AI: Real Empathy?
Most of the attempts to ‘socialize’ robots, to smooth human-robot interactions, only try to emulate what humans would do. The robot smiles, raises its eyebrows, coos. What we will need is quite different: a robot that actually relates to us, such that it has our best interests in high regard.
Such a robot requires a ‘Theory of Mind’ — it must understand that it is a thinking being, embodied, and that we are also such beings, with our own feelings and motivations. It must also have a context for our values in its own; the robot must have personal aesthetics and ethics, and from those it can recognize and appreciate the values of others. These are nebulous and challenging hurdles. I argue for an additional, subtle requirement which will act as a check upon many errors: the robot must prefer the real over the imagined or contrived.
Suppose an artificial general intelligence received a reward signal whenever it made people laugh or smile. By Goodhart’s Law, the machine, while attempting to optimize its reward, might keep humanity in a perpetually drugged state! It would seem to have achieved its goal of joviality, yet this way of reaching that end is something we consider ‘outside the spirit of the goal’. Suppose we wanted world peace, bounty, togetherness, and the AI simply attached us to virtual reality systems where we were presented with peace in a fabricated world. We would be force-fed happiness, unable to discover real joy. If a robot empathizes, it must also prefer that our emotions happen ‘naturally’.
Neuroscientists have uncovered only some components of empathetic behavior in the human brain, yet a preference for the real may be even more elusive. The most ‘natural’ emotions would occur if the artificial intelligence did nothing; what happens happens, unforced. Yet, we want to create an intelligence which intervenes upon events in a way that aids our best interests, without crossing the line into ‘forced’ feelings. If AI can strike that balance, it will be our good steward. Otherwise, I wouldn’t trust it with any goals outside human oversight.
