Who Supervises the Supervisor? A Riff on Machine Learning
“Supervision” implies benign intent, but it can be cruel and coercive.
WAUWINET, AUGUST 6, 2026. My effort to learn how machines learn has me mulling over “supervision.” A lot lurks behind that bland term.
In Why Machines Learn: The Elegant Math Behind Modern AI, Anil Ananthaswamy notes that until recently machine-learning programs have needed human supervision to recognize shapes, sounds, sentences and other patterns.
You train a program to distinguish ducks and geese, say, by feeding it images you’ve labeled “duck” and “goose.” These images correspond to mathematical patterns. The program devises two sets of equations: one set approximates “duck,” the other “goose.”
Then you feed the program unlabeled images of ducks and geese and ask what it sees. When you tell the program it’s wrong, it tweaks its equations and guesses again. In this way, the program learns to perceive ducks and geese.
Human-supervised machine learning is laborious, expensive. Recently, Ananthaswamy reports, engineers have devised programs capable of “self-supervised learning.” The key is giving the programs access to an enormous database, like, say, the Internet.
A self-supervised program doesn’t need humans to label data and tell it when it’s wrong. The program trains itself, making guesses and checking its answers against the database. In this way, a program can teach itself to tell ducks from geese and, far more impressively, to chat with humans about quantum mechanics, Brazilian jazz, whatever.
Self-supervised learning underpins recent advances of large language models. Ananthaswamy quotes AI researcher Alexei Efros: “The revolution will not be supervised.” Gil Scott-Heron would no doubt find this quip super funny.
As I say above, Ananthaswamy has me contemplating “supervision.” Although the term implies benign intent, supervision can be coercive and cruel. Who, then, supervises the supervisor? Below I consider various kinds of supervision, ranging from natural selection to divine intervention.
NATURAL SELECTION
My early education was supervised by my parents and teachers, whose education was supervised by their parents and teachers and so on. Ideally, each of us becomes self-supervised, but do we ever really escape our societies’ meta-supervision?
And how did our species learn to learn in the first place? The ultimate cause of our capacity for learning is natural selection, which cares only about one thing: reproduction. Pardon the anthropomorphism, but you can think of natural selection as evolution’s supervisor, nudging species to be faster, stronger, smarter or whatever it takes to pass on genes.
Natural selection has produced a dazzling variety of species. Forget bacteria and bugs, just look at birds! Natural selection also created a species that created language, music, poetry, mathematics, democracy, science yada yada.
But natural selection is a brutal supervisor, akin to the sociopaths running Squid Game. For every winner who survives long enough to reproduce, countless losers perish without spawning. Whole species vanish. Surely there must be a better system of supervision.
HUMAN SUPERVISION
We humans have been supervising the evolution of species, including our own, for thousands of year. We breed animals and plants for food and other purposes. We invent sewage system, vaccines, antibiotics and other life-extension technologies.
Impressive, but now there are too many of us. We have driven other species to extinction and polluted the land, air and water. We have made the means of our own destruction, including nuclear weapons, bioweapons and, perhaps, self-supervised AI.
Throughout history, some groups of humans have felt superior to and hence entitled to supervise other groups. This supervision has ranged from indoctrination to enslavement, sterilization and extermination. The thinking behind this sort of malign supervision hasn’t exactly vanished.
My point is that the record of human supervision, like that of natural selection, has been mixed.
THE MARKET
Back to self-supervised AI. It doesn’t really exist yet. Humans still oversee the design, construction, release, re-design, re-release and marketing of ChatGPT, Claude, Gemini, Grok and their Chinese counterparts.
But who, really, is supervising AI now, guiding its evolution? Tech moguls like Altman, Musk, Zuckerberg, Karp? Heads of state like Trump and Xi Jinping? These big-shots can nudge AI in this direction or that.
But AI’s ultimate supervisor, I submit, is capitalism, or the market. The market is far more powerful than any individual humans. It is as ruthless and monomaniacal in its way as natural selection. As Marx warned, the market cares about only one thing: profit.
As long as AI appears monetizable, the market will push it forward, even if the risks AI poses to humanity—massive job displacement, metastasizing data centers, wars waged by robots--outweigh benefits.
AI SUPERVISES US
Here’s a prediction: Through a perverse positive-feedback effect, social chaos triggered by AI might accelerate acceptance of it. The more AI disrupts our lives, the more we will rely on it, because we trust AI more than humans to supervise us. This will be true even (especially?) of people in power.
Already, many humans I know turn to AI for advice on career, money, relationships, health, you name it. Current AIs are too obsequious to be truly effective supervisors. I’m guessing the market will encourage AIs to become more blunt, and strict. When it thinks we’re wrong, AI will punish us, spank us, as it were, to ensure we learn our lesson.
At some point, neither the market, governments nor individual humans will supervise AI. Rather, AI will supervise us. Does this prospect alarm you? It alarms me. Once we grant AI the power of supervision, I’m not sure we can get it back.
GOD, THE ULTIMATE SUPERVISOR
Given the unreliability of these supervision schemes, it’s not surprising that many people hope a benign supernatural potentate is supervising us. Robert Wright considers this idea in his new book The God Test: Artificial Intelligence and the Coming Cosmic Reckoning.
The book is the culmination of Wright’s ferocious career-long effort to figure out why we are here. Surely it’s not to live in pain and fear. Wright proposes that maybe—not definitely, Wright gets mad if you say definitely—god exists. And not some cold, distant god but a god who cares about us, who wants what’s best for us.
Maybe, Wright says, AI presents us with a test: We can let AI divide us and risk annihilation. Or we can come together to form an AI-enhanced world of which a benign god would approve--if he/she/it/they exists.
Wright is speaking to my school October 6, so for now I’ll only say this: I too sometimes suspect the history of life, and of humanity, is too directional and dramatic to be a mere accident. So I don’t rule out the possibility of god.
But I do rule out—and I’m pretty sure Wright agrees with me on this--the possibility of a god who will save us from ourselves, from our own meanness and foolishness. God can’t save us. Neither can the market, or superintelligent AI.
We can only save ourselves, by getting better at self-supervision. We know what we must do. It’s simple, it’s a lesson we learn as children. We must stop hurting each other. If we can’t manage that, we’re fucked.
Further Reading:
Okay, AI Is Cool. But I Still Loathe It
An AI-Hater Tries to Learn the Math That Underlies AI