Part 1 of 2. Part 2: How to check your own writing over time
I started with a simple question. Could an AI notice early signs of dementia or Parkinson’s just from the way someone writes? I mean everyday writing: emails, chats, the messages people type to a chatbot.
This one is personal. My grandmother had Parkinson’s, and it’s what she died of. She was an English teacher in Miami, and I eventually got an English Literature degree — we have that in common. So when I read that early signs might show up in how people type and write, I wanted to understand it.
The short answer: sort of. Then I did what I always do with a new idea: I tried it on myself.
I’m in an unusual spot for this. I’ve run StarBase 118, a Star Trek writing community, for more than 30 years, so I have decades of my own writing to work with. A while back I backed up all of my email to a drive in my office: about 1.2 million messages, including three decades of StarBase 118 history. I also had old downloads of my Twitter and Facebook accounts and years of journal entries. It’s a lot of me, in my own words, over a long stretch of time.
What the research says
Changes can show up years early. For example, researchers studied Agatha Christie’s novels and found her vocabulary shrank noticeably in the books she wrote in her 80s. She used more vague words and repeated more phrases. Iris Murdoch’s last novel, written as her Alzheimer’s was developing, showed the same kind of drop.
But it’s a hint, not a diagnosis. One of the best studies used short written descriptions from the long-running Framingham Heart Study. It correctly predicted who would later develop Alzheimer’s about 70% of the time. It also missed about 4 in 10 future cases. That’s interesting science, but it’s not something you’d want telling you (or your mom) that something is wrong.
Context can fool it. Take Ronald Reagan: an early study said his press conference language declined over his presidency, but a 2024 reanalysis found the change mostly came from reporters asking him different kinds of questions. The words changed because the situation changed.
Parkinson’s shows up in typing mechanics more than in words. MIT researchers looked at how long you hold each key down, and how much that varies. That timing could pick out early Parkinson’s about as well as some clinical tests!
What I fed it
I pulled together about 82,000 pieces of my own writing from 1995 to 2026:
- School papers from high school and college
- More than 25 years of StarBase 118 forum posts and mailing list emails
- My personal and work emails
- Twitter and Facebook posts
- My journal
- The messages I type to Claude every day
All of it stayed on my own computer. A script measured the writing there, and only the numbers came out.
What it measured
These are the same kinds of things the research looks at:
- Vocabulary variety: how many different words I use
- Sentence length: decline tends to make sentences shorter and simpler
- Vague words: “thing,” “stuff” and “something,” which creep in when the right word gets harder to find
- Pronouns: more “it” and “this” in place of actual nouns
- Spelling: how many words aren’t in the dictionary
- Long words: decline tends to push people toward shorter ones
Then it compared my writing to my own past writing, not to anyone else’s. Everyone writes differently. What matters is whether you are changing.
What I found
No sign of decline. PHEW! A few things surprised me, though:
- My vocabulary held steady or grew. In my StarBase 118 forum posts, it barely moved (upward direction) from 2002 to 2024. In my personal email it went up.
- My sentences got longer, not shorter. My forum posts went from about 15 words per sentence in 2002 to about 24 in 2024. That’s the opposite of what decline looks like. (It might just mean I got more long-winded. I’m choosing to call it growth.)
- My spelling hasn’t changed in 25 years. About 2.5% of my forum words weren’t in the dictionary in 2002, and about the same in 2026. Most of them aren’t typos at all. Instead, they’re words like “email,” “wiki” and “simming,” which is what we call writing a Star Trek scene together.
Two small things drifted: I use slightly more vague words in the forum than I did in the 2000s. And my personal email uses more pronouns since 2022. Neither shows up anywhere else in my writing. That points to a change in who I’m writing to and what about. My personal email today is mostly chatty notes to friends instead of work.
I also compared the last two years against my “normal” from 2015 to 2023. Everything was within my usual range.
What I didn’t expect: AI is erasing our baseline
Since 2025, a lot of what I write starts as an AI draft that I then edit. That’s great for getting things done. It’s terrible for this kind of check. If an AI picks my words, my words stop telling you anything about my brain.
So I had to set aside a chunk of my recent writing. The cleanest record of how I write today turned out to be the messages I type to Claude. They’re fast, messy and 100% me.
Even if you never run a check like this, keep some writing that’s only yours. A journal. Texts to friends. The prompts you type to an AI. Ten years from now, that might be the only honest record of how you used to write.
What I’m doing next
Typing can show things that writing can’t, so I also set up a small typing logger on my laptop. It records how long I hold keys down and the gaps between them. (It never records which keys or what I type.) It needs a couple of months to learn my normal. Then it’ll show me if anything starts to drift.
I’ll rerun the writing check once a year.
What this isn’t
This isn’t a medical test, and a clean result isn’t a clean bill of health. Mood, sleep, medication, a new job or a new keyboard can all move these numbers. If you’re actually worried about yourself or someone you love, talk to a doctor. They have real tests that work far better than anything here.
Oh, and: Please don’t run this on someone else’s writing without asking them first. It’s their writing and their brain.
Want to try it on your own writing? Part 2 walks through how, step by step, with an AI helper doing the technical parts.
Sources
- Alzheimer’s Early Tell: The Language of Authors Who Suffered from Dementia, Cornell Human Development Outreach
- AI can predict the onset of Alzheimer’s disease, IBM Research (Framingham Heart Study)
- Did Reagan Decline?, Beckmann, Journal of Alzheimer’s Disease (2024)
- Computer keyboard interaction as an indicator of early Parkinson’s disease, Giancardo et al., Scientific Reports (2016)
