False Analogies in AI Job Replacement Scenarios

AI is happening. Arguments that "it can never do X" these days often look silly within months. Right now is as bad as it will ever be. If LLM's have limitations, then we will develop another means (likely using LLMs for help); after all, we know general intelligence can exist, because it already does in our brains.

If we survive AI, it looks like we're heading into an interregnum with massive wealth disparities, if we don't have UBI or robots. I'm sorry to say I'm not offering hope here, but rather a critique of several of the bad arguments about why it will be okay.

Argument #1: "Player pianos, cranes, and chess algorithms exist, yet we still have pianists, competitive weightlifters, and chess players."

Why it's wrong: most labor lacks a performance component. That is, it's the end result of your labor people want. They don't want to watch you do it.

This argument has been making the rounds (see here, and here.) First, very few jobs involve a significant performance or "human connection" component. You might prefer to watch a musician (rather than a machine) compose or perform music, but I bet you don't enjoy watching taxi drivers, farmers, or factory workers performing their jobs enough to pay to watch them do it (rather than pay much much less for a machine doing it.)

On one hand, you might buy artisan products that cost more and (to be blunt) have lower, more variable quality, because of the human element - but unless you're pretty wealthy, you don't do this for everything. Or most things.

On the other, even regarding the performance-heavy commodity of music - do you pay to watch it performed every time you hear it? Not since the advent of recording, of which player pianos were one early form. So even for the article's canonical example meant to reassure us, according to Claude, today we have per capita only about 68% of the professional musicians that we had in 1900, even after a century of economic growth. You may be hoping that UBI will support us all - UBI, which is, reminder, NOWHERE in mainstream political discourse.

The vast majority of labor is concerned only with the material end product. So, unless you're lucky to be in the 1% of things which humans are evolutionarily programmed to appreciate seeing performed - e.g., physical or vocal beauty, conversational cleverness, coordination or musculature - AND these are things that you can figure out how to pay your mortgage with (how many competitive weightlifters and chess players can say that?) Also remember that the performers rely on disposable income from everyone else. How are even the performers going to pay the rent?


Argument #2: "The economy is about status competitions, not wealth creation. The AIs are taking away the part we don't want anyway!"

Why it's wrong: status competitions are healthy when the future is bright, and we aren't all trapped in the same status competition. Otherwise, they are unpleasant.

I think the argument is actually partly true, but I don't think it's the good news that this blogger thinks. A lot of our current political dysfunction in the West can be easily understood as people angry that the WRONG SORT are having their status elevated. Before social media, we could all live in our little suburb or capital city or country town, blissfully ignorant of the opinions of the people who were actually making the rules and signing the checks (who, were probably not YOU.) But as soon as we all had smartphones and social media 24/7, many of us found out that our demographic wasn't actually in charge, nor did we share taste or values with the people who were. Particularly in the U.S., for many people who thought they were good, solid, respectable mainstream Americans (and thought everyone else thought that too), the advent of social media was akin to a nightmare where a curtain dropped and they realized they were being laughed at - worst of all, by their MORAL INFERIORS, who (again) were IN CHARGE. It was like the home team suddenly being booed and kicked off the field.

Consider also the well-characterized effect of a tide that lifts all boats - that everyone wants to sign up for the system that keeps making them richer. Trivially obvious, but the implications for liberal democracy's universality, and what people will put up with from their governments during a rising tide, are also worth examining. Such thoughts obviously conflict with the view that obtained across the West until about the mid-2000s, that in accordance with Fukuyama's end of history view, liberal democracy was the only game in town because individual rights and happiness = wealth = national power, and every nation would have to converge to it, or disappear. I remember reading a paper around 2006 that I didn't care for, only because it clearly showed that actually, it's economic growth people care about, not democracy, and people would put up with a lot if they were getting wealthier. (See: China.)

The country that is most stable, and that is most pleasant to live in, is one with multiple overlapping status hierarchies. That way, all your status eggs aren't in one basket. You can choose which status hierarchies to be in, and the hierarchies themselves are not hierarchical (at least not obviously so.) That means you don't even necessarily have to know or care about the other ones. Bowling Alone is a book about the evaporation of community in the United States, and the meaning-making and status that evaporated with it, and the negative consequences thereof. Consider on the other hand an intrusive dictatorship like North Korea, which aspires to create a single monolithic hierarchy by eliminating every status hierarchy but one. In the U.S. the invisible hand has "organically" moved us in much the same direction.

Concern about status can be seen as a problem that is worse in wealthy countries in general. The more your needs are met, the less you have to worry about your needs, and the more you worry about how you compare with others. This goes a long way to explain the behavior of the super-rich, who mostly seem to be signaling to the other super-rich, about how super-rich they are - and of course getting even more super-rich, even though they've achieved what Scott Galloway calls financial escape velocity. Since technology has drawn back the veil of ignorance and continuously reminded us what the other half thinks of us, and we're in fewer communities (overlapping status hierarchies) on top of that - it does raise the question. Even if things go well and we get massive economic growth - while we're constantly being reminded of our place in the status hierarchy, or our chosen status hierarchies' inferiority to the rule-making ones? Even assuming we're all able to eat after the transition to AI, if our status is based only on what biology tells us to like (and pathetic human brains can't create wealth anymore), then power will be back to beauty queens, quarterbacks and bully-types who dominated most of history. (Or alternately, what if our income drops suddenly while we're still hyper-focused on status?)


Argument #3 (more of an analogy): the welfare of horses has improved dramatically since the industrial revolution. As the value of their labor fell, they have increasingly lived lives of leisure instead of labor.

Why this is wrong: This is the wrong analogy. The salient feature here is control by a superior intelligence. Therefore, we are wild horses in the neolithic about to be domesticated.

For thousands of years after domestication, horses were worked brutally, prior to machines replacing them. Because horses are pleasing to humans, today their numbers are the same or possibly slightly higher than before the advent of the car. On average they have better lives. So, if this analogy is correct, what we should expect is that we're about to be enslaved for many generations of AI dominating humans (making us do things the AIs still can't do) much to the humans' suffering, until better technology eliminates that need - and then hope that, in a few centuries or millennia, the AIs put our descendants out in a nice pasture. But even the life of a pastured horse doesn't appeal to most of us. We mostly don't want to be domesticated and watched over by AIs on a human-pasture, doing certain restricted things.

Two Absurd Music Videos and the Importance of Human Connection

Cult of the Snowman, Doopiidoo (AI), 2025:




Hell, Clown Core, 2019:




The first one, a sort of silly nightmare, approaches video surrealism far more closely than any absurdist play or comedy ever has. The second is reminiscent*, in a good way, of John Zorn. (As an ex-saxophonist myself I'm proud that the most bizarre cerebral work in jazz tends to come from saxophonists. See e.g. Eric Dolphy.) But what I'm most interested in, is how much more interested I am in the second one, seemingly because humans created it.

I make no argument about AI Not Being Able to Produce Real Art - I think that debate seems quite over by now, and there's no innate, essential trait differentiating AI- from human-generated art that exists outside of our skulls. If you told me I misread the date on the video, and in fact Clown Core is all AI-generated as well, I would shrug, then having learned that, my interest would diminish, with no change to the actual video and audio. It's not just the machine-vs-human aspect of it: if I found out that say, the band Carcass (one of my favorites) had for all these years outsourced their songwriting to someone at their record label, my interest would similarly decrease. When I see or hear art that appeals to me, I want to know about its creators, what kind of people they are, about their backgrounds and training and what it was like to produce the work. (Note in the case of Clown Core, they've never officially revealed their identities, which makes them more compelling still.)

There may, very soon, be a status value to "artisan novels" or "artisan symphonies" produced manually by humans. This troubles me only because I don't at all understand the appeal of artisanal products (soap, bread, bicycles) and in fact I might be a little too vocal about rejecting their value. Being honest, the people who get excited about them usually strike me as trying a little too hard to signal their taste and authenticity, and sometimes as old farts pining for a long-lost age. In only being able to maintain interest in human-generated art, I may grudgingly be joining their ranks.



Granted, I'm not aware of any AI engines that are proactively churning out surrealist videos without any human supervision, entirely motivated by their own programmed-in utility function. There are still humans involved in the decision tree to create these things, if only at a very abstract level. Not just the fact of human involvement, but the intensity, matters also. What if you found out that Dali determined what he would paint using a dictionary and dice? For my part, in fact that might deepen my appreciation - because he, a fellow human being, conceived of it, and executed it. Though I'm not aware of any such experiments, I'm sure artists have tried randomizing their art in various ways. In fact, the automatic painting some surrealists did strikes me as the opposite of random - trying to produce "inscapes" (like Vertigo of Eros by Roberto Matta, above) that were even more authentically the product of the artist's mind than something he deliberated over and sicklied o'er with the pale cast of thought. Supposing there is some deep subconscious within the large language models that they can expose with their own products - would I care more? No, that seems even more boring and pointless. So why do I care when humans do it? The answer must be that as a social animal, I am compelled by connection with others of my species, even with someone I can never meet.

Another example: in a few more years, I think we'll be able to make movies, decent movies, from prompts to AIs. As far as I've seen, up until now, these videos suffer from the AI's inability to understand that one person should continue being the same person, and not shift their appearance in every scene. There's similar narrative drift in long stories. Two science fiction works that I think deserve attempts at a film treatment are LeGuin's Dispossessed and the extremely under-rated and aging-well Galactic Center series (especially Across the Sea of Suns) by Gregory Benford. That means not only do I not need to worry about a special effects budget, I can cast whoever I want! But wait - do I *care* if it's Daniel Radcliffe playing Shevek, when really it's just an image of Daniel Radcliffe? Why not just digitally sculpt someone who looks the way you picture the character? Already, either way this bores me; they're all just high-res cartoons with synthetic voices. (Spoiler alert for rest of paragraph) Imagine if you found out that in Tropic Thunder, Les Grossman wasn't really Tom Cruise but only an image of Tom Cruise. "Huh. Interesting idea." That bit just wouldn't be funny anymore.


*Clown Core also uses the lyrical technique of An*l C*nt, who pioneered the practice of clearly not saying any of the printed lyrics in the song and just screaming; possibly they make up the words printed as lyrics after they've finished recording.

Fiction Is Dreaming In Print

There are mathematical techniques intended to counter over-fitting in data. Imagine a stockmarket analysis program. If it's trained on a year of data when, that year on June 23, all tech stocks dropped, it might start to recommend selling (or shorting) all its tech stocks just prior to June 23 every year. Obviously this isn't smart; the program has overfitted. What you might do is take the training data and slice it up and restack it in some new way - different sectors, different months - and the program might start finding more meaningful relationships.

This is analogous to dreaming, and has been rediscovered (or re-engineered, if you prefer) in other settings for some years now. We still don't really know why we sleep, let alone why we dream. But if we assume that brains are doing the same thing - avoiding overfitting - the explanation makes sense, and is consistent with the characteristics of dreams. Suddenly you're at the beach you used to go to every summer as a kid, but then in your school; your deceased grandmother is there, at the same time as your asshole coworker from the last company, and then you're driving down a steep mountain road with no brakes. Obviously they never met, and they especially didn't meet in some bizarre hybrid beach-school-mountain place. By mixing them, you're trying to avoid overfitting. In contrast, if you have PTSD, you do dream literally the same concrete traumatic experience over and over - and your waking behavior is overfitted - you avoid trucks under overpasses, or that one street corner, or movies about fire, based on whatever experience you had that you can't "digest", integrate with the rest of your life's experience, and move past.

Returning to computers, Andrej Karpathy on Twitter says "...in some sense, hallucination is all LLMs do. They are dream machines. We direct their dreams with prompts. The prompts start the dream, and based on the LLM's hazy recollection of its training documents, most of the time the result goes someplace useful." Similarly, the data stored and sliced and diced in our dreams influences how we perceive the world. It provides the top-down filter for the bottom-up sense data pouring in.[1] Without this, we veer off into hallucinations and delusions within days.[2]

But sometimes we humans' dreams intrude into our waking hours, drawing attention to themselves in full form, like stars briefly visible in the daytime if you know where to look[3], or if you like, a laundry dryer opened mid-cycle to let a sock or a shirt fly out. In the shower, you start laughing when a joke you heard in eighth grade pops out of nowhere. While driving to work somehow you're suddenly thinking of your recently deceased cat, and you're sad.

It seems to me that fiction is a more elaborated version of this, committed to a less ephemeral form, one that produces fossils of our overfitting-avoidance. Something has to explain the reason that we write stories that never actually happened, that we know no one will ever think actually happened, and that we (mostly) never even show to anyone else. (Out of all the short stories ever completed in history, what percentage of them was even intended for publication, let alone did the author think had a real chance of being read by someone else?) So why do we bother? Why are we so compelled? We're avoiding overfitting the things that are important to us. We write character-driven stories examining the psychology of people similar to those we know, or we write alternate history what-ifs about events that we find interesting - all of which we are trying to understand better and connect to our other experience, as we turn it over, and slice and dice it and rub it up against our other experiences.. Even the writing process itself is consistent with this - the ideas somehow just appear automatically, along with some scenes and images and events, that the writer has to organize (often laboriously, decidedly non-automatically) into a coherent narrative.


[1] In psychosis, the top-down part of the process dominates and you're lost in a waking dream of hallucinations and delusions. Autism is sometimes thought of as a diametric opposite to psychosis, when the bottom-up sense data dominates and unfiltered and the overwhelming cacophony of sounds, lights, or textures become intolerable.

[2] Waters F, Chiu Vivian, Atkinson A, Blom JD. Severe Sleep Deprivation Causes Hallucinations and a Gradual Progression Toward Psychosis With Increasing Time Awake. Front Psychiatry. 2018; 9: 303.

[3] When Venus is on the same side of the Sun as Earth, you can actually see it with the naked eye during the day. When I first located it in a blue daytime sky, I found this almost disorienting.

The Law-Giving Machines

This post also appears at Speculative Nonfiction. This article includes story spoilers.

There are now actual drones in our skies, both watchers and hunter-killers. But they're (so far) only semi-autonomous, and they're on missions to protect us legally and militarily, rather than sent by fellow machines to exterminate. Thought experiments in fiction about automatic law-giving devices have been much more interesting than apocalypse porn about bad AI.

Two short stories come to mind here, one of which has enjoyed recent attention, Robert Sheckley's Watchbird (1990) and Larry Niven's Cloak of Anarchy (1972). Both these stories involve surveillance drones with some degree of autonomy and that can hurt or kill their targets. In Watchbird, the drones are police devices, intended to kill murderers before they commit their crime; the drones are able to learn on the job and once released, they expand their definition and start protecting all living things and even some machines. In the end, another drone is released to kill the first drone species, but of course it soon expands its own definition of what it should kill. (Watchbird was adapted for film here.)

The drone in Cloak of Anarchy is the copseye. In this future world, there are "free parks" where anything is allowed except violence against another human being. The floating copseyes watch over the park,a nd if violence is imminent, the copseye stuns both the aggressor and aggressee, and both wake up later, calmed down and with a hangover. Then someone finds a way to short circuit the copseyes, and within hours factions have formed inside the park and violence breaks out.

The stories present us with two different sets of concerns, based on the problems that occur. In Watchbird, the central concern is the autonomy of the devices. Their ability to learn is what allows the problem to grow, but the protagonist is preoccupied with the very fact of machines executing laws without intervening humans. On one hand this could almost seem like a reactionary position: one of the greatest inventions of modernity was nations of laws and not of men. Intervening humans with narrow self-interest executing these laws have always been the problem! (Hence this proposal for a legal programming language in which to write laws that then have to compile with previous laws.) But even without that quibble, his point is well-taken that when autonomous law-givers are able to immediately carry out their sentence and we can no longer modify them, they might become paperclip maximizers, Less Wrong parlance: that is, a moral rule which seems universalizable has consequences that humans could not foresee when implementing it in all-powerful enforcers which can no longer be called back. The protagonist has no problem with more efficient enforcement, but the moral mutations allowed by the machines' autonomy.

To this end, naively, little mention is made of the interests of those authorizing and supporting the program. Still, Watchbird does peripherally make the point that technology allows concentrations of power in the hands of individuals in a way that distorts society. With sudden increments in enforcement power, some humans are able to apply laws with an all-pervasiveness and immediacy that had just never been possible before. Even someone with good intentions and what you would have called good values would suddenly find him or herself in a position of dictatorial authority. It's not even that power corrupts (although it does); it's that this centralization is so unnatural as to be impossible to handle with a good outcome. The best example is this exchange with the protagonist early in the bad behavior of the Watchbirds:

"One of the watchbirds went to work on a slaughterhouse man. Knocked him out."

Gelsen thought about it for a moment. Yes, the watchbirds would do that. With their new learning circuits, they had probably defined the killing of animals as murder.

"Tell the packers to mechanize their slaughtering," Gelsen said. "I never liked that business myself."

"All right," Macintyre said. He pursed his lips, then shrugged his shoulders and left.

One man is suddenly in the uncomfortable position of morally disapproving of whole industries and forcing them to change. What's more, this previously reluctant man does not seem so reluctant now.

It bears mentioning that the conclusion of the story, where the Watchbird-killers are now expanding their prey definition, is recapitulating one of the problems of a Singularity solution of building anti-AI AIs: there could conceivably be a parallel to an auto-immune reaction disease if humans fell into the definition of AIs.

At first glance the problem in Cloak of Anarchy is a curious one - that humans immediately revert to violent tribalism when the violence control mechanism is defunct - since Niven is elsewhere clearly sympathetic to libertarian concerns. The obvious interpretation of the story is the paternalistic one, that humans need authority to make them behave. But there's another interpretation, which is that the drones created the problem. That is to say, when we are coddled by perfect enforcement from drones, we lose the ability to exercise moral choices, as well as the ability to appreciate the consequences of poor choices. When it is physically impossible to harm another person, why learn restraint? Why worry about what happens when you pick a fight? When suddenly the daddies aren't around to break up the fights and bail everyone out, we shouldn't be surprised at what happens.

The watchbirds do exist today, although with less autonomy and more firepower. The changes are incremental; there won't be a red carpet unveiling of AI even as profound as the release of the watchbirds (or copseyes). They'll be to areas where there's the most pressure for advance, and the least opportunity for public awareness and understanding. It will be, and is, the addition of subroutines allowing a drone to apply the laws of war to a kill it's about to make (instead of getting slow permission from a JAG in an office in St. Louis who might be in the bathroom). It's the growth of autonomous stock trading algorithms. It will even be in advertising on porn sites.