Who Owns the Machine That Writes Like You?

Jim VandeHei has been running an experiment on himself. The Axios co-founder, 55 and, as he puts it, "a writer by training," has been turning his own catalog into training material: every Axios column, memo and strategic document, plus four books, two of them unreleased. The result is a set of AI assistants that draft in his voice. One of them, an agent running overnight, scans publications and research while he sleeps and delivers writeups in his style by email before he wakes up.

His method is not exotic. It is three moves, and none of them require an engineering degree.

First, he writes rules into the model's memory, and the rules are about him, not the machine: "challenge me, never flatter me; press me with wise skepticism. Then write like me." Second, he makes the style mechanical and explicit. Short sentences. Fact-based emphasis. Supporting points stacked in order of importance. He tells the model to commit that style to memory and keeps a copy in a skills document for his agents. Third, and this is the step most people skip, he pours in examples. Not a sample. Everything. His advice for the rest of us is that it "can be as simple as diary entries, memos or a school project."

Does it work? He says the output is often "edited better than my favorite human editor," which is a remarkable sentence for a media founder to publish under his own byline. He also rates the overnight agent's drafts as "good, smart starters or enhancers," and "never camera-ready." Both things are true at once. The machine trained on Jim writes a very good Jim first draft. It does not finish.

He flags the risk himself: anyone lazily outsourcing their thinking is building what he calls a "blah brain." And the sharpest sentence in his column belongs to his wife, Autumn, who is more skeptical than he is: don't conflate conveying information with "soul writing that many of us need to live, breathe and understand the world around us."

The reason to bother

There is a second Axios piece that explains why training an AI on your own voice is worth the trouble, and it is not a productivity argument.

Left on their defaults, these models pull everyone toward the same middle. A University of Southern California study found writing style diversity dropped sharply after ChatGPT's release, across scientific journals, local news and social media. Researchers at the Max Planck Institute reviewed 740,249 hours of recorded content and found ChatGPT's pet words, "delve" among them, migrating into ordinary human speech. Morteza Dehghani, the professor who oversaw the USC study, describes the mechanism plainly: "even people who are not using it... start writing more like LLMs."

The linguist Emily Bender calls the destination the "LinkedIn average." The Poynter Institute's Alex Mahadevan now avoids em dashes in his own writing so nobody mistakes him for a chatbot, and admits: "I have been second-guessing myself, thinking, 'Well, sh*t, is someone going to think this was written with AI?'"

Read against that backdrop, VandeHei's project stops looking like a shortcut and starts looking like a defense. An untrained model averages you toward everyone. A model trained on your corpus averages toward you. The examples are the whole game, which is exactly why his third step is the one that matters.

The custody question

Here is the part the how-to columns skim past, and it is the part an accountant cannot.

VandeHei's voice model lives inside Claude projects and ChatGPT agents. So does the corpus that built it, including two unpublished books. That may be a fine trade for him; Axios publishes its own policy on AI use, and he owns the company. For the rest of us the question is worth asking in plain terms: the most complete machine-readable record of how you think, stored on someone else's computer, under terms you did not write. You are not the owner of that copy. You are its tenant.

There is a local route, and it stopped requiring a research lab in 2021. A Microsoft team published a technique called LoRA, low-rank adaptation: freeze the big model, train a small set of added weights on your data. The paper reports it cuts trainable parameters by a factor of 10,000 and GPU memory by two-thirds versus full fine-tuning. Translated out of paper-speak: adapting a capable open model to your writing became a job for one consumer graphics card and a weekend, not a data center.

The plumbing is now genuinely boring, which is the highest compliment infrastructure gets. Ollama's Modelfile format is a few lines: FROM names the open base model, ADAPTER points at the LoRA you trained on your own corpus, SYSTEM carries your rules, the same "challenge me, never flatter me" contract VandeHei writes into cloud memory. The whole thing runs on your desk. The diary entries never leave the house.

Honesty requires the trade-off stated at full volume: the local copy is dumber. A small open model with your adapter will not match the frontier models VandeHei is using, and his bar of "never camera-ready" applies twice over at home. What you buy with the lost polish is custody. Whether that is a good trade depends on what you are feeding it. For published columns, the cloud is probably fine; the copy of record is already public. For diaries, client memos and unreleased manuscripts, the question answers itself.

What I would actually do

Both routes start in the same place, and it has nothing to do with software: assemble the corpus. The work you are proud of, gathered in one folder, is the asset. It is also, not incidentally, a useful audit of whether you have a voice worth teaching. Then decide where the copy lives, the same way you would decide where money lives: convenience accounts and custody accounts are different products, and pretending otherwise is how people get surprised.

And keep Bender's warning taped somewhere visible. "There is value in the struggle of writing," she says, "because we learn to express ourselves, and we learn to do the thinking that happens as we're writing." The machine that writes like you was built from the struggles you already did. It cannot do the next one for you, and if you let it try, there will eventually be nothing new to feed it.

Sources: Jim VandeHei, "Writing with AI," Axios, June 18, 2026 · Josephine Walker, "AI is changing how we write and speak," Axios, May 2, 2026 · Hu et al., "LoRA: Low-Rank Adaptation of Large Language Models," arXiv:2106.09685 · Ollama Modelfile reference