Preserving Individuality in AI-Assisted Communication

Most writing tools quietly average you toward a mean. Ask a model to "improve" a message and it hands back something cleaner, blander, and less like you. My dissertation asked a narrower question: can a system help you write without erasing the fingerprint that made the writing yours?

What the pipeline actually measures

The Voice model reads 47 stylometric features across several tiers: surface patterns, lexical choices, syntax, psycholinguistic signals, character level habits, and function word bigrams. None of these is about what you say. They are about how you say it, the part that survives across topics.

Why prompting alone was not enough

A key finding was that prompt based conditioning has a structural ceiling. You can describe a voice to a model, but describing is not the same as embodying it. The interesting work sits between prompting and full fine tuning.

More soon. This note will grow as the paper does.