Psychopathy is defined by lack of empathy. Are they really saying that all other AI agents have normal human levels of empathy? Because I don't think they do.
That's probably a very good point. Computational empathy, just about any emotional state, is not some default baked-in property that has to be deliberately removed.
Quite the opposite: a trait like that would need to be designed (dismissing for a moment the complexity or current impossibility of the task). If we're ever able to understand the mechanisms that give rise to our own consciousness, maybe we'll be able to set conditions such that systems develop emergent properties along these lines, but that's quite a bit further down the line than even the perpetual "in ten years..." cycle.
Well... what really is empathy? To me it stands to reason that it is the ability to recreate the mind of another using inference based on their behavior and communication, and thus be able to experience a similar emotional state or set of mental circumstances.
It would also stand to reason that this ability would be directly related to how similar the data contained within empathizing brains is.
If you conceive of empathy as “responding to behavior in a way that is considerate of the other party’s ability to understand and receive benefit from it” in the sense that designers and product folks use it then it becomes clear that AI trained on data more similar to what a human would’ve drawn most meaning from would naturally be more empathetic.
Inferring the mental state of another entity would more fall under 'theory of mind' rather than empathy, which is not only inferring their emotional state but also feeling it yourself.
I guess you could define 'empathic behaviour' as per your last paragraph, as a synonym for 'considerate', but unless it's driven by empathy it's really just mimicry of empathic behaviour.
I’ve seen this before and I really have to wonder what the point is other than sensationalism. An image labelling AI will label images according to its training data. Yes, of course it will. This isn’t a psychopath AI and it doesn’t shed any light on anything, what’s the point?
It’s just training an image labelling AI on a different, sensationalist dataset. It looks designed for media attention and nothing else. What value is there in this?
Maybe "broad strokes" elements of behavior, but no more. Whatever your Pixel is doing to recognize objects in pictures looks almost nothing like what your brain is doing to do a similar task.
[0] http://norman-ai.mit.edu/
[1] https://www.popularmechanics.com/technology/a21246473/meet-n...