The Machine Gave Me Homework

What teaching AI my writing voice taught me about authorship.

Some time ago, I came across a LinkedIn post from a writer who said she could always recognize AI-generated writing and would never use it herself. She enjoyed the process of writing, finding the right words, building sentences, and working through an idea until she was satisfied with the result.

I understood her position. I enjoy writing, too. And I agree that much of what AI produces is recognizable. Predictable structures, unnecessary explanations, and polished sentences that manage to communicate very little.

But her post stayed with me. I started wondering what we're actually protecting when we insist on writing every word ourselves. The act of writing? Our ideas? Our individuality? And if someone else helps us construct the sentences, at what point do those words stop being ours?

When an Editor Changes the Writer

In 1980, Raymond Carver received extensive revisions to his manuscript Beginners from his editor, Gordon Lish. Lish cut more than half the material, changed endings, and rewrote passages. Carver was so troubled by the changes that he wrote Lish a letter asking him to stop publication.

The book eventually appeared in 1981 as What We Talk About When We Talk About Love and helped establish Carver's literary reputation. Lish's editing contributed to that success, but Carver didn't accept every change. He later published a longer version of one of the stories, The Bath, as A Small, Good Thing in his 1983 collection Cathedral.

Readers can compare the versions in the Library of America's edition of Carver's collected stories.

Some prefer Lish's tighter prose. Others prefer Carver's fuller treatment. An editor can improve a manuscript while removing something its author values. And what readers consider better isn't always what the writer wants.

Today, that editor can be a machine.

The Machine Gave Me Homework

I kept running into the same problem when using AI to help with my writing. I'd give it an idea, some research, perhaps a few rough paragraphs, and ask it to develop the material.

The result would be grammatically correct, logically organized, and considerably longer than anything I would have written.

I'd spend the next several minutes removing elaborate introductions, artificial transitions, rhetorical questions, and explanations that repeated what the previous sentence had already established. Sometimes an ordinary observation became a grand statement about the human condition.

After correcting the same problems over and over, I decided to see whether I could teach AI to recognize how I write.

I started a project called Pavel Humanizer. Apparently, I needed AI to teach AI how to sound less like AI. Eventually, I renamed it Authorial Layer, which sounded a little more respectable.

I approached it as I would any design problem. Before trying to improve something, I needed to understand its behavior. I wanted to compare how AI naturally wrote with how I wrote myself, identify the differences, and establish a way to evaluate the results.

That's when AI started giving me homework.

It asked me to complete writing assignments without using AI. Those samples would provide evidence of my vocabulary, sentence structure, rhythm, and recurring patterns.

I was building a writing assistant, and the assistant was telling me to sit down and write something myself.

I hadn't considered how difficult it might be to write naturally when I knew every sentence would be analyzed. Was I writing the way I normally would, or demonstrating how I thought I wrote? I couldn't be certain, but the samples gave me a starting point.

Once I had those samples, I moved to a different exercise. The model would produce a draft from a set of notes, and I would edit it, recording what I changed and why.

That process revealed more about my writing preferences than I expected.

When More Personalization Made Things Worse

I tested six writing exercises using an AI model without my personalized profile, then repeated them with the profile applied.

Without the profile, I made 32 corrections. With it, I made 44.

I'd spent time teaching AI how I write, and now I had more editing to do than before.

It was a small experiment using one model and synthetic writing scenarios, with corrections counted from my own edits. Some profile runs may also have been influenced by ChatGPT's memory, while the memory settings for the baseline weren't recorded. I couldn't isolate exactly what caused the difference, but I could examine the mistakes.

The personalized model reduced some patterns I disliked, particularly artificial contrasts. Unfortunately, it also became more talkative, adding explanations, narrative framing, and language I kept removing.

My preference for conversational, story-driven writing seemed to encourage the model to elaborate. It had taken a few recognizable characteristics of my style and exaggerated them.

One of those personalized exercises involved a fictional governance review. Fourteen company policies, nine without assigned owners, and three controls marked complete without supporting evidence. The assignment was to turn those notes into connected paragraphs without changing the facts.

The model produced about 220 words. I cut the passage to 113.

Here's one example.

AI wrote:

Once someone changed a cell to “done,” that answer effectively became the record.

I changed it to:

Once someone typed “done,” the spreadsheet treated the claim as fact.

Across the passage, I removed unnecessary framing, invented objections, and explanations that repeated what the evidence already showed.

Looking at this and other corrections, I noticed that I tend to name who does what. I prefer concrete actions over abstract descriptions, and I usually want to make my point without explaining it a second time.

But shortening everything wasn't the answer either. In a mentoring exercise, I actually added words to acknowledge the person's frustration and offer some reassurance. The model had been concise, but the response needed more empathy.

That distinction mattered. I like writing that flows naturally, and sometimes an idea needs a little more space. What works in a short professional email may feel rushed or incomplete in a personal essay.

I could recognize these differences almost immediately when reading something, but describing them to a machine required considerably more thought.

Learning From the Edits

The first version of Authorial Layer relied heavily on descriptions of my writing style. Conversational, direct, personal, reflective.

Those descriptions were too broad. The model could reproduce certain characteristics, but it didn't know when they were appropriate.

The corrections provided better evidence.

Every time I rejected a sentence, removed an unnecessary explanation, or changed the emphasis of a paragraph, I was making a decision that the system could examine.

I began recording those decisions, along with accepted writing examples and explanations of what I wanted to preserve. The system could then use that evidence to guide future drafts, rather than relying entirely on a general description of my style.

I also introduced checks to protect factual meaning. In one evaluation, I accidentally changed a claim while revising a sentence. A preservation check caught the change, and I corrected it. Apparently, the human needed supervision, too.

Later evaluations showed some progress. In a new set of six exercises the profile hadn't seen before, shorter pieces required between zero and two corrections. Longer articles and explanations still needed considerable editing. I reduced their lengths by 41% to 56%.

The results were encouraging in some areas, but I was still spending time correcting problems the system was supposed to prevent.

What interested me more was how the process changed my understanding of my own writing.

I could recognize a sentence that didn't sound like me almost immediately. Explaining why was harder. Sometimes I disliked the wording. Sometimes the model had changed the emphasis or introduced an interpretation I hadn't intended. Other times, it had simply used too many words.

There were also moments when a technically imperfect sentence communicated exactly what I wanted. Improving its grammar or structure didn't necessarily improve what I was trying to say.

I had been making these decisions for years, usually without examining them closely.

The clearest evidence of my voice was often in the edits I refused to accept.

Recognizing My Own Voice

I still think about the writer on LinkedIn who said she could recognize AI-generated writing.

Looking at my early experiments, I suspect she would have recognized plenty of it. I certainly did, although I had the advantage of knowing how I would have written those sentences myself.

As Authorial Layer improves, the obvious AI patterns become less frequent. The writing follows more of my preferences, and some tasks require fewer corrections.

But there's a question worth considering. If a machine can reproduce my writing patterns, does that mean it can reproduce my voice?

I don't think the patterns alone are enough.

A model can learn that I prefer direct language, concrete examples, and connected paragraphs. It can observe which expressions I remove and which sentences I leave untouched. With enough examples, it can produce something that resembles my writing.

But it doesn't necessarily know why I chose a particular expression or what experience led me to make an observation.

Those decisions come from somewhere. Years of work, conversations, mistakes, things I've read, places I've been, and ideas I've changed my mind about.

Even my writing habits aren't fixed. Sometimes I deliberately break them because the subject calls for something different.

A convincing imitation of those patterns can exist without any of the thinking that produced them.

For me, authorship involves developing the ideas, examining the evidence, deciding what matters, and taking responsibility for the finished work. AI can participate in that process. The extent of its involvement matters less to me than whether I'm still making meaningful decisions about what gets said.

And I don't think that requires typing every word personally.

What Makes the Words Ours?

There's one part of writing I don't want to lose.

Sometimes I begin with a clear idea and need help expressing it. Other times, I have an observation I don't fully understand. I need to work through the sentences, question my assumptions, and occasionally discover that what I originally believed was wrong.

When I write, I'm often figuring things out as I go. A sentence can take me somewhere I didn't expect, and an argument can fall apart halfway through a paragraph because I suddenly realize it doesn't make sense.

AI can make that process faster. It can also make it easier to accept a reasonable-sounding explanation before I've finished thinking about the problem myself.

I started building Authorial Layer because I wanted AI to recognize and preserve my writing voice. Instead, I found myself examining choices I'd been making for years without giving them much thought. The machine could identify patterns, but I still had to decide which ones mattered, which ones needed improvement, and which ones I wanted to keep.

Joan Didion wrote in her essay Why I Write:

“I write entirely to find out what I’m thinking, what I’m looking at, what I see and what it means.”