Etiquette Lessons for AI
You’ve just built an AI system. What will you do if mistakes happen?
Here’s one mistake already: when! “When” mistakes happen, not “if” mistakes happen!
When B.J. May stepped out of his Georgia home on a sunny day, he experienced one of my favorite AI whoopsies. The “smart” facial recognition-based lock on his front door wouldn’t let him back in because it was protecting the residents from…
Luckily, the team that built the smart lock planned for mistakes, which is why B.J. had the option to use his pin to get back inside. But what if they hadn’t? Our (super)hero would have had a very different day indeed. And a very dark night.
When, not if
Just as responsible policymakers create plans and failsafes for human error, responsible practitioners of machine learning and AI always make sure that there’s a plan in place in case the system produces the wrong output.
The image above is a slide from one of my talks. When I ask reliability engineers what the mistake is on this slide, I’m amazed how often they ignore the directionally challenged road markings because they’re too busy hollering “when, not if!” Good. I’m with them: the if is the bigger issue. Mistakes will happen.
A good reminder for all spheres in life is to expect mistakes whenever a task is difficult, complicated, or taking place at scale. Humans make mistakes and so do machines; mathematics doesn’t trump the rules of common sense. The most dangerous mistake you can make is forgetting that mistakes will happen with AI.
When we attempt to automate complex tasks and build complex systems, we should expect imperfect performance. This is true for traditional complex systems and it’s even more painfully true for AI systems.
When it comes to AI, expecting perfection is not only unrealistic, it’s dangerous.