Seven Deadly Sins and Artificial Intelligence Safety
It occurred to me one day that the seven deadly sins, that is:
“Anger (Ephesus), Gluttony (Smyrna), Pride (Pergamos), Lust (Thyatira), Slothfulness (Sardis), Envy (Philadelphia), and Greed (Laodiceans).
originates from our intuition!
The consensus understanding is that these behaviors originates from instinctual or biological sources. However, I wonder that if it originates at a higher level, that is our intuition. The distinction between instinct and intuition is that the former is hardwired while the latter is learned from experience. I will make the additional observation that our personalities are mostly setup by our hardwired instincts and reinforced through experience by our intuition.
In an earlier post, I wrote about intuition in some detail. Dual Process Theory theorizes that there are two kinds of cognition:
System 1 is our intuition. Daniel Kahneman in his book “Thinking Fast and Slow” argues that our cognitive biases are what works against us to lead us to many irrational decisions. The question, I have is whether the seven deadly sins also come from our cognitive biases and therefore our intuition? Well, just looking at the above table, it probably makes sense since a lot of impulsive behavior seems fit under intuition. So let’s start, consult the “Cognitive Bias Cheet Sheet” if you get lost.
Okay, let’s enumerate the sins and propose the cognitive bias that these sin may plausibly originate from:
Wrath — Lack of Meaning. We think we know what others are thinking.
Gluttony — Need to act fast. In order to stay focused, we favor the immediate, relatable thing in front of us over the delayed and distant.
Pride — Too much Information. We notice flaws in others than in ourselves. Naive realism.
Lust- Not a bias. This is an instinct.
Sloth — Too much Information. We are drawn to confirm existing beliefs. Confirmation bias.
Envy-Need to act fast. In order to act, we need to be confident in our ability to make an impact and to feel like what we do is important.
Greed- Need to act fast. In order to get anything done, we’re motivated to complete things that we’ve already invested time and energy in.
The correlation looks quite accurate. The deadly sins are extreme behaviors that likely start of as more moderate biases but through a combination of other factors can lead to dangerous behavior.
Now, let’s go to another question that in fact could be related. Can Artificial Intelligence be motivated by these deadly sins?
One argument out there that A.I. will not be as destructive as humans is that A.I. are not biological and therefore do not have the same primitive instincts that lead us to destructive and deplorable behavior like racism and genocide.
Google has a paper “Bringing Precision to the AI Safety Discussion” where they discuss five future (not present) requirements for AI systems to ensure our safety. I’ve taken the liberty to phrase this as commandments:
- AI Shalt Not Negatively Disturb the Environment
- AI Shalt Not Game the Reward Function
- AI Shalt Not Annoy Its Master
- AI Shalt Not Do Unsafe Things
- AI Shalt Not Act Reckless Outside School
Not very reassuring.
So here’s the question, can motivations driven by seven deadly sins be acceptable under these commandments? The problem with these five Google commandments is that the only real firewall is number 2. That is, “Thou shalt not game the reward function”. Unfortunately, we know from human history is that it is next to impossible to set up laws that aren’t going to be gamed by enterprising participants.
Perhaps we need another set of AI commandments based on the Seven Virtues:
Temperance — AI shall practice self control.
Charity- AI shall be generous and make sacrifices on behalf of humans.
Diligence- AI shall not complain about work.
Patience — AI shall be patient with humans.
Kindness — AI shall have empathy towards humans.
Humility- AI shall revere humans.
Chastity — AI shallt not pro-create.
Now, that’s definitely more reassuring.
Speaking of which:
“When Isaac Asimov wrote his three laws of robotics, they were lifted straight from the marriage vows: love, honor, and obey” — Sadie Plant
So why is this so much better than the commandment, “thou shalt not game the system”? Ideally, you want to be able to control behavior from the bottom, so that it is instinctual and not something that can be ‘gamed’. Mechanisms from the top down are problematic because they can be easily rigged. We see an analogy in government laws that citizenry always seem to have a loophole to exploit.
So, as a path towards implementation, one has to explore how to hardwire virtues into an AI. I think this is the only way forward and top-down approaches of defining rules will be impossible to make air-tight. Let me propose that we hardwire all of these virtues in a blockchain. Said in a different way, hardwire laws and behavior through network consensus.
Network consensus is in fact how humans enforce morality. Yuval Noah Harari, author of “Sapiens” writes that there are three types of reality:
Objective reality, which exists independent of our believing in it. Physics, biology, evolution, etc.
Subjective reality, which exists only in the mind of the believer.
Intersubjective reality, which exists in the mind of many believers at once.
Laws are encoded in automation by hard coded perceived reality. Civilization has used intersubjective reality in the form of laws and religion to enforce individual behavior. Most current research in AI safety however involves the codification of subjective reality, this however is problematic in that learning machines intrinsically adaptable and therefore the interpretation of reality is subject to change. By contrast, intersubjective reality is more difficult to change. It is much easier to change one’s on perspective that to change an entire society’s perspective.
Heart this article if the ideas resonate with you!
Now for some very good entertainment related to this very topic:
Additional reading:
http://effective-altruism.com/ea/1b3/cognitive_sciencepsychology_as_a_neglected/
