Even the undisputed (human) world champion in chess, Norway’s great Magnus Carlsen, would not stand a chance in a match against any of the best AI-based chess engines. This fact has led some to foresee a development where machines rapidly outperform humans, soon making most human job functions redundant. This is not at all the case, according to Thomas Bolander:
“There is no doubt that AI will have a profound impact on many aspects of our lives, jobs, and society as a whole. However, even though the goal of AI is to simulate aspects of human cognition, machine intelligence is fundamentally different from human intelligence and has a very different set of strengths and weaknesses. Therefore, many functions can still be solved by human cognition only.”
The ability to play chess serves as a perfect example of what AI excels at, illustrates Thomas Bolander:
“Chess is extremely well-defined and clearly delimited. There are only a few and strict rules to obey, and there is a very precisely formulated goal: to mate the adversary king. For a human being, chess has an extreme combinatorial complexity, but modern computers are far from overwhelming.”
Why traffic is more complex than chess
So then, where does silicon intelligence get into trouble?
“The less well-defined the task at hand is, the more difficult will it be to solve it through AI. Take automated cars. As in chess, there are rules to obey in traffic, but there are many more rules than in chess, and they are much less formally specifiable.”
Things get even more tricky when the machine is not only required to make the best decision but also explains the reasoning behind it. “For instance, machine learning is a relevant tool when a bank is to consider an application for a loan. The problem here is, that bank customers are in their good right to expect an explanation of the decision made, but this is what many of the algorithms currently used can’t give.”
Human-machine interplay will be key
In machine learning, the program is typically trained by looking at historical data. Here, the machine will often outperform humans in finding patterns – for instance regarding which bank customers failed to pay back their loans. However, this introduces another kind of problem.
“If historically, all bad customers shared the same last digit of their phone number, and no good customers had that number, the algorithm would then give a low scoring to any new customer with that digit. Obviously, a human bank employee would dismiss the correlation as irrelevant, but an algorithm doesn’t have the same overall perspective and would act on correlations alone.”
This does by no means imply, that machine learning is unfit for the purpose but highlights how we to an ever-higher degree will see tasks performed jointly by humans and machines. The challenge here is to strike the right balance, where the respective strengths of humans and machines are applied.
“So, when trying to predict which human competencies are needed in the future, we need to think about which of our tasks are least well-defined, least clearly delimited and least repetitive. Since linguistic and social intelligence are very hard problems for AI, these might become the most important human competences of the future. Interestingly, this could even be the case in technical areas like engineering!”
Recently, Thomas Bolander was awarded the prestigious H.C. Ørsted Medal for his extensive efforts in communicating AI research to the wider public. The participants at the AI track at High Tech Summit 2019 will be able to enjoy Thomas’ communicative skills first hand.
Want to know more about Thomas Bolander and the H.C. Ørsted award? Click here.
“It’s no use closing your eyes to the changes. We need to constantly keep up and try to influence the process.”
Professor Jan Madsen, DTU Compute