How smart is Edward Snowden

Michael Strube was impressed when he read the study. A colleague had given the children's book "Alice in Wonderland" to a piece of software to read, which then extracted the network of relationships between the characters in the story from the text alone! However, the 51-year-old computational linguist from the Heidelberg Institute for Theoretical Studies was a little irritated when he looked up who had financed the study: It was the Defense Advanced Research Projects Agency (Darpa), the powerful research agency of the US Department of Defense.

The study was an example of the so-called dual-use problem, the fact that the same research can be used for civil as well as military purposes. High-speed centrifuges can be used in pharmacology or to enrich uranium for atomic bombs. Algorithms can fully automatically clarify the relationships between Alice, the March Hare and the mad hatter - or read out from real mail traffic which terrorists are in which relationship to one another. The children's book study shows what artificial intelligence can do today: It can capture the content of texts and draw conclusions from them - more or less reliably, for good as well as questionable purposes.

"How the NSA got so smart so quickly"

Strube still remembers how the problem first suddenly became clear to him. On June 9, 2013, ten days after the Snowden revelations, he was at a conference in the United States when he happened upon a headline in the Wall Street Journal found on the breakfast table: "How the NSA got so smart so quickly". It was about the question of how the largest foreign intelligence service in the USA had been able to collect and analyze its masses of data.

The article explained in detail how Strube's discipline had made the spies strong: With so-called Natural Language Processing (NLP), computers learn to understand human language in terms of content. Most people have to do with this technology when using services like Google Translate or operating their smartphone using voice commands. The dark side is less obvious: "The public knows that the secret services access metadata," says Strube, that is, data such as the sender or subject of an email. "But very few people know how well we can analyze unstructured data." Human language, for example, is unstructured and has long been an opaque thing for machines. And even if you can already draw a lot of conclusions about a person and their environment from the metadata: If you can automatically evaluate the content, you will of course find out a lot more.

Suddenly, the conference seemed secondary to Strube, which, among other things, dealt with how the algorithms could be made even more precise. Doubts gnawed at him: is my work really only good for society? He spent the day in front of the doors more than in the lecture halls, discussing with colleagues: Our research is being misused, we have to talk about it!

The more publications he read in the weeks that followed, the more uncomfortable he became. He could no longer just be amazed at how smart the computers already were and what psychological subtleties they could read out of people's texts. He no longer read an essay on the question of how to influence an Internet forum out of pure curiosity for researchers. "Colleagues wanted to build a machine that manipulated people's opinions on the Internet," he says, "and they weren't even aware that it was dangerous."

In the meantime, Strube has found colleagues, such as the computer scientist Dirk Hovy from the University of Copenhagen. "So far our research has been mostly academic, it was not assumed that individuals could be affected," says Hovy today, "so we did not see the need to question it ethically. Now the algorithms have come to a point where they can have an impact. "