In an interview, mathematician and philosopher Rainer Muhlhoff talks about misleading data protection regulations, discrimination with poorly protected data, and the business model of social media.
Mr. Mühlhoff, when I talk about data protection, it always causes a yawning reaction in my circle of acquaintances. Why dedicate your scientific career to this topic?
For me, the topic of data protection and new digital technologies is closely related to growing social inequality. We are witnessing a rapid imbalance of power between companies and individuals. Anyone who yawns does not have enough on the screen.
But can you understand that people get upset when data protection complicates things? Just think of the endless cookie banners.
Of course, many are annoyed by cookie banners, which are also a big burden for me. But it is more than an example of bad and misguided regulation. A real disservice to data protection, as it does not result in even less data being collected because people are forced by banners to consent to data processing. But the negative impact that data protection prevents is far greater than the friction it sometimes creates.
Do you have examples of these negative effects that data protection protects us from?
Data protection is actually a wrong word. Data protection doesn’t protect data, it protects people and their fundamental rights. Since we started processing machine data, we’ve had to deal with the resulting force asymmetry. In the past 10-20 years, we have increasingly noticed that the data of many users is being combined into large databases, the so-called big data. Then it is no longer about individuals, but the information content of the data set.
And why is this bad?
With a huge amount of data, you can treat people with different characteristics differently. They can sell more expensive insurance to people who have higher health risks than they do. They can manipulate people through personalized advertisements regarding their political beliefs. You can sort people into groups and discriminate against individual groups. In the end, information about us that we will never reveal can be extracted.
How does this work? How can a company know more about me than I want to disclose?
Learning algorithms are used here. There are business models whose goal is only to collect as much data as possible from as many people as possible, for example when browsing the Internet. The algorithms then use this data as a basis for learning. For example, if there is a dataset of people with prostate cancer, the algorithm can learn how those people behave online. What they are looking for, and what medicines they are ordering online, for example against symptoms that may indicate cancer. If a person who does not appear in this data set behaves in the same way, then the algorithm can predict a certain probability that this person will also have prostate cancer.
How are these algorithms used in practice?
For example, if someone wants to buy life insurance, the insurance company can do risk scoring…
So an estimate of how high the risk is that the person will die soon and the sum insured will be due…
exactly. They can use their phone number, for example, to find a person’s social media account and use that data to let the algorithm know if the person has prostate cancer. If this is the case, the risk score gets worse and the insurance is more expensive.
Does this mean that group discrimination does not only affect people who already have a characteristic that makes them vulnerable to discrimination in our society?
yes. This does not mean that existing forms of discrimination no longer play a role. Big data and artificial intelligence, as they are also called learning algorithms, learn existing discrimination patterns from the data and then reinforce them.
to a person
Rainer Mullhoff (39) Professor of ethics in artificial intelligence at the University of Osnabrück. In addition, he is conducting research at the Free University of Berlin.
Among other things, he studied Mathematics, Computer Science, Philosophy and Gender Studies.
in his research It deals with, among others, data protection and data ethics in the context of artificial intelligence and big data. yes
But prostate cancer is primarily a disease of the “old white man”.
AI doesn’t stop at recognizable patterns. You can discriminate according to any pattern and no one can understand why. Big data can also be used to find out who has a predisposition to prostate cancer, and these are primarily older white men who, in this example, would not be able to get life insurance.
How do we protect people from this kind of discrimination?
In order to better control discrimination through learning algorithms, data protection and anti-discrimination laws must work hand in hand. It is important to note that by combining different characteristics and characteristics, algorithms can create new patterns of discrimination that we do not expect and do not yet have a name. In this way, groups that are already disadvantaged in any case often become victims of mutually reinforcing multiple discrimination.
But if there was a system that could tell me if I was at increased risk for a certain disease, that might help me, too. I can do more tests. Wouldn’t this be a useful app?
There is a difference between making predictions about a person with their consent in a doctor’s office and using them for other purposes. Of course, there are positive uses for data and artificial intelligence. When this is the case and when it is not, a good organization should make it clear. Because all too often data companies combine a seemingly beneficial and harmless business model with a harmful one. Google offers a free email service and search engine and sells your data to insurance companies and ad groups. This is how the company makes its profits.
Will politicians have to step in to prevent this action?
We have a blind spot in current regulation when it comes to algorithms that can make predictions about people. For example, companies use an anonymity loophole. The data that algorithms learn from is often impersonal. However, anonymized data is not subject to any data protection restrictions. The General Data Protection Regulation does not apply to them. Data becomes personal again only when prediction models are applied to individuals, for example in the case of insurance. In my opinion, it’s too late.
But is this really such a big deal when insurance gets more expensive?
It’s not just about insurance. It is about community and its cohesion. Predictions are not made primarily about individuals, but about groups. We have to get rid of the idea that one person should be spied on. It is about population management. We’ve already noticed this on Facebook in the past: political ads are tailored to the psychological sensitivities of users, who have been almost completely manipulated in this way. This is a danger to democracy.
Why don’t we get the regulation right?
The risks of big data and artificial intelligence are not yet well understood at the political level. Our data protection rhetoric is heavily influenced by liberalism, the idea that data protection is sufficient if everyone has individual control over their data. This fits the economy well. Industry can support data protection regulations that result in an obligation for individuals to give their consent, because these rules do not call their business models into question.
As long as we still have individual data protection laws: what can everyone do?
It is important to realize that the choices I make about how I use my data affect others. If you always click accept in the cookie banner, this also allows the algorithms to — with the help of my data — make predictions about other people being more keen on their data.