Ever since the victory of the IBM Watson computer system over its human teammates in the American quiz show “Jeopardy” in 2011, it was clear what computers are capable of thanks to artificial intelligence (AI).

The electronic brains can not only recognize complex patterns, faces and language, they are - as Watson showed - after appropriate training they are also able to correctly interpret the meaning of spoken words.

AI has also found its way into materials research, biology and medicine.

But how trustworthy are the smart computers when it comes to the results, and where are the opportunities - but also the technical limits?

Those were some of the questions from the panel discussion “AI - Promise or Threat” last Thursday afternoon in Lindau.

Manfred Lindinger

Editor in the “Nature and Science” section.

  • Follow I follow

    The American internet pioneer Vinton Cerf urged caution right from the start.

    Often you don't know how an AI calculates with machine learning and how reliably it does it.

    Sometimes there are surprises with pattern recognition.

    It does happen, for example, that she confuses a dog and a polar bear - for example, when the algorithm only classifies the animals according to the color of their fur.

    Even nonsensical correlations without any logic are possible, according to Cerf: "Car tires can give birth to babies", an AI could claim because an expectant mother is driven to the hospital.

    Meaningful AI for materials research

    Applications in medicine should therefore also be assessed with caution. Incorrect results as a result of incorrect correlations could have fatal consequences for patients. Certain knowledge of statistics and computer science are essential for the medical profession, said Cerf. The user has to know what an algorithm should calculate. As the last resort, humans are indispensable, affirmed Bernhard Schölkopf from the Max Planck Institute for Intelligent Systems in Dresden. Often the right solution would be more intuitive. In the case of very complex questions, for example in mathematics and physics, that could be solved with a classic computer program, AI is extremely useful. But then it would be clear what the algorithm should calculate, said Schölkopf.

    The chemist Marco Eckhoff from the University of Göttingen reported how machine learning can help in the development of new materials. The chemist uses AI to calculate complex quantum chemical synthesis reactions in which many particles are involved. This often involves the search for configurations in which the entire system is in the lowest state or in chemical equilibrium. If many atoms or molecules are involved, a classic computer program quickly reaches its limits. In the end, however, the calculated models always have to be checked with an experiment. Eckhoff uses AI to manufacture high-performance catalysts.

    At the end of the discussion, a young scientist wanted to know who is to blame when a self-driving car controlled by an AI causes an accident - the person, the AI ​​or the software company. An interesting question, answers Cerf. Many courts will also have to grapple with it.