ChatGPT as it stands has already been a wonderful tool for generating text, especially when it’s taking the place of myself, but lately, I’ve been coming across the issue of how time-consuming it can be to rewrite the text provided to make it sound as if I wrote it, especially as an Australian English speaker with a, well, unique writing style. Right now, you can use custom GPTs or make new chats that can somewhat imitate your tone, but they will still feel & read as AI-generated instead of matching how the user-provided training material actually reads.
I would like to suggest a persistent, trainable profile that can be invoked as necessary, perhaps under the same drop-down menu where you choose what model you are using. Something that would allow a user to feed ChatGPT as many examples of their writing as they please & answer direct example questions to help the LLM understand better how the user responds to input. For example, it could ask the user to answer a complex but concise question in their own style. This could be repeatedly updated with new information & active feedback from the user, as well as making direct changes, if possible, to the deterministic part of an LLM—that is, changing the probability of what word the model chooses next, adjusting for things like user sentence structure, vocabulary, punctuation preference, overall rhythm & preferred word choice. It wouldn’t just try to mimic broad categories like “formal” or “casual” (a trap I’ve noticed existing models fall into when I try to train them) but instead generate responses that are indistinguishable from the user’s actual writing.
The custom profile would ideally dynamically weigh its responses. It would judge things like approximate sentence length, whether the user prefers “&” over “and,” (such as myself) or how they structure transitions. The ideal result would be that users could manually tweak responses to correct things that feel off, with ChatGPT learning from those adjustments in real time, allowing for the complete elimination of the verbal (textual?) uncanny valley of AI-generated text.