Member-only story
Designing probability, little known secrets to great AI experiences
AI doesn’t need to work every time, just enough times to be valuable.
AI is hyped as superhuman, but 85% of AI projects fail to deploy.
Multiple reasons: (1) Media hype, misleading people to envision incredible things AI cannot reasonably do, (2) choosing the wrong problems for AI to solve (3) AI literature is mostly technical, think GANs, BERT, QML, reinforcement learning, self-supervised learning… there’s little about capabilities AI can do for people: detect objects, predict trends, generate images, make suggestions… (4) biases & unintended harms, (5) AI just doesn’t work well enough, confidence scores too low to be useful.
More on steps to keep AI projects on track, and foundational challenges
But AI doesn’t need to work every time for great user experiences. It just needs to work enough times to be valuable. Even moderately performing AI systems can be delightful, when the below points are considered & built-in:
(1) Fallbacks,
(2) Gaps & Overlaps,
(3) Expectations & New Directions,
(4) The Mundane,
(5) The Magical
Fallbacks
Help people achieve desired outcomes regardless of methods getting there.
- Face ID doesn’t work every time but feels like it does, because of built-in fallbacks. When it fails, the system tries again, then again. If still fails, enter your password manually. Access granted.
- Speak to type is a low-risk example. All your exact words are rarely captured. When not, say it again, cancel and redo, finish your thought in speech then type to edit, or switch to manually type.
- Self-driving cars, autonomous robots etc. rely on sensors to perceive surroundings and make decisions. Systems are designed with sensor redundancy for fail-safe. If one sensor fails, signals come from other sensors. If the system isn’t confident enough, control is handed back to the driver.
People ultimately care about outcomes, and tend to forget methods to achieve those outcomes, unless they were either super…
