Using Current and Precise AI Vocabulary is Key to Success
I’ve been throwing around the term “AI” much too loosely. As someone who spends considerable time scrutinizing words, diving into semantics, and communicating the ideas of experts, I was struck by a recent conversation with Dr. Anthony Scriffignano, Senior Vice President and Chief Data Scientist for Dun & Bradstreet.
One of his many mandates is to bring clarity and leadership to organizations who are struggling to develop a cohesive big data and AI strategy. As it turns out, many of these problems stem from our expectations and misconceptions based on how we understand (or fail to understand) the term “artificial intelligence.”
Scriffignano mentioned that one of the underlying sources of confusion about AI is our tendency to apply the term “artificial intelligence” without discretion, often using it inaccurately and incorrectly. In fact, the term AI more accurately refers to what would be described as augmented intelligence.
The Evolving Definition of Artificial Intelligence
“The problem with the terms that we use to describe the field that is largely called ‘AI’ or ‘artificial intelligence’ is that those terms are old, and the field continues to advance at an ever-increasing rate,” he said. Scriffignano went on to note that many of the early definitions of artificial intelligence involved humans: serving humans, mimicking human intent, projecting human behaviour, behaving ethically.
“Those are all anthropomorphic terms,” said Scriffignano. “When we anthropomorphize, we project the properties of ourselves onto our machines. The reality is that what’s going on inside those machines is as inhuman as you can get. Take the term ‘machine learning,’ for instance: a lot of machine learning is just regression.” He explained that as a result, those algorithms – quite powerful in the right context – fail at the simplest human tasks. “But you also have other methods that are what’s called neuromorphic methods; methods that are intended to behave more like how we understand the brain to behave.” These types of algorithms can perform sophisticated “human tasks” such as recognizing objects and faces, driving a car, or flying a drone and can easily defeat top human chess players.
How Neuromorphic Algorithms Shape AI
The problem with those new methods, however, is they’re less explainable, even though they produce better results. “And the new methods often do produce better results for more complex problems, and we like them more,” said Scriffignano. Unfortunately, the very complexity that allows these newer algorithms to work makes it difficult to understand what’s happening inside them. We do not know why a neuromorphic algorithm reached a conclusion. This lack of understanding is a relatively minor problem when we do not understand why a chess program made a brilliant yet entirely unpredictable (to a human) move; it’s a much larger problem when we do not know why a driverless car hit a pedestrian.
The Benefits of Augmented Intelligence
The term “augmented intelligence” is starting to be used for applications where users do not have to give up their thinking, intent, or morals to the algorithm. Instead, the algorithms help the user do something.
With augmented intelligence, “the idea is that you’re still the intelligent one,” said Scriffignano. “Your intelligence will be augmented by something that can read everything that was published in the last hour, which you can’t. Or it can look at all the things you’ve done in the past and help advise you do to something similar to or different from what you’ve done in the past – which you really can’t [on your own] because you have all kinds of human bias in recollection.”
Augmented intelligence will help users do things better. “So, I would not say that it’s just a more in-vogue term than artificial intelligence. I would say it’s a nuanced approach to artificial intelligence that applies in certain situations,” said Scriffignano.
To read more about Scriffignano’s thought leadership on AI, check out this article on AI World 2018.