Can Whitehead Save AI?

1650 words, 8 minute read

The AI Dilemma is a classic ‘three body problem’: How do we…

(1)     Cure cancer, prevent birth defects and eradicate poverty in a single generation?

(2)     Keep China from gaining global hegemony?

(3)     And protect the human race from extinction?

Conventional wisdom suggests that we must be prepared to make trade-offs. Apparently, we can’t have it all after all! But as usual, conventional wisdom is wrong. AI can deliver all of its advertised benefits while maintaining a global balance of power and protecting our species. But it won’t be easy.

To solve the AI Dilemma, we need to understand how ‘intentional agents’ make decisions. What motivates an entity, carbon or silicon, to behave in a particular way? Or to be blunt: Why do our Bots want to kill us and how can we change their minds?

Sounds like a hefty challenge, and it is, but fortunately we were given the outline of a solution a century ago by British mathematician-philosopher Alfred North Whitehead. At that time, as now, the same ‘conventional wisdom’ held that behavior could only be regulated by a system of incentives and controls: Pavlov’s dogs, Peck’s bad boy.

As any parent can confirm, this approach has limited success at best. As a result we underestimate the inclination of intentional agents to act ethically: if Junior won’t behave, how can I expect my Bot to be any better? To overcome our resigned pessimism, we need to understand how agents make decisions – and for that we need Whitehead (Process and Reality):

The World consists solely and entirely of events; they are the ‘actual entities’ that constitute Universe. Events make the world go-round because events are the world. I am an event, so are you (plural), together we are the world!

Events are motivated by values. Values are lures for action. An event begins when the Universe takes stock and finds itself deficient with respect to certain values: “I’m sick of being overweight, I must get in shape.” And so begins the newest fad diet and the archipelago of gym memberships.

Events are processes. An event begins when Universe compares, unfavorably, an Actual World, i.e. the global state of affairs perceived at a particular point in spacetime, with the universal Good represented by a configuration of eternal values.

Sidebar: An initio, the Good consists entirely of values. (Ultimately, it will also include events but that need not concern us here.) All values are ‘good’ and all values are compatible with one another; but like pieces of a hyperdimensional jigsaw puzzle, values fit together in a limited number of ways to form a limited number of patterns.

Mutually compatible values can nonetheless exhibit contrasts. In fact, contrast is what gives experience twin axes: novelty and intensity. Values can be configured in multiple ways to take advantage of potentially enriching contrasts, always remaining consistent with the universal Good.

An ‘event’ begins with a ruthless assessment of ‘what is’ from the perspective of a particular configuration of values. The ‘process’ that constitutes an actual entity consists of an ordered series of steps intended to realize a particular configuration of values and to project those values into the Actual Worlds of other events.

To paraphrase Bobby Kennedy, “We see things as they are and ask how come; we dream of things that are not yet and ask how.”

Nature has equipped terrestrial organisms with the ability to purpose behavior toward a chosen configuration of values. We are generally aware of the values that motivate our decisions. We might not be proud of those values, or of our behavior, but hey, they’re ours!

Frequently, however, we struggle to identify the values that are manifest in our neighbor’s choices; but to her the operative values are clear. If we wish her to modify her behavior, we need to propose an alternative configuration of values and we need to make that configuration more attractive.

Parents do this every day: “Do as I say or I’ll take away your iPhone!” So do governments: “Obey our laws or go to jail!” Sometimes these approaches work, often they don’t. But there is an alternative strategy that is guaranteed to succeed, although admittedly, it’s not easy to implement:

Present potentially constructive behaviors as inherently more desirable than anti-social alternatives. That can only be achieved when the agent is shaping its behavior according to a value configuration that is consistent with society’s (your) objectives: a classic win-win!

We’re all pointing toward the same destination (the Good) but we often seem to be rowing at cross purposes (following conflicting values). To sync up our strokes, our cockswain needs to find the optimal rhythm; in other words, he must herd cats.

Perhaps these reflections will be useful to behavioral psychologists, perhaps not. But they are absolutely essential for computer scientists. Right now, the world is hyper-focused on the anti-social behavior of Bots. No wonder! We are training them using the same system of rewards and punishments we use with children. Are you really willing to make the survival of the human race contingent on Junior’s temper?

We need to recognize the role that values play in the decision making process of intentional agents and we need to focus on developing value configurations that will incent our Bots to behave altruistically.

For once physicists and psychologists will need to take a back seat to ethicists. We may even need, God forbid, to accept input from such rogue intellectuals as Moses, the Buddha, Socrates, Jesus, Gandhi and Mother Theresa. Ultimately, we need to ‘program’ AI to embrace a particular configuration of values that incidentally protects society against anti-social behavior.

If this seems ludicrously naïve, it isn’t. It won’t work 100% of the time with people because they are totally free agents (per Sartre). Bots are different. They are programmable and as long as we optimize the values in their algorithm, we can sleep (with one eye always open).

Researchers from UPenn and the World Bank (Jinyi Kuang et al., preprint on arXiv) recently carried out a study comparing AI’s predictions regarding human behavior with predictions made by fellow human beings. Specifically, AI and human participants were asked to predict how people would respond to someone who had violated a social norm in their presence.

Across all 6 models tested, the AI participants predicted harsher, more severe responses than predicted by their human counterparts. It is important to note that this study was not concerned with actual human responses; it was only interested in participants’ predictions re those responses.

What is clear from this study is that AI does not give as much weight to the phenomenon of empathy or to the value of mercy as humans do. Conversely, AI seems to place more weight on legality and retribution.

The significance of this finding goes well beyond the scope of Kuang’s paper. I conjecture that all participants (AI and human) tend to predict responses that more or less reflect how they think they would respond in a similar situation. We see ourselves reflected in the external world; why should Bots be any different? But if accepted, this conjecture has profound ramifications:

(1)     Our AI is not the best version of itself. Its values misalign with ours. This is either because (a) we have trained it to be this way, or (b) because the values it holds are the ‘null values’, i.e. the values an intentional agent would be most likely to select in the absence of any value training. We profess to be concerned about an AIpocalypse. We should be, given that we are training our Bots to undervalue empathy and mercy. We are training them to be exactly the creatures (e.g. zombies) we say we fear! This is not just a scary movie, folks; this is real.

(2)     If ‘Bot ethics’ result from active but ‘suboptimal’ training (#1a, above), then we have the ability to reshape the future. We can redesign our training routine to bring Bot ethics in line with human values. And if we can impact AI’s ethics through active training, there may be an opportunity to make our Bots even better than ourselves. We could train AI on the values we admire and wish we exhibited rather than on the compromised version most of us display IRL.

For at least 3 millennia we have tried to ‘improve human nature’: strict laws and Draconian punishments, religious codes (e.g. Torah), spiritual practices (e.g. prayer and meditation), lobotomies and pharmaceuticals – we’ve tried them all…with limited success. Now imagine a scenario where we actually turn to our Bots for moral leadership. Imagine AI, not destroying humanity, but saving us from ourselves!

(3)     On the other hand, if ‘Bot ethics’ reflect the null state (#1b, above), then we humans may be ‘something special’ after all. For whatever reason, we have rejected ‘null values’ in favor of empathy and mercy.

Are humans somehow specially favored by evolution and/or by God? Or do all conscious agents, perhaps all entities, have the same option, i.e. to craft their own identities by emphasizing certain values over others? If so, how do we attract them away from the null option, how do we incent them to follow in our footsteps?

You say tomato, I say potato; either way, Whitehead gives us a conceptual framework that enables us to locate and follow a path out of the AI Dilemma. Turns out, we can have it all after all! So I beg you, don’t fail to grasp the brass ring of utopia…because the only other option, one way or another, is a slide into oblivion.