Artificial Intelligence still has a long way to go
Loading...
When most of us think 鈥渁rtificial intelligence鈥 our minds may turn to science fiction: the endearing android Lieutenant Commander Data from 鈥淪tar Trek: The Next Generation,鈥 or the murderous Cylons 鈥 the robots who turned on their human creators in 鈥淏attlestar Galactica.鈥
On a more mundane and 鈥 to some 鈥 more realistic level, many fear that AI will eliminate jobs currently occupied by humans. (Think driverless vehicles making human truck drivers obsolete.)
But scientists working on artificial intelligence spend more time struggling with the field鈥檚 limitations than they do exploring its possibilities, dystopian or otherwise. At least that鈥檚 the premise of Janelle Shane鈥檚 鈥淵ou Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It鈥檚 Making the World a Weirder Place.鈥 Shane, an optics research scientist and the author of the blog , is thoroughly familiar with AI鈥檚 shortcomings: her book鈥檚 title was produced by an algorithm she wrote to compose pickup lines. Drawing on her own and other researchers鈥 experiences with machine learning programs, she鈥檚 written a funny and fascinating introduction to what exactly scientists mean when they say 鈥渁rtificial intelligence鈥 鈥 and how it is and isn鈥檛 intelligent.
鈥淎rtificial intelligence鈥 is a catch-all phrase for the field of research that seeks to create smart computers. Modicums of computer intelligence manifest most easily in well-delineated tasks 鈥 the more focused the task, the better the computer鈥檚 performance. But trouble starts when a scientist programs an algorithm to do tasks that require a wealth of knowledge humans take for granted.聽
In 2016, for example, a human driver decided to use a Tesla car鈥檚 autopilot function 鈥 which is designed for highway use 鈥 on a city street. The vehicle ran into a flatbed truck while it was crossing an intersection because the autopilot could only recognize trucks coming from behind the car. Instead, the AI identified the truck as an overhead sign.
But while Shane acknowledges the very real dangers posed by the limits of AI, she focuses on the lighter side of the subject. In lively, conversational prose, she provides accessible illustrations of basic AI concepts. For example, take two of Shane鈥檚 鈥淔ive Principles of AI Weirdness鈥: AIs don鈥檛 understand the problems you want them to solve, and AIs take the path of least resistance to their programmed goal. Case in point: Shane later discusses an algorithm that was supposed to write programs that would perform basic computing tasks using minimal power. The algorithm鈥檚 solution? It created programs that were permanently asleep so they wouldn鈥檛 use any resources.
Shane occasionally returns to the darker side of AI. A commercial algorithm used by many states to recommend prisoners for parole was found to be much less likely to recommend African-American prisoners for early release, regardless of their conduct or age. That鈥檚 because the algorithm was trained using datasets compiled from the historical records of this country鈥檚 justice system, whose judges disproportionately condemned African American offenders to harsher sentences than white offenders. In other words, the algorithm recorded and repeated human bias. When AIs don鈥檛 work, sometimes the problem isn鈥檛 AI 鈥 it鈥檚 humans, Shane notes.
AI鈥檚 myriad failings stand in stark contrast to its cutting-edge cachet. Despite the technology鈥檚 vaunted promises, there are still many tasks 鈥 even technical ones 鈥 that humans do better. But because of the fanfare that surrounds AI (and the field鈥檚 ability to attract investors), companies sometimes misrepresent work done by humans as the fruits of AI. Shane points out that in 2019, an estimated 40 percent of European startups claiming to develop AI didn鈥檛 use any at all.
鈥淵ou Look Like A Thing鈥 provides an eye-opening look at a serious but sometimes oddly funny subject. And I don鈥檛 know that I could have wished for a better guide than Shane; I never expected reading about computer science to be so much fun. Ultimately, I found this book comforting: Shane makes a convincing case that AI isn鈥檛 going to be stealing jobs in the near future.