Computer Science : Giving AI Some Common Sense

  • -Aktualisiert am

Man meets machine Bild: dpa

Lurking behind all the excitement about AI, there is a problem. Despite being trained on immense amounts of data and often showing uncanny abilities, current AI systems make bizarre, lamebrained mistakes.

          9 Min.

          The Age of Artificial Intelligence is upon us. Articles on the topic are everywhere, even the cover of Cosmopolitan. Siri and Alexa offer us AI in our pockets and our kitchens. AI in the cloud handles online chat inquiries, makes product recommendations, and translates between languages. And AI technology is at the heart of one of the more exciting promises of our era: self-driving cars.

          After its birth in the 1950’s, AI’s first half-century had its ups and downs, with occasional dramatic achievements, like beating the world chess champion, countered by many failures to live up to its hype. But now, largely in the form of neural net-based machine learning (ML), it seems to have finally arrived. Industry and governments are investing billions of dollars and fighting fierce battles for AI-savvy talent. Recent progress even has some believing that machine intelligence is close to rivaling that of humans: Elon Musk tweeted that he’d be surprised if we didn’t have “artificial general intelligence” by 2029.

          But lurking behind all this excitement, there is a problem. Despite being trained on immense amounts of data and often showing uncanny abilities, current AI systems make bizarre, lamebrained mistakes.

          Striking examples appear regularly. Image recognition software has mistaken school buses for ostriches and turtles for rifles. A Tesla in “smart summon” mode drove itself into a parked jet on a relatively empty tarmac. Language systems sometimes spew potentially dangerous babble; one even responded “I think you should” when asked, “Should I kill myself?” And stunningly, Alexa instructed a 10-year-old child to hold a metal coin on phone charger prongs plugged halfway into an electrical socket. If a person did these things, we’d question their intelligence, if not their sanity.

          The errors are unpredictable and unexplainable

          Why is this happening? Why do these otherwise successful AI systems sometimes flub so badly? It's not that they are still missing some sort of ultra-advanced ability or expertise. What they seem to lack is something much more basic, something that humans constantly rely on: ordinary common sense.

          But why are occasional failures to use common sense by AI systems such a problem? It’s simple: the errors are unpredictable and unexplainable. Failure to see and act on things that are obvious to the rest of us means that we cannot trust the systems to act reliably on their own. If at any point, without warning, an AI system can misfire in a stupefying, unhuman way, how can we know it won’t be during the very next mission-critical or life-dependent action? Without common sense, an AI system does not have the capacity to recognize that it is about to do something absurd.

          Maybe if it could explain why it did something, we could at least tell an AI system how to avoid the mistake the next time around. But we can’t do that. “Don’t do that again!” would have no effect on a system that has no idea what “that” is or how to avoid doing it. The reasons AI systems choose to do what they do are opaque and untraceable, and ML-based AI systems, trained on reams of data but not general principles or rules of thumb, are incapable of taking advice. I wish I could say to my car in self-driving mode, “Wait! That’s a pedestrian!” and have it respond appropriately. But as good a driver as it is, it simply could not understand what I was talking about. My car has no concept of what a pedestrian is or why stopping would matter.

          Weitere Themen


          Protestbewegung in Iran : Freiheitsbewegung in einem erschöpften Land

          Der Mut der Frauen inspiriert die Männer: Die Proteste in Iran zeugen von der tiefen Kluft zwischen Gesellschaft und Regierung. Sie markieren einen Wendepunkt auf dem Weg zur Freiheit. Ein Gastbeitrag.


          Immer auf dem Laufenden Sie haben Post! Die wichtigsten Nachrichten direkt in Ihre Mailbox. Sie können bis zu 5 Newsletter gleichzeitig auswählen Es ist ein Fehler aufgetreten. Bitte versuchen Sie es erneut.
          Vielen Dank für Ihr Interesse an den F.A.Z.-Newslettern. Sie erhalten in wenigen Minuten eine E-Mail, um Ihre Newsletterbestellung zu bestätigen.