redesign I suggest cannot be accomplished with the existing data, but it’s easy to see what data will be required. Good AI design is more than just using existing data, it’s more about understanding what problem users need solving, how the solution should be presented, and determining what data is required to provide that human-centered AI solution. AI systems must include UX perspectives to evolve to the next level of effectiveness.