Technology slowly learns about people

Artificial intelligence: The computer disagrees

Let's jump back mentally a good hundred years to get a different view of the rapid progress of the present. Back then, modern physics was scientific hot shit. At the turn of the century Max Planck had formulated the quantum theory, and shortly afterwards Albert Einstein had formulated the theory of relativity. A little later, Niels Bohr showed what that meant for the atoms. During this intoxicating time (at least for physicists), the understanding of nature expanded enormously - everything was the history of science.

Let us now add a thought experiment to the time jump. Let us assume that during this time the interested public adhered to the obsession that the seven-mile steps of the physicists were aiming at a fantastic goal, namely that, tata !, to enable time travel. The novel, for example, could be such an idea The time machine by H. G. Wells, published in 1895.

The punch line behind this journey of thought: If you had reflexively wanted to see a step towards the time machine in every breakthrough in physics, you would have been very disappointed at some point - all the progress, but still no breakthrough ...

And with that back to the present and to yours hot shit,artificial intelligence (AI). It was on the front page of the leading science medium last week Nature done it again. For the entire decade, computer scientists have reported success after success: the supercomputer Watson defeats the human champion in the guess show Jeopardy! (2011), in the board game GoBeats the self-learning system AlphaGo Pro Player (2015/16), while playing poker the AI ​​Libratus bluffs human gamblers (2017), an AI from the Chinese Alibaba group does better than humans in the reading comprehension test (2018), and the software GPT wrote last year -3 Factual texts and poems that sound as if they were written by people (ZEIT No. 54/20). In the past decade, AI systems have composed new songs in the sound of the Beatles, calculated works of art in the style of great painters and Beethoven Unfinished, well, we say completed.



Be there live online when our podcasts are created and meet your favorite hosts at the first ZEIT ONLINE podcast festival on Sunday, June 20, 2021.

With your registration you take note of the data protection regulations.

Many Thanks! We have sent you an email.

Check your mailbox and confirm the newsletter subscription.

All of these advances generated attention beyond the professional audience. Bit by bit, computer systems acquired capabilities that were once considered the sole domain of humans. Some, like the British-Swedish philosopher Nick Bostrom, have long warned against a "superintelligence" that would usurp all control (ZEIT No. 21/16). For laypeople, the question remains: will the machines take over the shop in the end?

The majority of experts say no, there is at least one instructive distinction between weak and strong AI. While a weak AI system was designed for a specific task and trained with sample data, a strong AI would be an all-rounder. She could do complex tasks from different fields at least as well as humans. Weak AIs are indeed making breathtaking advances, and the above list could be extended almost indefinitely. There is just as little strong AI today as there was a time machine a hundred years ago. Nevertheless, the idea of ​​a universal machine brain triggers diffuse fears and provides shrill warnings.

That’s where the one comes in NatureArticle by computer scientist Noam Slonin and his team from the IBM research centers in Haifa and Dublin on "an autonomous debating system". Within ten years, 53 researchers have developed a computer that can argue with people. He searches for information on a topic from various data sources, identifies arguments, presents them in natural language and takes up counter-arguments from people. IBM's "Project Debater" has proven all of this in front of an audience.

Z +: Exclusively for subscribers

Sauerkraut-Schnitzel: Schnitzel in sauerkraut heaven

Harald Martenstein: I liked a perpetrator

Bob Dylan: In the voice of a grave robber

More subscription texts

But the article does not report a breakthrough, at most an interim result. In 2018 and 2019, the program competed against three skilled human discussants, and the audience had to vote on who was more convincing. Once the computer won, once the people, once the audience voted in a draw. Astonishing progress, yes, but not a triumph like at Jeopardy or Go. Debate is just not a game based on simple rules that fall into the "comfort zone of AI", the authors write. "Debating with people, on the other hand, lies in other areas where people still have the upper hand." To change that, you need completely new approaches in order to achieve "real progress". A remarkably humble statement from top exponents of this hype discipline.

Just as the idea that the time machine was waiting around the next corner a hundred years ago would have obscured the view of the rapid advances in physics, so much the fixation on (over) strong AI clouds the view of the practical benefits of this computer science discipline.

Project Debater also combines techniques that are each useful in themselves, without any super intelligence. One of the project's spin-offs is called Speech by Crowd. Noam Slonin explains the tool to ZEIT with a fictitious example: "If Michael Müller, the mayor of Berlin, wanted to create more green spaces, he could ask his 3.6 million citizens to take part in a survey, in everyday language. And speech by Crowd could summarize these millions of answers for him in a few paragraphs - including the most important points and their weighting. "

Another part of Debater is already helping chatbots understand flowery phrases. And Debater's automatic argument search, says Slonin, can be used against disinformation campaigns.

Long before the first computers perhaps debated people in the ground, they were already being used as tools in debates. There is no need to travel into the future for this.