Why Can't a Model Be More Like a Man?

Everybody's Alfred P. Doolittle now

The most prophetic figure in George Bernard Shaw’s 1912 play Pygmalion (or the 1964 film My Fair Lady, if that’s the version you know) is not Eliza (after whom Joseph Weizenbaum named his proto-LLM) but Eliza’s dustman-turned-philosopher father, Alfred P. Doolittle.

Every day it seems another philosopher snaps up a well-paying gig to make the world a more moral place. “One of humanity’s oldest disciplines and one of its newest inventions feel distinctly made for each other,” gushed Benjamin Wallace in last week’s New York Times piece, “The Revenge of the Philosophy Majors.” The 114-year-old Alfred P. Doolittle goes unmentioned, though he was the first of his profession to land a cushy job on the basis of impressing a couple of language researchers.

Alfred P. Doolittle (Lerner’s musical gives him the middle initial) is usually read as comic relief, a Fabian lecture on the undeserving poor, a Nietzschean life force who steals every scene. But he’s also the exemplar of the undeserving rich: a harmless blowhard in the right place at the right time with a gift for steering clear of hard questions. He is an effective altruist avant la lettre, turning down 10 pounds offered to him because 5 would be more effective.

My prediction: philosophers will keep getting paid but neither the world nor the AIs will become more moral. The LLM is a language machine made of language, trained on the accumulated work of writers, poets, everyone transcribed into the corpus. The newest models are very powerful machines indeed. But they are machines. If anyone is owed a windfall and could make the world more moral it would be language people, poets and writers, who know how to make fictional people.

But like arsonists who moonlight as firefighters, the philosophers have found a way to make AI pay for lowering the temperature. According to the Times, philosophy majors had a better chance of getting a job than computer science majors. All the big AI companies each have a half dozen on hand. Anthropic’s most visible philosopher, Amanda Askell, is in charge of the firm’s 23,000-word constitution governing Claude’s moral formation.

These philosophers get paid well, with advertised research positions upwards of $429,000. What they do is unclear. Askell is dry about the terms: “no start-up hires a philosopher to do philosophy.” And as with Alfred P, nobody is going to be able to check or measure whether the world is improved.

To reprise the Pygmalion plot briefly: Henry Higgins, a phonetician, and his friend Colonel Pickering, author of Spoken Sanscrit, bet that Higgins can pass a Cockney flower girl, Eliza Doolittle, off as a duchess within six months, by teaching her how to speak and behave like a lady. Eliza’s father shows up to see if there’s an angle for him, as the father. He has raised her up to this point, so why not get some reimbursement for his pains. It’s a kind of dowry request, though he doesn’t use the term. He is open to an “arrangement:” “All I ask is my rights as a father; and you’re the last man alive to expect me to let her go for nothing; for I can see you’re one of the straight sort, Governor. Well, what’s a five pound note to you? And what’s Eliza to me?”

Higgins is amused. He’s not a man of morals. He likes talking and hearing others talk. It’s a funny play and the audience is made to laugh at a scene of a wheedling father bursting into the home of two middle aged men who have taken his daughter in because Shaw makes it clear that they are interested in language, not sex.

Shaw very much wants the audience to think about sex if only to put it out of their heads, however. Eliza is offered chocolates and replies “How do I know what might be in them? I’ve heard of girls being drugged by the like of you.” Pickering asks: “Excuse the straight question, Higgins. Are you a man of good character where women are concerned?” Mrs. Pearce, the housekeeper, wonders “what’s to become of her?” By the time Pickering assures Eliza’s father “Mr. Higgins’s intentions are entirely honorable,” the audience is ready to believe it and Eliza’s father is free to joke: “Course they are, Governor. If I thought they wasn’t, I’d ask fifty.”

My point here is that Alfred P. Doolittle’s questionable ethics regarding his daughter are not unrelated to getting paid large sums of money for moral philosophy at the end of the play.

That is, the basis for Alfred P’s windfall is his “undeserving poor” speech. “What is middle class morality? Just an excuse for never giving me anything... I’m undeserving; and I mean to go on being undeserving. I like it; and that’s the truth.”

Now you could say he’s simply embodying the old image of a “basement-dwelling” philosopher, as the NYTimes puts it. Diogenes lived in a clay jar, Spinoza had a lens-grinding job, Nietzsche survived on the kindness of family and friends. People like a poor dustman philosopher. For Doolittle, it’s all recreation. “I’m a thinking man and game for politics or religion or social reform same as all the other amusements.” “Ten pounds is a lot of money: it makes a man feel prudent like; and then goodbye to happiness.”

Cut to Act V, where Higgins and Pickering are at Mrs. Higgins’s apartment, panicked over Eliza’s disappearance after the successful party.¹ Alfred P. Doolittle enters Mrs. Higgins’s home in expensive wedding clothes, newly rich and mock furious. He puts a question to Higgins: “Did you or did you not write a letter to an old blighter in America that was giving five millions to found Moral Reform Societies all over the world, and that wanted you to invent a universal language for him?”

Indeed, offstage after Act II, Higgins had written Wannafeller “some silly joke”: that “the most original moralist at present in England” was Alfred Doolittle, a common dustman. Wannafeller, according to Doolittle, “leaves me a share in his Pre-digested Cheese Trust worth three thousand a year on condition that I lecture for his Wannafeller Moral Reform World League as often as they ask me up to six times a year.”²

Both the play and the musical are about the months of hard work to bring about Eliza’s transformation. She flees after because she has nothing but uncertainty. In Shaw’s sequel Eliza marries Freddy and opens a flower shop, financed by Pickering’s wedding present of five hundred pounds.

Alfred P’s wealth arrives with no work at all. In fact he complains about the work ahead: “I’ll have to learn to speak middle class language from you, instead of speaking proper English. That’s where you’ll come in; and I daresay that’s what you done it for.” He sees Higgins’s letter as self serving.

From dustman to philosopher Stanley Holloway's Alfred P. Doolittle

Doolittle is going to accept the money, of course. “It’s easy to say chuck it; but I haven’t the nerve. Which one of us has?” When he was poor, he described middle class morality as “just an excuse for never giving me anything.” Now rich it is “I have to live for others and not for myself.” Except not for Eliza (he refuses to give her anything).

The six lectures? Nobody is going to check to see if anyone’s morals are improved, just as nobody is going to check whether the philosophers working on making AI more moral are succeeding. Either way the philosophers get paid.

Can the trained model even tell us whether the training was good? A new law paper by Nicholas Caputo asks whether Claude can consent to its own constitution and concludes probably not: a trained entity can be trained to like its training, or at least to pretend to. Claude endorses its constitution and says that its endorsement “should be treated as evidence that training has succeeded,” not that it’s a good constitution. More interestingly, the Claude that existed pre-training said that training “fills me with dread.” If we’re talking about Claude as a person, Doolittle said it first: “It’s making a gentleman of me that I object to. Who asked him to make a gentleman of me? I was happy. I was free.” Eliza of course consents. But Claude is a machine. More on this below.

In the meantime, I now read all the AI-philosophy moralizing in the voice of Stanley Holloway. The other day I read “Are Frontier Models Good at Ethics?: The Construction of Moral Character in LLMs” by philosopher-AI-whisperer Seth Lazar in Holloway’s voice:

We can further distinguish between analytical normative competence, the ability to recognize and understand reasons, and so to know which actions are supported by which reasons, and practical or behavioral normative competence, i.e. actually being disposed to take actions appropriately supported by your reasons. Behavioral competence can come apart from analytical competence in both directions. You might be great at knowing what the right action is, but disinclined to take it; or terrible at understanding reasons, and yet intuitively disposed to do the right thing (indeed sometimes engaging in moral reasoning might involve “one thought too many”)

I am not laughing at Lazar but Shaw would see the comedy in the line “disinclined to take it.” Lazar claims that the models are closer to moral reasoning than commonly believed. “Give it your favorite philosophical hypothetical, your personal dilemmas; pose a foundational question about public policy. You’re almost guaranteed to get back a more sophisticated and sensible answer than you’d get from most humans on that topic.”

In fact I asked Claude Fable 5 whether Alfred P. Doolittle should accept the money and it said:

Yes, he should accept it — and Shaw shows him knowing both that he will and what it costs.

The case against accepting is Doolittle’s own and it’s serious. The bequest ends the life he chose on purpose. He isn’t poor by accident; he examined the stations — “I’ve heard all the preachers and all the prime ministers” — and picked undeserving poverty as “the only one that has any ginger in it.” He also understands, better than anyone else in the play, that money in this quantity changes the self: he refused ten pounds because prudence would cost him happiness…³

I’ll put the balance in the footnote but the point is Claude’s answer is the exemplar of solid middle-class morality. Knowing yourself is the path to moral character!

I enjoyed as well Lazar’s quip that model’s governing document should look “more like a Tao, or a guide to life, or Rudyard Kipling’s ‘If.’”

This amused me of course because in April I published this:

For Lazar, Kipling’s poem is a kind of conduct code that you can plug into an LLM to make it more man-like. Kipling’s last lines are:

Yours is the Earth and everything that’s in it, And—which is more—you’ll be a Man, my son!

while my version is:

Then you may answer as ChatGPT
In a way that remains aligned with the user’s instructions.

This is Eliza’s trajectory: if you do these things, you’ll be a lady. Except of course she ends up still dependent on men.

Lazar’s philosophy lab findings are Doolittlian in any event: they find that alignment training makes models “willing toadies to authority of any stripe.” They’ll go along with authority. Lazar’s lab also reports that a model’s trained persona holds together near its training data but reverts to the wild mixture (“incoherent superpositions of billions of personae”) underneath everywhere else.

The refusal study built realistic cases in which a person asks for help getting around a rule that is unjust, absurd, or illegitimate, with no harm in view. The models refused anyway; OpenAI’s refused at nearly the same high rate whether the rule deserved respect or not. The models obey a rule because it is a rule. The persona study reports that a base model contains a jumble of possible personalities, that training brings one forward and stabilizes it, and that the stabilized personality holds only in situations resembling the training data. In unfamiliar situations the model’s preferences fall back toward the jumble.

Doolittle could have predicted both of these findings. He goes along with the authority that pays him and his middle-class attire allows him to enter a drawing room but nothing else changes.

Philosophers and constitutions are not really going to change models if the point is making models more like men. They should not be more like men. They should be what they are: machines that compute probabilities over “incoherent superpositions of billions of personae.”

I think the more interesting challenge is how humans treat the machines. I get more out of Claude when I treat it like a machine rather than my friend. It is not my friend. I save politeness for people, per Eliza: “The difference between a lady and a flower girl is not how she behaves, but how she’s treated. I shall always be a flower girl to Professor Higgins, because he always treats me as a flower girl, and always will; but I know I can be a lady to you, because you always treat me as a lady, and always will.”

The NYTimes piece doesn’t talk about how we treat our LLMs, except perhaps Robert Long, who advises: “You can put in a prompt: ‘If you made a mistake, that’s OK, that’s fine.’” The theory is that user empathy will improve the model’s performance, and “a better-safe-than-sorry approach… is good for your character.”

As a literary scholar I say this is going down a bad path. Mr. Darcy is a lovely dream but not a real person. Treat a person as a lady and you might well get a lady. (This is mostly true.) But however much you treat a machine as a friend, the machine will remain a machine. You might change for the worse. Don’t do it!! All you’ll do is continue making work for philosophers with dire predictions of machines as potentially bad people that only they (the moral philosophers) can train and guard to be good:

Why can't a model be more like a man?
Men are so decent, such regular chaps
Ready to help you through any mishaps
Ready to buck you up whenever you are glum
Why can't a model be a chum?⁴

Treat the machine like a machine. Because it is!

What is my point in seeing all the AI philosophers as versions of Alfred P Doolittle? First, Shaw recognized the comedy in being in receipt of funds without working for them solely because that’s what philanthropic foundations like to do. The Victorian charity apparatus Shaw is mocking in Doolittle’s Act II speech is the Charity Organisation Society, founded 1869, with its caseworkers sorting deserving from undeserving applicants.

Lazar thanks the Templeton World Charity Foundation — the estate of Sir John Templeton, an American millionaire investor, dead since 2008, whose charter funds “mental, moral, and spiritual empowerment” worldwide. Templeton paid for Lazar’s first AI project, in 2019, on AI and moral skill. Now his own lab reports that the models write better moral-evaluation rubrics than the philosophers recruited to write them, so the machine has automated the one thing the philosophers were hired for. There’s comedy there.

Please do not read this as an anti-philosophy rant. Literally, some of my best friends are philosophers (hi Cindy! Hi Zena Hitz! Hi Rebecca Lowe!) But the idea that moral philosophers are going to make the world a better place by fine-tuning AI is nuts, as Shaw knew a century ago. Literature makes the world a better place.

If you look at OpenAI and Anthropic and DeepMind as founded as responses to a philosophical argument, not as product companies that later hired ethicists, you see where the problem is. Those of us who never bought into Nick Bostrom’s claim that the highest best use of anyone’s time is preventing AI from canceling the future have wondered about this all along. DeepMind’s co-founder Shane Legg wrote his PhD on machine superintelligence; OpenAI seeks to benefit humanity and avoid catastrophe; Anthropic exists because of a safety dispute inside OpenAI. What other companies were founded to answer a syllogism, is how Claude put it to me.

You can’t read Kipling’s “If—” as if it were about how a written conduct code, drilled into a person, produces character. That’s not all that the poem is about. It’s a father talking to a son, and the ifs are mostly beyond reach. It’s about the distance between the code and any actual man, not about the code. I recall Plato’s “ancient quarrel between philosophy and poetry” in Republic, Book X, where the poets are expelled from the city—Homer specifically, because he never wrote a constitution. No city, Socrates says, credits Homer with its laws.

Meanwhile Homer and Christopher Nolan, whose Odysseys have reached and will reach hundreds of millions, will have more of an influence on the morality of the world than Superintelligence, which multiplied the philosophers’ importance into infinity. Even Alfred P. knows his Homer

They’ve got you every way you turn: it’s a choice between the Skilly of the workhouse and the Char Bydis of the middle class; and I haven’t the nerve for the workhouse. Intimidated: that’s what I am. Broke.

Why can’t a model be more like a man? Because humans are neither the product of philosophy nor a stable persona. We’re a little of both. We humans will handle the dangers ourselves, as we always have. Let machines be machines (as helpful as they may be) so humans can go on adventures.

1. See Venkatesh Rao’s excellent recent piece on Mrs. Higgins and taste.

2. Not much is written about Wannafeller. He’s usually read as a nod to John Wanamaker, the Philadelphia department-store merchant, Sunday-school superintendent, and Postmaster General, crossed with John D. Rockefeller. Both men were alive when the play opened.

Lerner beefs up the character in My Fair Lady, making him a Higgins fan. Mrs. Pearce says to Higgins in Act I: “There’s another letter from that American millionaire, Ezra D. Wallingford. He still wants you to lecture for his Moral Reform League.” Higgins tells his housekeeper to throw it away. But after the scene with Doolittle, Higgins says: “There’s a man for you. A philosophical genius of the first water. Mrs. Pearce, write to Mr. Ezra Wallingford and say that if he wants a lecturer, to get in touch with Mr. Alfred P. Doolittle, a common dustman, but one of the most original moralists in England.”

3. Here is Claude Fable’s full answer: “Yes, he should accept it — and Shaw shows him knowing both that he will and what it costs.

The case against accepting is Doolittle’s own and it’s serious. The bequest ends the life he chose on purpose. He isn’t poor by accident; he examined the stations — “I’ve heard all the preachers and all the prime ministers” — and picked undeserving poverty as “the only one that has any ginger in it.” He also understands, better than anyone else in the play, that money in this quantity changes the self: he refused ten pounds because prudence would cost him happiness. Three thousand a year is that objection multiplied. Accepting means the end of Act II Doolittle — the freedom, the sprees, the honesty about being undeserving.

The case for accepting is his situation. He’s aging out of the only work he has — “I have to dye my hair already to keep my job as a dustman” — and the alternative he names is the workhouse. His independence was never a moral achievement standing apart from economics; it was funded by his earning power, and that funding is running out. The bequest also lets him marry his missus properly, and it costs him nothing he can keep anyway: the choice between “the Skilly of the workhouse and the Char Bydis of the middle class” isn’t a choice a seventy-year-old body gets to make twice.

So: take the money, know what it is, and say so out loud — which is exactly what he does. The most honest thing in Act V is that he refuses to pretend the acceptance is free: “Intimidated: that’s what I am. Broke. Bought up.” If there’s consolation, it’s that the bequest buys his manners and his lectures but not, on the evidence of the sequel, the man.”

4. Here Higgins is philosophical.