Machines can get it wrong, too

Faster, more effective and perhaps even smarter than humans: artificial intelligence often causes concern. Is this justified?

When Professor Olaf Landsiedel asked during the "Night of the Profs" in mid-November, who feels discomfort at the thought of artificial intelligence, around 50 percent of those present in the jam-packed large lecture hall in the Audimax raised their hand. This highlighted two facts: there seems to be all kinds of uncertainty regarding the topic, and the interest in it is very, very big.

Whether it is for self-driving cars or nursing robots, according to Olaf Landsiedel, artificial intelligence is viewed "significantly more critically" in Germany than in many other countries. "The risks are assessed as greater than the benefits," is how the computer scientist put it, but in his view this is much too pessimistic. After all, according to the scientist, artificial intelligence is by no means an uncanny power, but rather a phenomenon which follows a fairly simple principle: learning on the basis of a very large volume of data. If a computer is supplied with enough images, it can soon independently distinguish cats from dogs, with a high degree of accuracy. However, artificial intelligence can do much more, and to some extent it really has become detached from people. For example, it can develop the algorithm (i.e. the method) by which cats are identified completely by itself. In doing so, it draws upon a neural network, which - albeit not entirely, but certainly in principle - is modelled on a digital replica of the human brain.

This is in no way new, emphasised Landsiedel’s colleague Professor Reinhard Koch. As far back as 50 years ago, the postal service used artificial intelligence for automatic letter sorting, and fed handwriting pixel by pixel into a database. Today this system works with an accuracy rate of 99.8 percent, and thus is more accurate than staff made of flesh and blood. However, one thing remains true: the more differentiated the task, the harder it is for artificial intelligence. Although computers can learn to distinguish between breeds of dog quite well, based on snout, fur and other features, according to an experiment quoted by Koch, they can only identify a collie with 77 percent accuracy. 20 percent of the matches were an Alsatian (German shepherd), and three percent were actually a cat. "To err is human, and this also applies to machines," said Koch succinctly.

However, in his assessment artificial intelligence can be a useful form of support, and not only for sorting letters. For example, image analysis for mammography screening is, in his words, "pretty accurate", and thus a valuable addition to the human eye, which also has its own weaknesses when reviewing large series of images. But with regard to autonomous driving, the Kiel computer scientist is more sceptical. Even though it is certainly possible that sensors in conjunction with software can identify pedestrians, cyclists or other drivers, as well as buildings or other objects, the reality of unexpectedly configured roadworks and detours, and also of irrational human behaviour, is often still too complex to be able to speak of fully functional and reliable systems.

Many knowledge gaps in artificial intelligence are expected to be closed in the short or long term, so that the topic will soon be part of everyday life. This is certainly the opinion of Professor Agnes Koschmider, who also researches and teaches at the Department of Computer Science. Of course, she emphasised, rules for dealing with artificial intelligence must then also be determined. "But this is nothing new," she said, and pointed out that legal experts already debated whether cars possess morality around 100 years ago. "As we did then, we will find laws and regulations in order to adopt the new technology," prophesied the business information technology expert, who - like her colleagues - sees no reason to fear artificial intelligence. As such, reliable estimates predict that around 25 percent of all jobs could be replaced by artificial intelligence, but equally reliable estimates predict that more new jobs will arise through this form of digitalisation than old ones which will disappear. Aside from this, in Koschmider’s opinion there is no escaping artificial intelligence because of its significant beneficial effects. If doctors have more time for patients again, thanks to digital assistance, this is in everyone’s interest. Last but not least, it is a reason why Kiel University has joined forces with the universities of Bremen, Hamburg and Lübeck, to form a competence centre for artificial intelligence in medicine.

Author: Martin Geist