Sophia: Uncanny, and untrustworthy?
Hanson Robotics demonstrated a new human-like robot called “Sophia” at SXSW recently. CNBC covered Sophia in this video interview with David Hanson. At the time, I was not on Medium, but I wrote a longish Facebook post in response to the video. It seems timely, so I’m reposting a slightly updated version here, now.
I’ve spent the last year studying things like machine ethics, machine learning and AI, and machine/human interactions, and there are two trends demonstrated here that I think are big mistakes.
First, it falls right over the cliff to the “Uncanny Valley”. This robot is too close to human, but is still not human. That sets off all the instincts and subconscious fears that cause us to fear and hate things that are inhuman. Vampires, changelings, and zombies have always been hated and feared because they are not us.
Now, communicating with people is done not merely through our words, but our tone of voice, posture, gestures, facial expressions and so on, and it is important that as robots and other autonomous systems become more and more like persons, they use the full spectrum of communications. That I buy. Its abilities in that area are a definite step up, and I would encourage that. But, to counter the uncanny problemthat, I would do something like make it more cartoon like or just make her skin blue rather than fleshtone. I would also desexualize “her”. It isn’t an animal. It has no sex. The implications of making this thing that is clearly a thing, clearly sub-human, a woman is wrong on many levels. And that brings us to:
Second, and more importantly, if robots, autonomous and personified systems are going to be integrated into our as pseudo- or near-persons, it is absolutely crucial that they be truthful and Sophia is not. It pretends not only to be female, but to have feelings, and aspirations. That’s just not true. It, or rather its creators through it, are lying to us, and that makes it not only an inhuman in the uncanny valley, but a liar and a fraud. Deceptive inhumans are monsters, things we dread, things we hate.
This problem is tied in many ways to the uncanny valley. One of the most important is that our best attempts at AI and Artificial General Intelligence (AGI) actually are doing things that are increasingly similar to biological systems in general and humans specifically. For instance, as the 9 dan professional go player and AlphaGo developer were discussing the way that the system plays Go, the clearest thing that they could say about the “policy network” part of the system was that AlphaGo had articicial opinions as to what an expert player would most likely do in a specific situation.
If we are going to understand and trust the systems that we interact with, they need to be honest with us. They must not claim to have emotions, aspirations, feelings and such, that they do not have, both because it makes them liars but because it makes them like the Boy Who Cried Wolf. Some day they may have these things or such close analogues that there is no clear way to call them anything else. We will need to trust them then, and that requires them to be candid now.
It is OK for AlphaGo to say, “I believe that…” or that “It seems to me that a human would…”, even “I unsderstand that…”, in that those are reasonable ways to describe the judgements that its policy and value nets make. However, when Sophia says, “I feel like I can be a good partner to humans…” unless and until it is reporting that it has evaluated millions of interactions between humans and robots like itself and humans, and is striving to put its conclusions into plausible human words, it is telling a falsehood.
This is why I chose “Candor” as one of the handful of virtues that are critical in making personified systems worthy of our trust.
