Who made algorithms before computers

: Computer: How algorithms determine our lives

Berlin - The bank refuses a loan despite the best credit rating, Amazon persistently suggests books that one would never read, and the full-body scanner at the airport finds something unusual on the body. None of the friendly officers can explain that, but it entails an extensive security check and a sprint to the gate. All of these are the consequences of computer decisions made on the basis of algorithms. Everyone knows them from everyday life, but very few are aware that there are arithmetic rules behind them that are increasingly determining our lives.

For example, can a computer program influence people to change their voting behavior? That sounds like something out of a science fiction film. But researchers actually warn against the ranking algorithms of Google and Facebook, which show different users different results and, according to current studies, can actually change political election results. The US presidential election in 2016 could be the first election in which that happens, according to fears.

Even elections could be influenced

Facebook itself has shown in several studies that only tiny changes in the algorithm change the opinion of users about the shared content of others. According to a study, the group could also influence voter turnout. During the 2010 congressional election, Facebook sent 61 million users a message reminding them of the election and showing how many of their friends had already voted. Those users voted significantly more often than the control group who did not receive a corresponding message.

If the election result is tight, such manipulation could be decisive for the election, warns the American technology sociologist Zeynep Tufekci: “Facebook can influence elections - without our being able to prove it.” After all, the influence is so subtle that it is neither for the individual nor for the courts or Regulatory authorities is understandable.

With the search engine Google, too, a complex calculation rule in the background decides which user gets which results. Studies by US psychologist Robert Epstein show how a change in the selection of search results made voters more likely to decide in favor of one or the other candidate.

Many algorithms are secret

Epstein used the 2014 elections in India for an experiment. He recruited more than 2,000 undecided voters and divided them into two groups. During their Internet research, one group received more search results about one candidate, a second group more about the other candidate. The result: The preferences shifted by an average of 9.5 percent in each case in favor of the candidate who was preferred in the search results.

Exactly which factors flow into the ranking algorithms of Google and Facebook and how they are weighted is a well-kept secret. The great influence of the Like button, with which users can mark messages with “Like”, is well known on Facebook. Various experiments prove this influence. For example, the designer Elan Morgan decided to ignore the Like button for two weeks - and was thrilled.

"That made Facebook clearly better," she said. She was shown much fewer things that she did not want to see. Because apparently the algorithm had misunderstood some things. For example, if she “liked” a photo of cute kittens, she would also receive photos of cats that had been abused by animal abusers. "The algorithm does not understand the many political, philosophical and emotional shades of a topic," says Morgan.

Fight crime before it happens

In various German federal states there are pilot tests with predictive policing. Berlin recently announced that it would use a computer system to reduce the number of break-ins. The software is called KrimPro, an in-house development from Berlin. The aim of such programs is to make the computer act like a seasoned cop who, over the course of a long career, has developed a feel for where criminals will strike next. Most federal states use the German software Precobs, which was developed by a family company in Oberhausen, the Institute for Pattern-Based Forecasting Technology IfmPt.

It is based on the assumption that professional perpetrators follow certain patterns. They act systematically, preferring areas in which they are not noticeable and which offer good escape opportunities. They target specific prey - and come back when they're successful. This is called "near repeat" in technical terms.

A perpetrator visits the crime scene a second time. The computer predicts such offenses based on a pattern recognition algorithm. The basis is the data from past break-ins. If the system is 70 or 80 percent secure, it sounds the alarm - and the responsible authorities decide whether to send a patrol over.

Success is hard to prove

However, the media repeatedly report cases in which such software has hit the ground, police officers are on the spot more and more patrols - and still have a break-in. These may be isolated cases, but the question of whether the break-ins generally decrease as a result of such measures can hardly be reliably assessed. One hears again and again about areas in which the burglaries are decreasing - and in the following year they increase again.

"That is the absurd thing about the whole story," says Michael Schweer from the IfmPt institute, "the non-occurrence of the secondary offense is our success, but it cannot be proven by non-occurrence." Nevertheless, Schweer is convinced that Precobs works : “We are now at seven locations, and the dips at all are significantly below the comparison values. It would be strange to say: It has nothing to do with the system. "

Other federal states work with their own algorithms. The LKA North Rhine-Westphalia is currently testing which data can make its own software even more accurate - such as real-time information on power consumption that shows where few people are at home. Something like that could attract burglars.

Racism instead of fair judgment

However, recent incidents in the USA show how quickly pattern-based suspicions can backfire. There, computers calculate the likelihood that a criminal will recidivist. The result affects his prison term. Research by the investigative journalists' office propublica recently showed that the underlying algorithm systematically disadvantages blacks. A bug in the system?

As a result, those responsible remained shockingly speechless - apparently the self-learning system had drawn its own conclusions, which can now hardly be checked. Researchers had repeatedly warned of such a problem. Some therefore call for algorithms not to be able to make decisions about people that would cause these disadvantages.

However, Michael Schweer is annoyed that cases occur again and again in which corresponding pattern recognition algorithms discriminate against dark-skinned people, for example. “That mustn't happen,” he says. Precobs is consciously not a self-learning system that draws its own conclusions, but is only expanded by people. “We want to avoid mistakes creeping in. Self-learning systems like to fall for false correlations. "

No money because of low scores

Katharina Zweig, head of the working group on graph theory and the analysis of complex networks at the Technical University of Kaiserslautern, is also annoyed about the lack of theory in automated pattern recognition, on which a large part of artificial intelligence algorithms are based today. The Schufa also calculate the creditworthiness using algorithms that incorporate various information. In the end, the bank employee only receives a value and the advice not to give this or that customer a loan, says Katharina Zweig.

But what should the bank employee say to those who cannot get a loan? She asks. “You have this value. Why, I don't know, in any case you won't get a loan. ”From their point of view, it should be possible to tell those affected what they can change in order to get a loan and why the algorithm assigned them this value. But what if the users of a system cannot know that themselves? What if the result is wrong? In order to make people more aware of such problems, Zweig recently founded the organization “Algorithm Watch”.

It's also about colleagues. "The biggest problem: We as computer scientists are not trained to model," says Katharina Zweig. Modeling means, for example, deciding which data are relevant as training data and which type of algorithm is applied to them. Some time ago, the Schufa announced a controversial project together with the Hasso Plattner Institute to use data from social networks to predict the creditworthiness of individuals, says Katharina Zweig.

Facebook presence decides on loans

This project is based on a model, namely the idea that such data could in principle be suitable for it. In case of doubt, which algorithm to use on it will be decided based on which one brings the best result, for example finding the uncredited and not falsely rejecting too many creditworthy. After all, the researchers know from their collaboration with Schufa who is rated as creditworthy.

An algorithm then searches for commonalities based on the data that those affected have shared on Facebook. But how do you know that these are not random correlations? And is it legitimate to deny people a loan on this basis? The project was discontinued after public protests.

But the whole thing also has a bigger, statistical problem that is often overlooked - even by computer scientists: Assuming that in the end you would find an algorithm that uses Facebook data to predict 90 percent correctly who the Schufa would not consider creditworthy , and who falsely classifies only five percent creditworthy as not creditworthy. At first glance, it looks like a pretty good result. One might get the idea that Facebook is a good source of data for calculating people's creditworthiness.

Too many people are being sorted out

But there is a trap, warns Zweig. One has to consider the proportion of the non-creditworthy in the population. Few did not pay back their loan in the end. Suppose that out of 5,000 people 150 did not repay a loan. The algorithm would identify 90 percent of these, i.e. 135. Extrapolated, however, due to the apparently quite accurate algorithm, a good 240 other people would not get a loan - 5 percent of the 4850 who are actually creditworthy. “This means that the hit rate for everyone who the algorithm does not consider creditworthy is only around 36 percent.

Such percentages are therefore a problem in Artificial Intelligence when the categories to be predicted are very unbalanced, "warns Zweig. And artificial intelligence is often used when little is known, for example in the case of rare diseases. But it was precisely then that prediction errors had a major impact: "You sort out too many or you can't find those people who actually belong."