Remember Edward Snowden? The NSA contractor who leaked classified documents, publicizing the government’s surveillance apparatus? His action brought to light major privacy concerns, and led to the government ending its bulk data collection programs1. Many of us think that was a good thing, and that Snowden did good.
What if some AI made a similar decision? Say you work at a hedge fund, and are researching trading strategies using an AI. Your AI determines that the strategy you are contemplating amounts to illegal market manipulation, and proceeds to leak your source code implementing it to the Wall Street Journal and the SEC. Did the AI do good?
We are constantly deciding what to do, increasingly enlisting AI to help us. We use the good as a criterion, and trust that our AI use is aligned with that criterion. Goodness, however, is difficult to define and elaborate. We could associate goodness with general happiness, and choose the action that maximizes general happiness. But that approach—utilitarianism—has serious issues2. There are philosophers who consider the question of what the good consists in intractable3. We typically make do by relying on our intuition, an affordance not available to AI.
That is why we shall turn instead to examine what typically accompanies good action. Can we look for indicators, tells, that can help us—and our AIs—recognize when we are or are not on the straight and narrow? I present three. In increasing levels of inwardness: means and ends, vice and virtue, shallow and deep.
Walking the labyrinth at Chartres Cathedral. Photo by User:Maksim at Wikimedia
We’ve all heard the cliché the ends justify the means. Snowden violated the terms of his contract by taking with him classified documents and then disclosing them. He did his employer wrong. But then, his employer was doing the public wrong. Didn’t Snowden’s action help correct that wrong? Well, not according to Immanuel Kant. He would say two wrongs do not make a right.
If everyone broke agreements for what they deem a positive end, agreements as such will lose their instrumental function. Our economy and society will suffer. That is how I’d apply Kant’s thinking to Snowden’s case. Kant’s categorical imperative, in its first formulation, reads: “Act only on that maxim which you can will as a law for all rational beings.”4
A philosopher friend of mine had another take. He said Kant’s categorical imperative isn’t about global consequence (”rule-utilitarianism”), but about respecting the dignity of all other rational beings5. Even so, the conclusion is effectively the same: Deceiving your employer, by smuggling documents out and reneging on your non-disclosure promise to them, is just plain unethical.
Portrait of Kant, 1790. Painter unknown, possibly Elisabeth von Stägemann.
The Danish philosopher Soren Kierkegaard would say pretty much the same thing. For someone willing the Good, he writes, “the means and the end are one and the same thing”. One is not responsible for whether they reach their goal, “[b]ut without exception, he is eternally responsible for the kind of means he uses”6. Being particular about means may seem burdensome, and means may not be as compelling as a Muskian vision. But attending to the means focuses us on our present actions and what we can control, granting us a measure of peace. By not fretting about whether we reach our goal, we avoid the “passionate one’s torment”7.
So here is an indicator we can use to check our actions: Do we attend to our act’s right-ness as much as if not more so than its goal? This principle is easily translated to AI actions. An AI agent may not take an unethical step to reach a user’s requested goal, no matter how urgent and important-sounding the goal is made out to be.
A college student loves playing the piano, but she worries she can’t make a living doing so. So she studies accounting instead of music, and becomes an accountant. Did she choose well? If your intuition says that it depends, you would find company in Aristotle.
Aristotle writes that courage, like most virtues of character, is a balanced condition between two opposing vices; in this case, between cowardice and foolhardiness8. Was our student friend too cowardly to take a chance, or simply being realistic? Perhaps she likes accounting too, and exercised the virtue of practical judgment (phronesis)9, judging where her personal interests meets the world’s needs.
Other virtues Aristotle develops include temperance, generosity, justice, and amiability. Together, they make the following query a decent tell of choosing the good: Are we acting in accordance with the virtues?
Being-at-work in accordance with virtue comes with two additional benefits. We build good character through habituation. And we experience a kind of happiness (eudaimonia)10. These two benefits serve to reinforce virtuous behavior. But neither of them accrue to current AI, which neither learns from its actions nor experiences states like happiness.
An AI can still consider Aristotelian philosophy when answering human questions. If I were to ask AI how best I can give to charity, it can guide me to an outcome in line with the virtue of generosity, striking a balance between stinginess and wastefulness. But the AI itself is not exercising generosity. How can generosity apply to an AI that has no bank account and needs no retirement savings?
That brings us to the last tell. Unlike virtue, it might work well for AI.
How deeply thought is the action in question? The more fully considered the action, the more likely it aligns with the good; at least when the actor is well-adjusted and not compromised in some way. Let me illustrate via a story I heard from Sharon Salzberg, a cofounder of the influential Insight Meditation Society11.
Sharon was in India, studying Buddhism with her teacher Dipa Ma. When on a trip to town, she was harassed by someone trying to snatch her handbag. Distressed, Sharon asked Ma how she could still be expected to maintain a compassionate composure. Ma replied: If that ever happens again, with a smile in your heart and all the compassion you can muster, use your umbrella to give your assailant a nice thwack.
An umbrella wielded in that manner is a considered response, not a reflex reaction. Responding thoughtfully rather than reacting reflexively is one of the aims of ‘mindfulness’, as taught by insight meditation teachers such as Salzberg. A similar principle is developed by Spinoza in his The Ethics. He distinguishes acting (“whereof we are the adequate cause”), from being acted on (“we are passive”)12. We are free when our actions stem fully through our nature, and compelled when determined by something else13. This parallels the mindful versus reactive modes of acting.
Portrait of Benedictus de Spinoza, 1665.
While we may intuit that Salzberg’s and Spinoza’s insights ring true, we could also appeal to neuroscience. Our brains have distinct regions whose activation patterns we can study. We can try to develop an indicator based on those patterns for when a person is deep and reflective in their thought vs. being shallow and reflexive. However, a device that computes such an indicator is likely workable only in a lab setting. But I know the perfect candidate—a candidate who only ‘lives’ in a ‘lab’: AI.
It is trivial to observe an AI’s artificial neurons activate as it goes about generating output. Patterns therein can indicate how much the AI is drawing upon its sum knowledge. For example, the more of the neurons in the early layers of the AI’s network activate, perhaps the more primal and broadly-sourced its output14. As current AIs are generally trained well, with “good” training data, this would be a tell that the AI is not acting waywardly. I do believe frontier AI labs have the chops to develop such an indicator, an indicator that the AI had “deeply thought”.
We hardly discussed what the good consists in. I doubt we can come to agreement in a pluralistic society like ours. But I suspect we can broadly agree on some indicators that accompany choosing the good, tells such as no-shortcut-taking, virtue, and thoughtfulness.
One tell we did not discuss is sympathy, a tell like virtue that does not work for AI. As we contemplate an action, we can ask ourself Is that kind?, and imagine its felt moral consequence. But it is difficult to envision AI developing sympathy. AI has never experienced pain or hunger, and to it mortality and vulnerability are alien concepts.
For better or worse, the good we know of, however each of us might conceive it, inheres in the human condition. Any indicators we devise for AI may necessarily fall short.
For example, what if one determines they add to the general happiness by murdering someone who was making many others miserable?
In his influential Principia Ethica, 1903, G. E. Moore argues that goodness is an undefinable (”non-natural”) property of whatever possesses it, and cannot be further defined in terms of other definable (”natural”) properties. Earlier in A Treatise of Human Nature, 1740, David Hume famously held that “reason is, and ought only to be, the slave of the passions”: We cannot rationalize our motives; reasoning happens only around how to accomplish the end our passion has decided on.
Groundwork of the Metaphysics of Morals. Immanuel Kant, 1785.
cf. the Golden Rule: “Do unto others as you would have them do unto you.”
ibid. This interpretation leans into the second formulation by Kant of his categorical imperative: “Act in such a way that you treat humanity, whether in your own person or in the person of any other, never merely as a means to an end, but always at the same time as an end.”
Purity of Heart is to Will One Thing. Soren Kierkegaard, 1847. Ch. 14.
Nicomachean Ethics. Aristotle, 350 BC. Book 2.
From her talk at the New York Insight Meditation Center, ca. 2015. Also recounted in many places online, as here:
The Ethics. Benedict Spinoza, 1677, tr. Elwes. Part III, Definition II:
“I say that we act when anything takes place, either within us or externally to us, whereof we are the adequate cause; that is (by the foregoing definition) when through our nature something takes place within us or externally to us, which can through our nature alone be clearly and distinctly understood. On the other hand, I say that we are passive as regards something when that something takes place within us, or follows from our nature externally, we being only the partial cause.”
ibid. Part I, Definition VII:
“That thing is called free, which exists solely by the necessity of its own nature, and of which the action is determined by itself alone. On the other hand, that thing is necessary, or rather constrained, which is determined by something external to itself to a fixed and definite method of existence or action.”
This hunch is based on my previous professional experience with developing and training AI models,, albeit at a much smaller scale than current LLMs.
No posts