The Science of Why AI Still Can’t Write Like You

TL;DR: Why does AI writing still sound like AI writing, even as the models get smarter? In his first piece since joining Every as Spiral’s general manager, Marcus Moretti explains why the answer is more complicated than you’d think. The most reliable fingerprints of your personal style come from the words you write subconsciously: articles, pronouns, and function words that emerge in a distinctive pattern as you focus on the meaning of a sentence. His piece explores what new research in machine learning and stylometry—the study of style—means for the future of writing tools like Spiral. If you want to go deeper, Spiral has several updates, including creating a writing style from your website or X account (even taking post engagement into account) and a cleaner, faster editor.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

OpenAI models demonstrate Ph.D.-level knowledge across physics, biology, and chemistry. Anthropic staff have claimed its Opus 4.5 model “largely solved coding.”

Yet AI writing remains stubbornly detectable: “It’s not an idea. It’s a breakthrough.” “Delve.” Lists of threes with no “and.”

If you’re a regular Every reader, you may already know why this is. LLMs are trained on an unfathomable amount of words and learn generally how to speak. Post-training, which refines a model after initial training on large datasets, makes the models friendlier and safer, so they end up speaking in a kind of generic politeness. Ted Chiang’s description from a few years ago remains apt: “ChatGPT is a blurry JPEG of the web”—a tool that approximates human insight without ever landing on the mark.

I’m interested in the relationship between LLMs and writing style because I’m the general manager of Spiral, Every’s AI co-writer. Writing sessions in Spiral begin as a chat: You describe what you intend to write, and Spiral helps you hone your message and gather relevant research. Then it produces one or more drafts, offering several approaches for your piece.

Our aim is for Spiral’s written output to reflect your personal writing style, not the generic politeness of the foundational model. To this end, I’ve been reading papers on natural language processing, linguistic forensics, and stylometry—the study of writing styles. It wasn’t until I started working on Spiral that I became aware of the century-plus history of stylometry, or of the fastidiousness with which researchers have catalogued the elements of style. In recent years, researchers in these fields have flocked to LLMs, finding new ways to expand our understanding of human writing. Here are some findings that I found interesting and even counterintuitive, and that provide a hint as to where AI writing might be headed.

Looking for an AI notetaker for your meetings?

Granola is a lot more. Most AI note-takers just transcribe what was said and send you a summary after the call. Granola is an AI notepad. And that difference matters. You start with a clean, simple notepad. You jot down what matters to you and, in the background, Granola transcribes the meeting. When the meeting ends, Granola uses your notes to generate clearer summaries, action items, and next steps, all from your point of view.

Then comes the powerful part: You can chat with your notes. Use Recipes (pre-made prompts) to write follow-up emails, pull out decisions, prep for your next meeting, or turn conversations into real work in seconds. Think of it as a super-smart notes app that actually understands your meetings.

Download Granola and try it for your next meeting. Three months’ free with the code EVERY.

Want to sponsor Every? Click here.

Subconscious decisions define writing styles

Stylometry has had a few moments of glory. In the 1800s, stylometrists gave sold-out lectures about whether William Shakespeare wrote those plays. In the 1960s, two stylometrists isolated Alexander Hamilton’s contributions to The Federalist Papers based largely on the presence of the word “upon.”

In the 2020s, LLMs have introduced new ways of studying style. Last year, two Cornell University researchers systematically manipulated text snippets to see how it affected LLMs’ ability to guess their authors. They removed an attribute of the text one at a time—such as proper nouns or capitalization—and measured the effect on attribution accuracy.

They found that removing the more functional features of the text caused the models to misattribute authorship more often, proving that those features are most helpful for attribution. In particular, removing “stop words” made it a lot harder to guess who wrote something. In natural language processing, stop words are common, functional words like articles (“a,” “the”) or pronouns (“I,” “she”). These words are often filtered out of text analysis because they don’t convey much meaning, but it turns out that they appear in patterns that can help identify who wrote something. This is why Hamilton’s use of “upon” tipped off those researchers to his Federalist contributions.

Things like stop words and word order turn out to be some of the most distinctive markers of someone’s writing style. These purely functional aspects of writing mostly reflect subconscious decisions. When we write, we focus on choosing meaningful words, and our subconscious tends to fill in the rest. But the way our subconscious contributes to our sentences is to be distinctive.

Most people are wildly inconsistent writers

“How Well Do LLMs Imitate Human Writing Style?” asked another paper co-authored last year by a researcher at Bucknell University. The authors used various methods to get an LLM to copy someone’s style, finding that just a few writing samples increased style fidelity over the base output by about 23 times. A smattering of examples went a long way.

They then tested whether AI text could be identified as such even if it accurately reproduced someone’s style. They discovered it can be. Natural language processing has something called a “perplexity” score, and the greater the linguistic variance, or the diversity and unpredictability of word choice and sentence structure across a writing sample, the higher its perplexity. The researchers found that, on average, humans are twice as varied in their writing as machines.

Writing styles change dramatically over time

An LLM is just a string of numbers. When you interact with ChatGPT or Claude, you’re “talking” to a static set of digital files. LLMs are point-in-time snapshots of human language, which is why they need to search the web for information after their training cutoff date.

Language itself, however, rapidly evolves. In the book Algospeak, the self-styled “etymology nerd” and linguist Adam Aleksic argues that our vocabulary is evolving faster than ever, due to social media and hyperconnectivity. This poses a problem for LLMs. What good is a model if its training run ended before we started saying “skibidi”?

In October, Sushil Khairnar, a graduate student at Virginia Tech, tried to quantify models’ “temporal drift.” He found that GPT-2 and GPT-3’s manner of speaking lagged behind the general lexicon by about 15 percent a year after its release and 28 percent after two years.

Ironically, LLMs themselves are altering language, including in academic research. Post-ChatGPT papers include significantly more AI-ese: words like “underscore,” “highlight,” and “showcase.” “Delve” is the biggest culprit, with usage in papers skyrocketing by more than 2000 percent between 2022 and 2024.

What’s next

As foundation models and methods of style transfer improve, computers will get better at mimicking individual writing styles. It’s an open question of how close they can get. The better we understand the science of style, the more we can bridge the gap between model output and manual output.

The research guides Spiral’s roadmap. As an example, we recently updated how Spiral generates a style guide from a user’s writing samples, which then stylizes Spiral’s drafts. We previously generated hundreds of descriptions of the user’s writing, but now we focus on the key textual identifiers and quintessential phrases from the source material. And we’re building connections to writing sources—blogs, newsletters, and social media—so Spiral can keep up with how your personal style evolves over time.

For any LLM-generated writing, though, there will always be some gap—after all, a person didn’t write it. That fact may be harder and harder to detect in the output, but it’s always worth considering what your reader would think upon learning that your piece was AI-assisted or -generated. This piece, for example, was written the old-fashioned way, despite the em-dashes, which Every’s style guide allows. Unless you want to analyze the stop words, you’ll have to take my word for it.

Spiral has been busy the past few weeks. New to the app: connect your Twitter account and get an engagement-weighted style guide, allowing Spiral to draft bespoke tweets for your audience. Link your website or RSS feed to teach Spiral your style via bulk post import. Workspaces now make it easy for you to share styles across your team, so you can write in one unified voice.

Marcus Moretti is the general manager of Spiral (@tryspiral).

To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization. Discover Every’s upcoming workshops and camps, and access recordings from past events.

For sponsorship opportunities, reach out to [email protected].