For anyone who writes regularly, the rise of generative AI creates an interesting dilemma. AI can produce text remarkably quickly. It can summarize information, suggest headlines, rewrite paragraphs and help turn a collection of ideas into something coherent. But can it actually learn to write like you? That is the question behind a fascinating experiment by Feather in his blog post Teaching AI My Writing Voice. Instead of simply asking AI to generate more content, Feather tried to teach it something much harder to define: his personal writing voice.

More than copying a style

At first sight, teaching AI how you write might sound relatively straightforward. Give it a collection of previous articles and tell it to imitate them. But a writing voice is much more complicated than that. It is partly vocabulary and sentence structure, of course. Some writers prefer short sentences. Others use long paragraphs in which ideas gradually develop. Some use humour frequently, while others prefer a more factual approach. Some start almost every article with an anecdote or personal observation.

But voice is also about the choices behind the words. What do you consider important? What deserves an explanation? When do you add context? Which subjects do you approach critically? And perhaps equally important: what kind of language would you never use?

That is what makes Feather’s experiment so interesting. He isn’t merely trying to make an AI reproduce the superficial characteristics of his writing. He wants to see whether it can understand the patterns behind it. As he explains in his original blog post, the process becomes an exercise in defining what actually makes his writing his own.

AI as an apprentice writer

One useful way of looking at this is to think of AI as an apprentice rather than an automatic content generator. Imagine working with a new editor who has never seen your work before. Initially, that person might make perfectly reasonable suggestions that nevertheless don’t sound like you.

Over time, however, an editor learns. They discover that you dislike certain expressions. They notice that you prefer examples over abstract explanations. They understand that you don’t want every article to end with a grand conclusion. Eventually, they start anticipating your preferences.

Something similar can happen when working with AI. The difference is that we need to make many of those preferences explicit. And that can be surprisingly difficult. We often know immediately when a piece of text doesn’t sound like us. Explaining why is much harder.

Your archive becomes valuable

There is another interesting aspect to Feather’s experiment: the value of an existing writing archive. Someone who has been blogging for years may have hundreds or even thousands of articles online. Together, those posts contain an enormous amount of information about that person’s interests, language, opinions and writing habits.

Until recently, that archive was primarily intended for human readers. AI changes that. A collection of old blog posts can also become reference material for a personal writing assistant. Instead of starting every conversation with an empty AI system and explaining your preferences again, the technology could potentially learn from years of previous work. For bloggers and other people who write frequently, that is an intriguing possibility.

But imitation isn’t identity

There is an obvious limitation. AI can recognise patterns in writing, but those patterns are ultimately the result of something the AI doesn’t possess: a life. A personal blog reflects conversations, travel, work, mistakes, discoveries, books, music, people and countless other experiences. Those experiences influence how someone writes.

An AI can learn that a writer frequently connects technology with everyday experiences. It can recognise that personal anecdotes often appear at the beginning of articles. It might even become remarkably good at reproducing that pattern. But it didn’t have those experiences. That distinction matters, particularly with personal publishing.

A different way of thinking about AI

Perhaps that is also why Feather’s experiment points towards a more interesting role for AI in writing. The discussion around generative AI often focuses on replacement: can AI write an article instead of a journalist, blogger, copywriter or author? But that may be the less interesting question.

A much more useful question is whether AI can become sufficiently familiar with an individual writer that it can help that person write better and more efficiently without removing the personality from the result. It could remember previous stories. It could point out that you wrote about a similar subject three years ago. It could suggest connections between ideas. It could produce a first draft that already resembles your normal writing instead of delivering the familiar, polished-but-generic AI prose. The writer would still decide what is worth saying.

Making AI more personal

That is ultimately what makes Feather’s experiment worth reading. It turns the usual AI-writing discussion around. Instead of asking how writers should adapt themselves to AI, the question becomes: how can we make AI adapt itself to the writer? That sounds like a subtle difference, but it is an important one. The most interesting future for AI-assisted writing may not be a world in which machines generate enormous quantities of anonymous content. It may be one in which our tools become increasingly familiar with the way we think, write and communicate.

Not an AI that writes for us, in other words. But one that gradually learns how to write with us.