What can AI learn from elephants about compassion?

Yesterday I saw a nature documentary with my daughter. A baby elephant got stuck in the mud, and its trunk and mouth were trapped under the surface. The baby’s aunt came to rescue it, using her legs to dig and move the baby, and using her trunk to lift the baby’s head and clear the thick mud away from the baby’s mouth and nasal openings.

While the aunt worked to free the baby, the rest of the herd watched in what could only be described as anxiety. As a human might, they shifted their weight from one foot to the other as they watched with eyes fixed on the baby.

As I watched, I realized that I was holding my own breath until the baby could finally breathe again. It was very uncomfortable to watch, and there was a great sense of relief when the baby finally got free from the mud pit.

Let’s use this as an opportunity to dig into the meaning and mechanics of compassion, with the hope of uncovering some insights we could use in building benevolent machines with the capacity for some form of compassion.

We can start with a really basic question which may seem silly at first: with so many other entities present at the scene, how does one choose where to place attention? What if a buffalo was eating some grass nearby while all of this was happening? Would the death knell of the vegetation be as compelling? Certainly not to an elephant or a human, but why? Why do we pay attention to the baby elephant in peril but not to the dying grass?

Perhaps we have a hierarchy of importance that tells us where to attend? For a moderately altruistic person in western society, it might go something like this:

Family > friends > self > other humans > other mammals > other vertebrates > other animals > plants and other non-animal species > non-living things

So, these things become the focus of our attention, and our attention extends from them to anything that benefits or threatens them. Thus, a man telling your wife that she has just won the lottery will get your full attention, unless there’s also an uncaged tiger preparing to spring at your wife. Then the lottery guy may lose your attention.

Our attention generally goes to threats > benefits. Timing matters here, too. Generally, earlier > later. Experience tells us that things which are changing/moving tend to bring earlier benefits or threats, so changing > static. Things which teach us or bond us together can go into the benefits category, which is why we attend to one another in conversation. In general, interacting > not interacting.

To recap, here are some general observations about what gets our attention:

Before entity recognition:

• Changing > static

• Interacting > not interacting

After entity recognition:

• Family > friends > self > other humans > other mammals > other vertebrates > other animals > plants and other non-animal species > non-living things

After interpretation of interactions:

• Threats > benefits

• Earlier threats > later threats

• Earlier benefits > later benefits

If we want machines with benevolence and compassion, we’ll have to give them some core values like what we saw above. One open question is, where do we place the machine’s self in that hierarchy? This becomes especially important if the machine reaches generalizable superintelligence. Before we go down that rabbit hole, it’s important to recognize that there will be many forms of artificial intelligence which could benefit from some core values, and which completely lack the capacity to grow up to be a generalizable superintelligence. Let’s pull back and focus on those for now. I’ll leave the doomsday scenarios for another post.

Okay, so now we have some framework for where we place attention. As we process our visual inputs in this situation, we recognize the entities of the baby elephant, the herd of onlooking elephants, the aunt elephant, the nearby buffalo, the grass, and the mud. None of these are immediate threats to us and most are standing relatively still, so our attention is drawn to the movement in the mudhole, where we see the highest ranking entity in the scene: a baby elephant. Our attention is held even more firmly when we recognize that the baby’s life is being threatened. We allow our attention to extend to the aunt elephant, who is also moving and who is trying to remove the threat to the baby elephant. To me, this framework for attention seems to be working reasonably well.

But wait. After we recognize entities and start to focus on them, how do we recognize what is happening? DALL-E and its cousins have shown that neural networks can generate images of actions in progress as described in a prompt. For example, let’s give DALL-E the prompt: “A photograph of a baby elephant suffocating in a mudhole.” This is what DALL-E creates:

Then, if we prompt DALL-E with “A photograph of a baby elephant being rescued from a mudhole by its mother,” DALL-E creates:

DALL-E was only trained on still images, so it's unclear how well it actually understands the concept of "rescue," but it seems reasonable that a system trained on time series data such as video could more fully understand concepts of action such as “rescue.”

Let’s return to our initial reason for thinking about elephants and mudholes. We hoped to uncover some insights we could use in building benevolent machines with the capacity for some form of compassion. We have learned that machines need a sense of what is important, which entails recognizing an entity and placing it in a hierarchy of importance. Machines also need a way to recognize situations, changes to situations, a way to judge whether the change represents a benefit or threat to some entity, and a capacity to predict how soon the benefit or threat will take effect.

What insights have you had about the meaning and mechanics of compassion as you pondered the baby elephant’s rescue? I’d love to hear them!