Reuters
- A Max-Planck Institute study suggests humans couldn’t prevent an AI from making its own choices.
- The researchers used Alan Turing’s “halting problem” to test their theory.
- Programming a superintelligent AI with “containment algorithms” or rules would be futile.
- Visit Business Insider’s homepage for more stories.
It may not be theoretically possible to predict the actions of artificial intelligence, according to research produced by a group from the Max-Planck Institute for Humans and Machines.
“A super-intelligent machine that controls the world sounds like science fiction,” said Manuel Cebrian, co-author of the study and leader of the research group. “But there are already machines that perform certain important tasks independently without programmers fully understanding how they learned it [sic].”
Our society is moving increasingly towards a reliance on artificial intelligence – from AI-run interactive job interviews to creating music and even memes, AI is already very much part of everyday life.
According to the research group’s study, published in the Journal of Artificial Intelligence Research, to predict an AI’s actions, a simulation of that exact superintelligence would need to be made.
The question of whether a superintelligence could be contained is hardly a new one.
Manuel Alfonseca, co-author of the study and leader of the research group at the Max-Planck Institute’s Center for Humans and Machines said that it all centers around “containment algorithms” not dissimilar to Asimov’s First Law of Robotics, according to IEEE.
In 1942, prolific science fiction writer Isaac Asimov laid out The Three Laws of Robotics in his short story “Runaround” as part of the “I, Robot” series.
According to Asimov, a robot could not harm a human or allow them to come to harm, it had to obey orders unless such orders conflicted with the first law, and they had to protect themselves, provided this didn’t conflict with the first or the second law.
The scientists explored two different ways to control artificial intelligence, the first being to limit an AI’s access to the internet.
The team also explored Alan Turing’s “halting problem,” concluding that a “containment algorithm” to simulate the behavior of AI – where the algorithm would “halt” the AI if it went to harm humans – would simply be unfeasible.
Alan Turing’s halting problem
Alan Turing’s halting problem explores whether a program can be stopped with containment algorithms or will continue running indefinitely.
A machine is asked various questions to see whether it reaches conclusions, or becomes trapped in a vicious cycle.
This test can also be applied to less complex machines – but with artificial intelligence, this is complicated by their ability to retain all computer programs in their memory.
“A superintelligence poses a fundamentally different problem than those typically studied under the banner of ‘robot ethics’,” said the researchers.
If artificial intelligence were educated using robotic laws, it might be able to reach independent conclusions, but that doesn’t mean it can be controlled.
“The ability of modern computers to adapt using sophisticated machine learning algorithms makes it even more difficult to make assumptions about the eventual behavior of a superintelligent AI,” said Iyad Rahwan, another researcher on the team.
Rahwan warned that artificial intelligence shouldn’t be created if it isn’t necessary, as it’s difficult to map the course of its potential evolution and we won’t be able to limit its capacities further down the line.
We may not even know when superintelligent machines have arrived, as trying to establish whether a device is superintelligent compared with humans is not dissimilar to the problems presented by containment.
At the rate of current AI development, this advice may simply be wishful thinking, as companies from Baker McKenzie to tech giants like Google, Amazon, and Apple are still in the process of integrating AI into their businesses – so it may be a matter of time before we have a superintelligence on our hands.
Unfortunately, it appears robotic laws would be powerless to prevent a potential “machine uprising” and that AI development is a field that should be explored with caution.
Powered by WPeMatico
