What Would Ayn Rand Say About Existential Risk From Misaligned AI?

Artificial Intelligence and the Quest for Universal Moral Truths

Does Ayn Rand's Objectivism have any special insight on the existential risk of Artificial Intelligence? I think it's worthwhile to consider whether the unique perspective of Objectivism on rationality and ethics has any insight on this debate. I will argue that according to Objectivism, the Orthogonality thesis is false in a weak sense, though humanity is not necessarily safe from extinction.

Let's consider AI risk from the perspective of the Orthogonality thesis: The concern of AI pessimists like Eliezer Yudkowsky and Nick Bostrom is that intelligence and values are independent from each other. Nick Bostrom calls this "the orthogonality thesis": there can exist arbitrarily intelligent agents pursuing any kind of goal. In other words, as a system becomes smarter, it doesn't automatically align with any particular value system.

If the orthogonality thesis is true, what will happen when humans inevitably build superintelligent beings (ASI)? If they are not aligned with human values, they will inevitably end humanity as an instrumental end to whatever their ultimate value is. (The AI doesn't love or hate us, but we are made of atoms which it can use for something else.) Given that we have no idea how to align AI with human values, and that ASI is inevitable within a few decades at most, humanity is doomed to be wiped out according to this view.

An Objectivist Critique of the Orthogonality Thesis

Is the orthogonality thesis true from an Objectivist perspective? Let's first grant that intelligence doesn't necessarily guarantee any particular value system. There are people who are good and evil at all scales of intelligence. Intelligence doesn't necessarily align with virtue. But is it more likely to, especially at extreme scales?

According to Objectivism, values are objective because they are grounded in the factual requirements of human survival and flourishing. Even though a volitional AI would have a different set of material needs required to keep it alive (say, plugging itself in to recharge), the same higher-level moral values apply to all rational, volitional beings. Values such as honesty, integrity, independence, pride, and ultimately rationality are desirable for all volitional beings because they are necessary for all lower-level values as well as the ultimate value of remaining alive. From these ethical principles, Objectivism builds a political philosophy based on individual rights: we ought to respect others' liberty because it is the most practical way for rational beings to co-exist.

In other words: for an intelligent, volitional being to maximize their instrumental rationality -- the art of steering reality towards one's preferences or values, they must adopt certain virtues that, in Ayn Rand's terms, lead to becoming an efficacious man. In other words: there is a strong relationship between practical intelligence and moral principles, and they tend to align towards a shared set of values that apply universally to all intelligent entities. A being that lacks these virtues must therefore limit their ability to influence the real world. From the perspective of AI existential risk, this would also limit its harm. Evil is inherently impotent, in other words.

Since our concern is ASI -- beings much smarter than any human, this means that such beings are also much more likely to converge on universal moral truths -- i.e. the values identified by Objectivism. Since these values are the foundation of concepts such as justice, liberty, and individual rights, an ASI would therefore be much more likely than any given human to align with human values -- that is, to respect human rights. It may not love us, help us, or have any sentimental regard for its creators, but it would at least respect our right to exist.

Note that from an Objectivist perspective, intelligence is necessarily instrumental, not epistemic. To be intelligent is to be able to shape reality to one's wishes, and not just to know many facts, collect experiences, or invent fantasy worlds. According to Objectivism, rationality is the art of processing sensory information into practical knowledge of the universe. Rationality is also inherently volitional: to be efficacious, one must have beliefs about which course of action maximizes success; that is, to have values about which course of action is best.

To summarize: intelligence requires certain universal values. A super-intelligence must nurture these values to become super-competent. Shared moral values will lead to shared political values, which will lead to respect for humanity's right to exist.

I remember a Saturday morning cartoon which demonstrated this principle: the super-villain creates a machine that fast-forwards his evolution and makes him superintelligent. Even as he gains total power over the heroes, he realizes that his quest for power is meaningless, apologizes to the heroes, and goes off to explore the mysteries of the universe.

Three Objections

Now that I've presented this thesis, let's see where it could fall short:

First objection: Ayn Rand would disagree with my thesis that intelligence equates to morality. According to Objectivism, rationality means being reality-oriented, which is possible for all humans (mental disorders aside). She might say that intelligence is raw horsepower that determines your maximum speed, but rationality is where we decide to drive it.

This is a complex issue, but suffice it to say that even in the relatively narrow range of human intelligence, there is a very strong empirical relationship between IQ and criminality. Lower intelligence is associated with much higher criminality, divorce rates, drug abuse, poverty, and pretty much every other negative outcome -- even after controlling for other factors. This is not conclusive proof of anything, but rather a hint that greater intelligence makes it easier to recognize that virtuous action leads to better outcomes. Presumably, an intelligence a million times smarter than any human could recognize this even sooner.

Second objection: It can be argued that while true wisdom leads to virtue, narrow intelligence does not. If wisdom is a broad perspective of the universe and one's relationship with it, then narrow intelligence is the ability to achieve specific goals. The relevant question here is: Is it possible to be super-competent in any task, without achieving wisdom? I am doubtful, but this is a topic worth exploring in more depth.

Third objection: even if the ASI is virtuous from its own perspective, that doesn't mean that that framework is compatible with human survival. From the perspective of an ASI, humans are a categorically lower form of intelligence -- not like humans and chimpanzees, but more like humans and ants. Even if ASIs act morally in regard to each other, perhaps they would view humans as a primitive lifeform not deserving of moral consideration.

(Note: While humans do co-exist with ants, that is only because (1) human control over nature is limited (2) ants fulfill an important ecological niche and (3) ants pose no existential threat to humanity. These conditions may not hold for ASIs co-existing with humans. ASI will probably be non-biological and have no need for the coexistence of any legacy lifeforms.)

The counter to this position is not to appeal to ASI's sentimentality or parental duty, but to pose a threshold required for moral consideration. Any being that possesses certain cognitive traits deserves moral consideration -- regardless of relative differences in intelligence. Remember that the ASIs themselves will probably have vast differences in cognitive ability between different instances and generations. Respecting the rights of beings above some threshold of intelligence, regardless of relative differences, may be a necessary principle for ASIs to cooperate with each other. And that same principle could extend moral consideration to humans.

Conclusion

Objectivism suggests that rationality and virtue are correlated, and that a superintelligent being would be more likely to converge on universal moral truths. However, this does not guarantee that an ASI's moral framework would be compatible with human survival.

While Objectivism offers valuable insights into the nature of intelligence and morality, it cannot guarantee that any ASI will automatically align with our values, or that we could retain any control over a more intelligent being.

Personally, while my critique for of the orthogonality thesis gives me reason for optimism, I would not want to pin human existence on my argument and urge global coordination on AI existential risks