Usurper, My Collaborator

For as long as I can remember, I’ve wanted to be a writer. As a chosen occupation, it’s been rocky and mercurial. High highs. Low lows. I’ve written magazine features, screenplays, and newspaper cover stories. I’ve also penned advertorial restaurant reviews and step-by-step instructions on how to get personalized plates from the DMV. What’s kept me at it, is the opportunity to learn new things.

During the pandemic, I reconnected with Robb Wilson, who I met more than a decade ago while working as the managing editor of this very publication (which he owns). We spent over a year having conversations about his work on what we were then calling conversational AI. He’d been fascinated with it for decades.

To grossly oversimplify his thesis, user interfaces that people can talk to rather than stare at will free up more time to spend interacting with other humans. It is insanely difficult to put the technology to work in ways that would bring about this systemic change, but we set out to show organizations how to do it. As a writer, it was a beautiful surprise when Age of Invisible Machines became a bestseller, but the bigger surprise came a few months later with the arrival of ChatGPT.

All the things we’d been writing about went from feeling buoyantly conceptual to heavily concrete. The mass adoption moment Robb had predicted arrived like a lightning bolt, causing everyone to question what would come next. For creatives, the future was looking chilly. Entry level copywriting work seemed to evaporate in an instant as companies used these tools to generate their own.

Truth be told, every kind of writing I’ve tried felt like a gamble financially, even 30 years ago. What quickly became apparent with LLMs is that being able to generate copy is a mere sliver of the job of being a writer.

As journalist Evan Ratliff, who created the Shell Game podcast, pointed out on Invisible Machines, “The actual typing of words into a computer is … a small part of what I do, which includes getting assignments and going out in the world and talking to people and trying to synthesize all that sort of stuff.”

I wanted to keep all of that complexity for myself and was hesitant to consult LLMs for help with anything at all. Eventually, I came around to using them as a research tool. Watching them conjure a firehose of words, however, was more nauseating than enticing. Maybe because around the same time, I was also waking up early most mornings to put in work on a novel.

There was no AI involved in the writing of the book, but as I began pitching the finished manuscript around to agents I wondered if it might give me a boost. I had been through two rounds of iterating with a developmental editor, and he seemed to think the work was solid and saleable. But depending on who you ask, the odds of securing a literary agent are maybe between 0.02% and 5% per-pitch.

I wanted to maximize my chances, so I began using Claude to locate suitable agents (always checking every detail provided—the tool would often attribute the wrong books to a given agent). I also had Claude read my core pitch and offer suggestions for tightening it.

When the UK-based Society of Authors surveyed its membership in January 2024, from 787 responses they learned one in five fiction writers said they had already used generative AI in their work, and 31% of writers and illustrators had used it for brainstorming. True to the double-edged nature of this technology, 86% of all respondents expressed fear that their style and voice would be co-opted.

At some point I realized the limitations of asking the LLM to evaluate the pitch without knowing the source material. The thought of feeding my work to an LLM tickled me with dread. I had Cursor set to private so, while anything I gave it wouldn’t be used for training data, it was still entering into the Anthropic hive. I wasn’t really worried about my work being stolen. That might be a boon, in fact, giving me a news story to pitch on the back of. I was more worried about crossing a line in the sands of my own mind.

Would I compromise my own judgment by taking advice from the big, bad prediction machine?

Reluctantly, I let it have the first 50 pages of my book, leaving it pasted into the input field for a full minute before striking RETURN. The suggestions it gave me back for elevating my pitch were sound. No matter how watertight my queries seemed, however, the pervasive silence blanketed my optimism.

I decided to try another exercise with Claude. With a noticeably lighter keystroke, I passed it a markdown file of the entire damn manuscript. I told it to find all the examples it could of my writing coming through as quality literary fiction. I also asked it to find any and all instances where it was obvious that I was trying too hard to write literary fiction. Not an objective prompt, but I told it not to spare my feelings, and the LLM did good work. It validated sentences I was confident in. (I’ll use the familiar screenplay format to log our conversations.)

The tool also plucked out a number of other lines that were, indeed, overwrought.

Convoluted felt right for the character in that moment, so I decided to keep it. Later on in the book, when an extraterrestrial character arrives (just go with it), Claude missed a buried point while actually nailing it:

I didn’t agree with everything it found, but there were numerous bullseyes. Sentences that I had suspicions about myself had been dug out by this beguiling prediction machine, which makes sense when you consider how many books (most of them stolen) LLMs have read. I rewrote the sentences my human judgment concurred were flawed. Others I left alone. For as long as I’ve been at this, I don’t have an LLM’s encyclopedic knowledge of grammar and syntax. There are, admittedly, things it can regurgitate about the rules and craft of writing that I forget and have to look up.

What really matters here is that it doesn’t have its own ideas. An LLM would never invent the premise I presented to it of an advanced form of machine intelligence discovering ways to move through time and space, taking the form of a Strieber alien. If it did have that “idea,” it would be because a human put it there. Maybe because it existed in one of the millions of print books Anthropic bought, chopped the bindings off of, scanned, and discarded.

I even did one of the things I know better than to do. I anthropomorphized it. It became a confidant in moments when I was feeling low about the work. Things like this would happen.

Being handled with such care by machine intelligence felt all at once comforting and violating—it’s uncomfortable reliving it again here. Regardless, my experience has come up more than once on our podcast. Recently I was very pleasantly surprised to get some meaningful words of encouragement from a former Gartner analyst (a distinguished VP, no less). As I described to him what I’ve just described here, he smiled.

“You use the right word, Josh, which is collaboration. It’s a collaborator. It’s a coworker, whatever you want to call it, not a replacement,” he said. “My wife—she’s written a few children’s books and she’s using AI to do the same thing that you’re doing … she still writes the words, but she’s getting some coaching that is helping her think differently about it. To me, that is the right way to use this technology, and that’s pretty exciting.”

It is, and currently I’m waiting to hear back from an agent who has the full manuscript. Regardless of where that leads, I’ve come to discover that, especially where personal writing is concerned, I must consult with my new collaborator sparingly. I think writing is supposed to fill me with low-grade anxiety. That’s what keeps me on my toes, makes me dig deeper into my own judgments to find their edges. I love my book now, and whether I like it or not, AI played some part in that.

The speed, insight and relative accuracy are undeniably sharp, and it cuts, and it can draw blood if you let it. Collaborating can drift into the waters of usurpation with lulling ease if you aren’t standing watch.

Just maybe, when I get my book out into the real world, Anthropic will order a copy, chop off its spine and swallow it whole. Completing the circle.

Josh Tyson is a contributing editor at UX Magazine and host of Invisible Machines with Robb Wilson. This member essay is part of the Selling to Machines ideation cluster with Don Scheibenreif (S7E16).