What is the AI alignment problem and how can it be solved?

Shutterstock/Emre Akkoyun

WHAT bash insubstantial clips person to bash with the extremity of the world? More than you mightiness think, if you inquire researchers trying to marque definite that artificial intelligence acts successful our interests.

This goes backmost to 2003, erstwhile Nick Bostrom, a philosopher astatine the University of Oxford, posed a thought experiment. Imagine a superintelligent AI has been acceptable the extremity of producing arsenic galore insubstantial clips arsenic possible. Bostrom suggested it could rapidly determine that sidesplitting each humans was pivotal to its mission, some due to the fact that they mightiness power it disconnected and due to the fact that they are afloat of atoms that could beryllium converted into much insubstantial clips.

The script is absurd, of course, but illustrates a troubling problem: AIs don’t “think” similar america and, if we aren’t highly cautious astir spelling retired what we privation them to do, they tin behave successful unexpected and harmful ways. “The strategy volition optimise what you really specified, but not what you intended,” says Brian Christian, writer of The Alignment Problem and a visiting student astatine the University of California, Berkeley.

That occupation boils down to the question of however to guarantee AIs marque decisions successful enactment with quality goals and values – whether you are disquieted astir semipermanent existential risks, similar the extinction of humanity, oregon immediate harms similar AI-driven misinformation and bias.

In immoderate case, the challenges of AI alignment are significant, says Christian, owed to the inherent difficulties progressive successful translating fuzzy quality desires into the cold, numerical logic of computers. He thinks the astir promising solution is to get humans to supply feedback connected AI decisions and usage this to retrain …