During the WGA strike of 2023, I was shitposting saying the WGA should have their own AI, because studios were threatening to use AI to eliminate writer jobs.
Though I was a newbie with Large Language Model (LLM) chatbots, and I’d only read cursory descriptions about how they worked, it was pretty obvious that they were “writing machines”. It wasn’t a “general intelligence” as people called it, but a software that would produce texts.
Having played with ChatGPT, getting it to write small stories, and write computer code, I was blown away. Not only because it was producing texts, but that it was doing it without computer code that, specifically, produced these types of texts. (I may have been wrong about this.)
Rather, it was, as Chomsky said, a kind of statistical technique to generate a text. Given a prompt, and chat history, the machine will calculate the next word to write. The calculation is based on the probability of each possible word following the current words. The probabilities of each word are established through training data: by how often that word follows the current word, in the training data.
LLMs are trained on existing texts, across the internet. I suspect some high-quality texts are included, and given a higher value. So, something like a published book would be considered more influential than a web page.
The texts must be “labeled”, or “tagged”, so features in the text are identified. For example, if you ask ChatGPT to write something in the style of HAL9000 (the computer in 2001: A Space Odyssey, a movie from the 1960s that was popular with computer hobbyists) it will do exactly that. Odds are, someone (or some program) labeled some texts as HAL9000, and when you ask for something “in the style of HAL9000”, it’ll be influenced to pick texts associated with “HAL9000”.
I realize this sounds like a lot of work, but there’s a website that has HAL9000 dialogue, and there are numerous pages with quotes from the HAL9000. So the website author has “labeled” its data, using things like titles, attributions, and even quotation formatting. So the work has been done, and it’s available for free.
Read this example of text that’s like the HAL9000, and explains the value of training data. (Please read it. It’s impressive.)
Isn’t that amazing? I thought it did a better job of explaining the issues than I did.
The Writers Should Own their Model
So, we’ve established that an LLM writes, and can write in the style of any person labeled within the training data.
How does a writer get hired?
They get hired to produce original work in their style, or to rewrite or fix other peoples’ work, because they can meld their own text with the existing text.
They’re obviously at risk of losing work to LLMs, particularly the more routine work.
Imagine if the studios make a training data set based on scripts they own, labeled by writer. They could reproduce the writer’s style in an instant.
One way around this is to deny the studios the right to use the scripts as training data, and prevent the studios from labeling the texts in ways that make an individual writer’s style available to LLM users.
Yet, the labor-cost-savings of LLMs are huge. There’s no way we’re not going to use them.
So, I think the WGA should invest in training a model, with WGA members’ work, and labeled with members’ names. WGA members should use this model to produce scripts, dialogue, and so forth, but with restrictions, like disallowing writers to copy each other’s style.
In contract negotiations, the WGA should require the studios to use the WGA model, through the WGA’s web services.
All Writers and Intellectual Laborers Should Form Unions
LLMs represent a new form of capital.
Capital, in the Marxian definition, is money, land, and the “means of production”. In the industrial age, the means of production were factories and machines that produced goods. Today, the means of production also include personal computers and other equipment that help to produce both goods and services.
Old, silver-haired people will remember when word processors made the typewriter obsolete. (Typewriters were 19th-century technology that were the primary writers’ tool until the 1980s.)
Starting in the 1980s, the word processor was a new kind of capital, a piece of software that turned the PC into a writer’s machine. By the late 1990s, the word processor could correct both spelling and grammar.
LLMs in the 2020s represent the next step: a piece of software that will write for you (and put a writer out of a job).
Why LLMs are a New Form
The earliest word processors emulated a typewriter, down to some of the graphics and terminology. In one decade, though, they far exceeded the typewriter, by correcting spelling, then grammar. Not only that, but you could augment the spelling dictionary so it wouldn’t mistakenly flag the words you use.
When you added to your dictionary, you were creating a new form of capital. It would help the word processor work better, in the future.
When you made templates for common letters, you were creating a new form of capital. They would help you work faster.
Today, word processors are adding AI tools. (MS 365 Word’s tool sales pitch.) They can help you writer better, faster.
Imagine if every single thing you wrote were treated as a potential template, and used to train an LLM. Everything you wrote would be contributing to this accumulated, new form of capital, to create a personalized LLM.
You could start a letter and complete it in 5 to 10 minutes, with the help of AI.
Now, imagine that your personalized LLM were given out to 10,000 people. They wouldn’t need you to write letters anymore. You might be out of work.
If they each paid you $10, maybe you wouldn’t mind… $100,000 is still good money… but they won’t do that.
This is why intellectual laborers (writers) should form unions: to control ownership of the model, to create better models by working together to produce the best models possible for your purposes, and to use the model to have leverage against the employers.
At-Risk Jobs: Programmers, Lawyers, Customer Service, Educators
I went looking for articles that explained what jobs were at risk of being lost or significantly altered by LLMs. They were forward-looking, and seemed like science fiction to me. So I thought about what LLMs can do, and what they are lousy at.
In my experience, they are great at writing.
They are not good at reasoning, unless some sources of explanatory text exist.1
They are particularly good at writing computer code. This is because computer code is extremely “regular” or uniform. Programs are often “boilerplate”, where you use the same template over and over, and fill it with your small changes.
The parts that aren’t boilerplate are the program logic – typically, the code implements the logic, and an English language comment explains it. So, it’s possible to create a mapping of code to explanation, and vice versa. (This is the reasoning part: it can “reason”, because it can translate from a symbolic logic, to precise, plain English.) The explanatory comments are then mixed in with other training data from other logical texts within the model, like philosophy, law, textbooks, and so forth.
So, I tried to recall what articles used to say about jobs at risk: law, customer service, education, programming.
They all involve reading and understanding a lot of text, and then producing new texts.
| Field | Text Inputs | Text Outputs |
| Computer Programming | Manuals, references, code. | Code, documentation. |
| Law | Law books, laws, cases. | Contracts, arguments, laws. |
| Customer Service Chatbots | Technical info, knowledge bases, troubleshooting scripts, flowcharts. | Chat text. |
| Educators | Textbooks, tests, essays, lesson plans. | Lesson plans, paper grading, textbooks. |
| Writers, text and screen. | Books, scripts, research material. | Books, scripts. |
Secretaries were often mentioned as a job at risk, but I think they really aren’t at risk, because they’ve already automated the hell out of writing letters and emails, and most of their day isn’t about writing emails. They do phone calls, organize meetings, buy things (a task that takes a lot of time), and otherwise manage the boundary between the public and their boss. Many secretaries work for multiple bosses, which increases their job security.
Understanding the Terrain: Surveillance Capitalism and Technofeudalism
The ideas in this essay are not mine alone, they’ve been influenced by two texts, Surveillance Capitalism by Shoshana Zuboff,2 and Technofeudalism by Yanis Varoufakis.3 Since they were long books, I listened to them on audiobook, and missed a lot. So, ChatGPT can help me:
ChatGPT says: Shoshana Zuboff’s concept of Surveillance Capitalism refers to the economic system where companies collect, analyze, and profit from personal data, often without users’ full knowledge or consent. These companies monitor individuals’ behaviors, preferences, and interactions, using this data to predict and influence future actions. Zuboff argues that this practice commodifies personal information, creating a new form of power that manipulates individuals for commercial gain. Surveillance capitalism undermines privacy, autonomy, and democracy, as corporations gain unprecedented control over individuals by turning their private lives into valuable assets.
ChatGPT says: Yanis Varoufakis’ concept of Technofeudalism describes a new form of economic and social system where digital monopolies dominate and control much of the global economy. In this model, tech giants like Google, Amazon, and Facebook operate as modern feudal lords, controlling vast amounts of data and resources. Rather than offering traditional market competition, these companies exploit users by extracting personal data and using it to shape behavior and consumption. Technofeudalism sees the decline of traditional capitalism, as power shifts from workers and entrepreneurs to tech oligarchs, creating new inequalities and a lack of democratic control.
ChatGPT missed the role of Machine Learning (aka Artificial Intelligence) in Zuboff’s argument, which is that the surveillance produces a “behavioral surplus” of clicks and mouse movements and other sensor data, that is fed into software that turns the data into a kind of capital that trains Machine Learning systems.
ChatGPT missed Varoufakis’ idea about platform feudalism, where platforms or “marketplaces” like Amazon, Walmart, Ebay, Uber, AirBNB, Lyft, Turo, VRBO, Etsy, Depop, The Real Real, Thredup, OfferUp, Facebook Marketplace, and others, operate like feudal landlords, or sharecropping landlords, taking a portion of each sale.
To get an idea of the contemporary data collection landscape, here’s a table listing a lot of websites we probably know or use.
| Companies | Service (what you get) | Data Likely Collected (their new capital) | Approximate Transaction Fees (the taxes paid to the technofeudal lord) |
| Amazon, Walmart, Ebay | Retail marketplace | Listing and sales data. | 15% to nearly 100% on Amazon |
| Uber, Lyft | Rides | Travel routes and times. | 30% |
| Ebay, OfferUp, Etsy, The Real Real, Thredup, Depop, FB Marketplace | Retail marketplace | Listing and sales data, used products. | 15% |
| AirBNB, VRBO | Lodoging listings | Stays in rooms, listings. | 14% |
| Turo | Car rental listings | Listings, rentals, trips. | 23% |
| Gmail, Google Drive | Email, Office Software | Email and document content | 0% |
| Toggl | Time tracking | Your work habits | 0% |
| Github | Software development tools | Your code | 0% (pay for Copilot to use their model, based on your code) |
| Tinder, Hinge, Bumble | Matchmaking | Your love life | 0% (but you are pressured to pay) |
These techofeudalists combine the rent seeking of Technofeudalists, with the capital production of Surveillance Capitalism, to create new form of capital. These models can then advise the user how to improve sales. They can also manipulate the user to work harder, or work better, or to buy something.
The new landscape of chatbots extends both surveillance capitalism and technofeudalism by capturing the behavioral data, and the content data, transforming it into training data sets, potentially producing new models.
In Marxian terms, technofeudalism “socializes” the production of the training data, and socializes the production of new data models. By “socializes”, I mean that the production of training data and models is no longer an individual pursuit, but performed by groups of people. Socialization also implies that the work cannot be performed individually, for the scale required. The capital is collectively produced.
“Socialization” sounds like “socialism”, but in the current world, there is little “socialism” in the socialized production of LLMs. It’s largely privately owned by the tech oligarchs. Even if the code is “open source”, the training data is not. The training data is often not fully revealed, because the training data contains the human information and logic that is used to replace human labor.
The way to gain control of the training data is to form labor unions, which would then produce models that could be used to automate parts of their own labor.
Notes
- A side quest to get ChatGPT to explain the obesity epidemic and roast chicken prices demonstrates how bad ChatGPT is at making inferences or even referencing facts. The text reflects conventional wisdom, ignoring some lesser known facts. https://bsky.app/profile/johnkawakami.bsky.social/post/3li3ybdsci22b
However, LLMs do exhibit behaviors that show they do have “ideas” that exist beyond the text and can reason. I did a search for papers, and found a recent one that describes recent progress in Chain of Thought reasoning. The original Chain of Thought paper was written in 2022. Even in 2024, the LLMs solving math problem test data with an accuracy of 70% or less, most often solving fewer than 40% of the problems. The LLMs are a “C” student in math, and even worse at “common sense” reasoning. - Zuboff, Shoshana. The Age of Surveillance Capitalism book. Website.
- Varoufakis, Yannis. Technofeudalism book. Website. Varoufakis used to be an economist at Steam, which is a software sales platform.