The New Puritanism

It should just be about how you get there, not how you get there.

Anthropic is adding watermarking to Claude-generated content in response to the European Union’s AI Act. The idea is that AI-generated text should carry a machine-readable mark identifying it as such.

Ben Thompson, writing in Stratechery, recently objected to this:

“Implicit in the E.U.’s regulation is the idea that an AI is an independent entity that needs to be distinguished from humans; the alternative view — that I hold — is that AI is (at least for now) a tool that is wielded by humans.”

A tool, not an alternate entity. Somewhere in there, a human greenlit the content, or at least the process. This makes AI, or anything else in the pathway of content creation, a tool.

Let’s say we have an executive who has an idea for an article. She might explain the thesis to a communications employee, and provide some arguments and examples. When the draft of the article comes back, she reviews the resulting draft, makes some changes and then publishes it under her name.

This happens all the time. This is what communications departments have traditionally done. No one cares if the executive hand crafted each word: their name is on the final product, and it is attributed to (or blamed on) them.

But now, if we replace the communications employee with an LLM a different standard emerges. Suddenly we’re deeply interested in exactly which words the executive wrote, and which were AI generated. We want disclosure. And we may even want the text itself to contain an invisible mark informing the world that a machine participated in its creation.

Why is this different. And honestly what business is it of ours how the content was crafted? The author put their name on it.

The Signature Is My Endorsement

In software development, a developer might submit a pull request containing 500 lines of code. It is possible that she wrote bit of it herself, or perhaps she copied and modified an example from Stack Overflow. Or, an AI agent might have wrote almost all of it.

Who cares. When the developer submits her code request, it has her name on it. The developer is saying: I understand this code. I believe it does what it is supposed to do. I endorse merging it. I take responsibility for the consequences.

If the code causes problems, then “the AI wrote it” isn’t a reasonable defense. The developer submitted the PR. (And a reviewer approved it. It’s also on them.)

Why is writing different? I might write an article from beginning to end. I might dictate it. I might ask an LLM to challenge my argument. I might have it produce a first draft, and I might rewrite that draft heavily. I might employ a human ghostwriter, communications department or editor to do exactly the same things.

When I publish the content I am saying: I endorse this. These are my claims. Hold me responsible for them.

You Must Do the Work

There is, however, a peculiar moral anxiety developing around this. There is an old moral substrate in our society that says effort equals morality. To simply arrive at a desirable outcome is not sufficient, it’s how you got there. You needed to sweat, or the act is immoral. You’ve taken a shortcut.

This is an extraordinary position for a technological civilization to take. Our long term history is largely built around rejecting it.

For example, when assembly language appeared, programmers worried that people would lose the ability to program directly in machine code. When higher-level languages appeared, programmers stopped needing to understand as much assembly. Gradually libraries replaced code that people once wrote themselves. Compilers increased automation and IDEs increased it more.

And yeah, skills disappeared. Try finding someone good at hand coding machine language these days. But so what?

Technological Progress Comes With Skill Loss

We accept the loss because the point of programming isn’t to demonstrate one’s ability to perform the maximum amount of labour. The point is to make computers do useful things.

And, if you stop to look for a minute you’ll see the same thing everywhere. Washing machines eliminate labour (but prevent us from getting really good at hand washing). Calculators eliminate labour (but limit our slide rule skills). Photography eliminates labour (wait, what about portraiture?). Search engines eliminate labour (spend much time in a card catalog lately?). Power tools eliminate labour (perhaps you don’t want to hand saw everything).

Often civilization advances in large part by discovering ways to obtain useful results with less human effort. Every time this happens, there is hand wringing over lost skills. Generative AI is producing this now: we have injected the moral value of suffering into writing.

Crouching Tiger, Hidden Classism

Yup, I’m going here. Rich and powerful people have always been able to outsource parts of intellectual production. Executives have communications departments, politicians have speechwriters, and professors have research assistants. These arrangements don’t generally provoke existential arguments about authorship. And yet, not everyone can afford a ghostwriter.

Generative AI has suddenly given millions of ordinary people access to some approximation of content production capabilities previously only available to people with money or institutional support. Now, someone with good ideas but weak writing skills can express them clearly. This also means that someone writing in a second language can produce fluent prose, and someone who simply doesn’t want, or can’t afford, to spend four hours polishing sentences can spend those four hours doing something else.

And yet the means of how the content is produced is the problem. I call this the New Puritanism.

Protocols and Pluralism

I’ve spent a lot of time thinking about protocol-based organizations. The basic idea is that an organization should make its important boundaries explicit: define responsibilities, interfaces, decision rights and constraints. But within a protocol boundary, leave people alone.

Again, to use a CS example, if I own an API and you own another API, we do need to agree about the interface between the two. But you don’t need to dictate how I implement my API. It is the interface that allows us to cooperate without mandating how it’s implemented.

This is also how pluralistic societies work. Let’s say we have a protocol that says “don’t steal my stuff”. (We would commonly call this a law.) The law does not generally require you to refrain from stealing my stuff because you share my philosophical theory of property. Maybe you believe private property is a natural right. Or maybe you want my TV, but are afraid of going to jail.

The law doesn’t care why. As long as you don’t steal the TV, we’re good. Why is your concern.

Purity Tests

The New Puritanism rejects this approach. It is not sufficient that you wrote a good article. It demands to know how you wrote it. Did you hand craft every sentence yourself? Did AI make a suggestion? Did you use a spell checker, or is your mastery of spelling such that you don’t need one? Perhaps it would be better if you used a typewriter instead of a computer.

It’s not about meeting the protocol (make a good article). It’s about the process of production, what happened behind the curtain. And now, everyone has to hold the same beliefs, because meeting the protocol is not good enough. And when everyone has to believe the same thing, pluralism dies.

One might argue bringing up “typewriter vs computer” is absurd, and they’re not wrong, but the point is that once we accept policing of the means of creating content, we have created a slippery slope. Human beings have spent enough of our history policing one another’s beliefs, motives, rituals and private behaviour that assuming we’ll politely stop at the first convenient boundary seems foolish. Over time, the “morally correct” way to write will expand in scope.