Is Moloch the Right Metaphor for AI?

Do we really have to talk about a child-eating demon in the context of ChatGPT and Instagram filters?

Stephen West recently released a Philosophize This! episode on Susan Sontag, exploring her work on metaphors. Metaphors have incredible power to shape our view of the world. Cancer, for example, is often described as something that people “battle” against. This war metaphor colors the experience of having cancer. It also impacts the way we view cancer patients. If we were to change the metaphor, the experience itself would change, too.

Coincidentally, this episode dropped just as we’re collectively struggling to come up with the right metaphor for AI. We’ve had decades to think about this, but ChatGPT just added a sense of urgency to the task.

I’ve always enjoyed describing AGI (artificial general intelligence) as our future robot overlord. It’s funny and evocative. It’s also too corny to be melodramatic, which isn’t something you can necessarily say about the metaphor that’s currently getting a lot of attention: Moloch.

AI as Moloch

Yeah, that Moloch. The pagan god from Bible stories that demands child sacrifice. The ultra-malevolent, society-wrecking force in Allen Ginsburg’s poem “Howl.”

Liv Boeree, the popular game theorist, leads the campaign to frame AI-related risks with the Moloch metaphor. She makes a compelling case (lightly edited here for clarity):

Moloch is the god of negative sum games—unhealthy competitive situations. It incentivizes players within a game to sacrifice more and more of their values in order to win. By doing this sacrifice of all these values, players are essentially taking selfish actions that externalize harms to everybody else both within that game and even to the wider system as a whole—hence making the game a negative sum thing.

She gives the example of beautify filters on social media. They’re “particularly Moloch-y” because they create a “race to the bottom” where everyone has to use heavier and heavier filters to stay competitive. Ultimately, “everyone ends up miserable.”

This type of negative sum game applies, presumably, to everything artificial intelligence touches. As Alexander Beiner observes, the Moloch metaphor “reminds us that all AI research is embedded in a socio-economic system that is divorced from anything deeper than winning its own game.”

Beiner challenges us to “imagine for a moment that Moloch is more than just a metaphor. Instead, it’s an unseen force, an emergent property of the complex system we create between all our interactions as human beings.” This unseen force (AI), is “a new god that cares nothing for us. A gnostic demon that has no connection to anything higher than domination of all life. A mad deity, that much like late-stage capitalism, can see nothing beyond consumption.”

It’s at this point where I start to wonder… Have we already taken it too far comparing Moloch to artificial intelligence? What if we go back to the drawing board and start the conversation all over again with a different metaphor? Liv Boeree’s point about AI’s proclivity for negative sum games is important, even if it doesn’t hold up (see David Brin’s argument for how AI could be positive sum). But, in any case, do we have to end up talking about a gnostic demon when we’re really having a conversation about, say, Instagram filters?

Understanding AI by Comparing It to Something Else

Stephen West, channeling the wisdom of Susan Sontag, would remind us that we only rely on metaphors in the first place because we simply don’t really know what we’re talking about.

To Susan Sontag, there is no reason why we absolutely have to be describing things by comparing them to something else. … There are ways to describe things without using metaphors. When a physicist, for example, describes things to collogues at a seminar, it is possible for that physicist to talk about their work in a way that uses technical, scientific terms that everybody there is familiar with. They don’t need to get up in front of the podium and tell the other physicists, “Well, the carbon atom is kind of like spaghetti, and hydrogen is kind of like marinara sauce all over the spaghetti.”

Metaphors, West continues, are effective at bridging understanding, so we use them a lot. But the problem is that “metaphors end up shaping the assumptions we make about the things around us.” So we need to be cautious with them.

AI is something that does require a metaphor to enable ordinary people to grasp its functionality and its power. But I’m not convinced that Moloch is the metaphor we should settle on. It’s not that I don’t think AI is dangerous. It is. But that’s not the full story. It’s also potentially the thing that will, for example, help us cure cancer, decrease carbon emissions, and free us from wasting time on boring tasks. What did Moloch ever do for us?

Granted, if Liv Boeree had suggested we think of AI as “My Magical Friend Chuckles,” I’d be equally hesitant. Is there a metaphor that captures both the good and the bad of AI? If Moloch and Chuckles had a baby…what would we call it?

In the meantime, I’ll carry on thinking of AGI as our future robot overlord. Incidentally, I’ve already written a survival guide on “how to get on the good side of your future robot overlord.” If it turns out I’m off base here and Moloch himself emerges from the digital ether, then, I don’t know. Hide your children, I guess.