My writing robot

A personal experiment with AI

Welcome to my personal newsletter. I’m publishing essays on digital technology and culture in the run-up to my January 2024 book FILTERWORLD: How Algorithms Flattened Culture. Subscribe or read the archive here.

---

Alert: Next Monday, I am going to send out a big update newsletter about Filterworld! Watch our for that and tell anyone who is interested to sign up.

Artificial intelligence is looming over everything right now. It’s less an abstract threat than an imminent reality. The Hollywood strikes, now both writers and actors, are motivated in part by the workers’ fears that they will be replaced by AI: the writers by generative text and the actors by generative videos, trained on all the movie material already out there and produced on demand. (One Hollywood contract proposal would have taken extras’ likenesses to be replicated in perpetuity by AI in exchange for just one day’s pay.) “We are all going to be in jeopardy of being replaced by machines,” Fran Drescher, the president of the Screen Actors Guild, said. Even novelists are having to negotiate contracts that ban publishers from using their work to train AI models.

But at the same time, people are also using AI, the way we use tools like photo filters or text-message stickers. It’s becoming part of the furniture of our digital lives. Friends have told me they turned to ChatGPT when it came time to write a toast or officiate a wedding. Students use it for homework and tourists to plan itineraries. AI chat bots are threatening to replace the long-established structure of Internet search, answering questions directly instead of driving traffic to websites that have published relevant articles. We’re in ambiguous territory, attracted to the novelty of the technology and the convenience it seems to deliver, but also scared of its consequences.

I wanted to better understand the state of these tools, what they can actually do right now. So I asked an AI startup called Writer to train a custom model on my body of writing, 150,000 words including dozens of articles and the manuscript of my first book. The bot was meant to be able to write in my style and copy my voice. My encounter with Robot Kyle is the subject of my latest New Yorker piece here:

>My A.I. Writing Robot

Spoiler: The robot was quite good at copying my style, but much less good at writing anything meaningful. It could predict which words I would be most likely to use but it couldn’t make an original argument for me. Using the robot made me think a lot about the nature of writing and what goes in to its peculiar form of labor. AI tools fundamentally promise to make things more convenient. They’re supposed to find or create information faster. But writing isn’t convenient. It’s not an instantaneous process — worthwhile writing by nature takes time and effort. There’s only so much you can speed it up. Here’s the pull-quote that many readers are grabbing from my article:

“The most unsettling aspect of A.I.-generated text is how it tries to divorce the act of writing from the effort of doing it, which is to say, from the processes of thought itself.”

While a painter might enjoy playing with the AI image generator Midjourney to try out ideas or visualize outlandish concepts, I found that Writer and other AI writing tools actually took the fun out of writing for me. I know most people hate it, but writing — stringing together words in a particular order, assembling sentences into paragraphs, alone at the keyboard — is my favorite part of my job. Writing isn’t just how I communicate ideas or discoveries; it’s how I come up with them in the first place. Writing an essay feels like assembling a vast puzzle that’s different each time. But rather than creative, my interactions with the AI felt combative. It turned me into a collagist or a frustrated editor instead of a writer as I decided which random texts it created were relevant and constantly reshaped prompts to get better results.

Working with an AI is a different form of mental labor than writing something on your own, and it is likely to be a less efficient one, at least until AI improves dramatically. Its efficacy depends, too, on what you want out of your writing — both for a writer and their reader. If what you want is a content-marketing blog post that serves more to incrementally improve Google search results than to be read by an actual human, AI might help. If you want to think deeply through a subject and synthesize knowledge, then communicate that knowledge to others, it’s going to be better if you just struggle through the painful process yourself. Most AI-generated writing is content that no one but the prompter wants to read. As it stands, the Kyle Robot has no place in my writing process.

---

Notes

Filterworld is about algorithmic feeds, a form of digital technology that has dominated the Internet over the course of the past decade. In some ways, with Twitter falling apart and the general ennui with social networking, it feels like Peak Algorithmic Feed might be over. We maxed out with TikTok. But AI is continuing all the problems with algorithms — where algorithmic recommendations automated the finding of content, AI automates the creation of content in the first place. Whatever the role of AI turns out to be, I think we’ll be able to draw a straight line from the rise of automated feeds to AI-generated content.

It’s hard to figure out how to end a book, particularly one that deals so much with the exact present. I finished writing Filterworld many months ago, and then finished editing as well. But I keep seeing it everywhere.