Many science fiction writers have written about how an AI can go rogue, typically in self-preservation. Stories like Terminator or The Matrix. Fortunately today, there is very little risk of an AI taking over the world, but there are ethical dilemmas that need to be considered when we work with Machine Learning.
Especially when the influence of individual factors is difficult to understand, the output of machine learning systems can be obtuse. Do you think Facebook can tell you exactly why they showed you an ad for shampoo? Do you think high-frequency trading agencies have full understanding of their systems?
As you are likely aware, in 2015, Volkswagen was accused of cheating on its emissions tests. Cars were found to behave differently in real-world conditions than when they were undergoing the EPA test regimen. The EPA fined Volkswagen for about $18 Billion. VW stuck with the story that a "rogue engineer" was responsible for this issue.
When he and his co-conspirators realized that they could not design a diesel engine that would meet the stricter US emissions standards, they designed and implemented [the defeat device] software [link]
Now, emissions software is actually a very likely candidate for machine learning optimizations. Many parameters can be tweaked and optimized to reduce diesel particulates to meet emissions standards. That standard, however, is a well-known test.
Let's look at an example, with a very simplified system, to show how a solution like the "rogue engineer" created might also arise with an "rogue" machine learning system.
This is egregiously simplified, but will demonstrate the concept.
Let's set up a scenario where some engineers are designing an engine system for a diesel car. They will put an EGR filter on the car to help it produce fewer NOx emissions. They use many complex equations to model the system. Let's mention a few facts that will be put into their model:
Next, a machine learning algorithm is thrown at the model, with these goals in mind:
Now, what do you think the exhaust gas recirculation rate will look like as a function of engine speed? The curve may look something like this:
Now, is the algorithm immoral? Is it "cheating"?
What if the engineers didn't actually see this plot, but just saw an obtuse set of constants in their code?
The point made above is that a machine learning system can generate behavior that might be considered immoral or "cheating" if it was intentionally designed by a person.
Note: this is based on a scenario I saw in a comment on Hacker News, but I can't find it anymore.
Let's imagine a man named Jim in the not-so-distant future. Here are a couple of facts about Jim:
Let's also imagine in this scenario that the airline he is flying on has revolutionized their customer service. They now offer personalized coupons, reminders, and notices delivered to your phone, based on advanced deep learning techniques. The airline's system looks at your Facebook likes to offer you special deals and coupons in the terminal. They even recognize your face to allow you to skip showing your boarding pass at various places, and can bill you for your food / souvenir purchases in the airport by recognizing your face.
Now, let's say Jim arrives at the terminal with about 15 minutes to spare. He is not looking forward to his 6-hour flight, and he heads over to the airport bar to grab a quick drink.
The system at the bar conveniently recognizes his face as he sits down, verifies he is of age, and bills him for his first drink without him even opening his wallet. After he sits for what feels like 5 minutes with his drink, his phone buzzes and offers him a coupon for another drink at half price. He accepts the offer and presents the coupon to the waiter and sips his second drink.
As Jim finishes his drink, he checks the time, and rushes to the terminal to board his plane. Unfortunately, he is too late, and the doors to the plane are already closed. The airline does not offer him a refund, because he was not on time to the doors. Unbeknownst to Jim, every seat on the plane was taken.
Jim blames himself for missing his flight. He is able to find a flight that leaves 2 hours later, and buys another ticket.
In hindsight, why did the system offer him a discounted drink? Why didn't it send him a reminder to board his plane when he still had time?
Machine learning algorithms have the ability to optimize for patterns that may not be obvious to humans working on the same problem. They can develop new strategies. In this case, the algorithm may have weighed a few facts, and acted in the airline's best interests:
|Jim was at the bar||recognized with camera|
|Jim likes to drink, and doesn't like flying||Jim's Facebook posts|
|Jim isn't very punctual||Jim booked his flight late|
|The flight is overbooked||Airline database|
|The airline will need to offer passengers vouchers or cash to "bump" them to a later flight||policy/regulation|
The cost of forcing Jim to take another flight would have cost hundreds of times more than a drink coupon. If the algorithm is optimizing for money, maximizing passenger throughput, and selling more products, this is one possible scenario.
Now again, is this behavior immoral? More importantly, how would anyone have known this behavior emerged at all?
This one will be brief, because I'm assuming you get the idea by now.
If a person chooses to give a 30-year, extremely high rate mortgage to somebody with absolutely terrible credit, it is predatory. But if an algorithm does it, what is it? Is it wrong? Is it illegal? Who goes to jail?
In 2015 we had Parallax scrolling. Okay, parallax scrolling still exists, but I think most people agree it's tacky and overused. Let's move on.
Speaking as a user, this is my official list of irritating web design trends in 2016.
Example: Smoothwheel Library
Yeah, smooth scrolling was terrible a few years ago, but it still hasn't gone away. Why??? I want my browser to work the same way no matter which page I go to. And implementations of smooth scrolling are usually painfully slow, making this even worse. Stop it!
Example: Apple Watch
Apple's watch page isn't the worst example, but demonstrates the idea.
This is similar to my problem with smooth scrolling. When I hit the scroll wheel on my mouse, I expect the page to scroll down slightly. This is everybody's expectation. I don't want it to jump down an entire screen. Even worse, if I discover that your site works like this, I expect it to scroll consistently, in other words, move once per mousewheel stop. Inevitably doesn't work this way, and the page scrolls twice when I wanted it to happen once, or it doesn't scroll at all with one motion of the mousewheel.
The way Facebook does infinite scroll works. It loads more content as you scroll, and doesn't remove content above you. Unfortunately, this isn't how infinite scroll is always implemented. Sometimes content is removed from the top of the page. This is infuriating when I want to just go back to the top of the page, either to get to the nav bar or see the early content
I am somebody who sometimes highlights text when I read a blog, news story et cetera. It's how I sometimes keep track of my place if I have to look away, or sometimes I just do it out of habit. What it doesn't mean is that I want to share a quote on Twitter, facebook, or anything.
This trend has even extended to include highlights of "most tweeted" sections of the article. But this is mostly on garbage trendy news sites anyway, so I don't run into it that often. I pray it doesn't spread.
Example: Node.js Docs
The progressive enhancement / graceful degredation / responsive design, whatever you want to call it trend is a good idea. Scratch that, it's a great idea. Make your webpage work on multiple different screen sizes. But that doesn't mean that a narrow viewport is a phone.
Windows, MacOS, and most linux distros have a viewport snapping feature which lets you snap a window to half of your screen. This is really useful for multitasking. But I don't want the mobile version of your website when I want to multitask.
Yes, even the nodejs docs are victims of this plague. Visit that link and make your browser about 800 pixels wide. The font size explodes and it becomes a chore to read. Move it over the threshold, and it becomed readable again.
The proper way to do scale-up on mobile is via the viewport meta tag, not a CSS media query.
Hamburger buttons do work fine with a mouse. It's a perfectly acceptable way to provide a menu on a narrow viewport on the desktop. But don't change your whole experience to blow up your icons