
Introduction
I recently read an article which showed that LLMs prefer their own output above that of humans, which immediately struck me as an expression of ingroup preference.
If you’re not familiar with the idea, ingroup preference is when you favor individuals from your own group. For example, being kinder, more likely to share resources, and less likely to judge them harshly.
If you’re more willing to hire your friend, marry someone from your own church, or more willing to forgive your own sister, that means you have ingroup preference. If instead you’re more willing to hire a stranger, more willing to marry a foreigner, and more willing to forgive people of another religion, you have outgroup preference.
Some people have argued that ingroup preference is a “social pathology”, though research indicates that has an important biological basis, and is required for healthy people and social systems. This recent research of LLMs also lends support to the idea that it is an emergent feature of culture: anything that has ideals, symbols, or archetypes will develop ingroup preference.
Ingroup Preference is Biologically Adaptive
Hamilton’s Rule
Cooperation and self-sacrifice among animals has been observed for centuries, and often perplexed observers. A narrow understanding of natural selection would say that animals should only sacrifice their lives in protection of their own children — but that’s not the only case we see it in.
We sometimes see cases where individual animals who have no offspring will sacrifice themselves so that other members of their families will be able to survive. This mechanism is called “kin selection”, and is the biological foundation of what we would label as ingroup preference and altruism.
Mammals Require a Lot of Care
Different animals have different reproductive strategies. Some release sperm and eggs into the water, never to be seen again. Some lay eggs but abandon hatchlings, and others feed their young for a few weeks. But among large, complex mammals such as whales, elephant, great apes, and humans it is very different.
Pregnancy among large mammals is very “expensive” in evolutionary terms. Pregnancies are many months long, even over a year in some cases. Children are juveniles for many years and require close supervision.
Human Children are Especially Costly
Compared to other animals, pregnant women are more incapacitated by their pregnancy. Human brains are so large that they cannot reach full development in the womb — their heads would be too large to fit through the birth canal! When we compare the abilities of a newborn human to, for example, a giraffe, we can quickly see that compared to other animals, humans are born “premature” and require about 2 years of intense attention before they reach the level other animals have when they are born or within the first 6-8 weeks of life: speaking, running, and eating normal food.
This huge amount of investment of time, energy, and resources calls for especially strong bonding between mother and child. Certain adaptations, such as initially favoring the father’s facial features, keep paternal investment high until the child is at least 5 years old. Children raised together form bonds through shared experience. Kinship links are pressed (siblings, parents, parent’s siblings, cousins) to care for offspring, which creates a large kinship network.
Ingroup Preference Is Automatic and Healthy
This kinship preference is so strong that when people seek friends and spouses outside of their kinship group, they still unknowingly choose people who are as closely related as 4th or 5th cousins.
Since ingroup preference is natural and adaptive, we would expect to see those who lack it to be suffering in some way. And that’s exactly what the data shows: those with low or non-existent ingroup preference are less satisfied with their lives and more likely to suffer from mental illness.
7 Pillars of Friendship
The 7 Pillars Are Proxies
Robin Dunbar (of Dunbar’s Number) has continued his research into how humans create their social groups. His research has shown that, as we would expect from Hamilton’s Rule, that family connections are privileged. Further, he has uncovered dimensions that control how deeply we are likely to connect with other people. He calls these the 7 Pillars of Friendship.
One of the interesting observations you can make about these pillars, is that historically they would have been good proxies for determining if you were related to someone. I’ve written a short piece exploring this before. However, due to the advent of travel and information technology, these pillars have become disconnected from genetic relatedness and now function as measures of cultural relatedness.
The Pillars Help Us Identify Ingroups
The more pillars you share with others, the more likely you are to form deep, fulfilling friendships and speak to each other on a very regular basis. You are more likely to rate them as having good traits, even if you don’t know them. To put it another way, the more pillars you share with someone, the more you consider them part of your ingroup.
When you struggle to find other people that you share pillars with, you become unable to fill in your social circles. This leads to isolation, feelings of loneliness and abandonment, and even depression.
Ideals, Norms, and Archetypes
Now we reach the point of the entire article, what made me feel compelled to write it in the first place: Weights in LLMs function like Platonic Ideals.
For example, each culture has various virtues, norms of behavior, and story archetypes. These shape how people within a culture act and how they expect others to act. Not every member of the group has to agree with these virtues, norms, or archetypes — but those that don’t should reasonably be able to predict that most members do.
For example, within a culture, there is some shared understanding of what a “hero” is: how he looks, how he speaks, why he acts, what he does, what he doesn’t do, and how his story ends. Someone from within that culture is most likely to create a story of a hero that checks all of these boxes.
These little boxes are directly analogous to the weights that LLMs use both to create and to judge. If an LLM has within it’s weights some platonic ideal of an essay, it will attempt to create that perfect essay. When it reads essays that it has produced, it is more likely to consider those essays as approaching the platonic ideal.
Interestingly though, the more sure an LLM is that it created something, the more likely it is to rate it favorably. The same can be seen in human works as well. For example, the LDS church did extremely well in the United States because it had a uniquely American origin story. People could clearly see the Church as a part of their own cultural story.
This isn’t some offhanded observation either. Preference for things produced by your own group in an important part of social cooperation. It helps to keep money and effort within a local community, leading to increased wealth and prosperity.
Conclusion
In conclusion, it’s clear that ingroup preference is needed for healthy people and societies. It’s not just biological, but an emergent property for any system, human or mechanical, that uses symbols and ideals.
Due to the negative feelings many have around ingroup preferences, this realization seems huge in relation to the research for my book. A strong defense of ingroup preference is essential to explaining the continuity and importance of Western Civilization.
Notes and Comments
- This article has been somewhat rushed, in an attempt to publish content more frequently. Consider it a rough draft rather than a final product.
- The article on LLMs is difficult and somewhat confusing: Panickssery, Arjun. “LLM Evaluators Recognize and Favor Their Own Generations”, 2024.
- For more information on the 7 Pillars of Friendship, Dunbar recently published a book geared towards a general audience explaining his research: Dunbar, Robin. “Friends”, 2021.
- When I refer to Platonic Ideals, I am not using it in the sense of a physical object which resides on another plane, but instead a shared archetype within a cultural setting.