This seems like a valuable utility for concealing writing style, though I feel like the provided example fails to illustrate the rest of the stated goal of the project, which is to “prevent biases, ensuring that the content is judged solely on its merits rather than on preconceived notions about the writer” and “enhance objectivity, allowing ideas to be received more universally”.
The example given is:
You: This is a demo of TextCloak!!!
Model: “Hey, I just wanted to share something cool with you guys. Check out this thing called TextCloak - it’s pretty neat!”
The model here is injecting bias that wasn’t present in the input (claims it is cool and neat) and adds pointlessly gendered words (you guys) and changes the tone drastically (from a more technical tone to a playful social-media style). These kinds of changes and additions are actually increasing the likelihood that a reader will form preconceived notions about the writer. (In this case, the writer ends up sounding socially frivolous and oblivious compared to the already neutral input text.)
This tool would be significantly more useful if it detected and preserved the tone and informational intent of input text.
You could argue that the point is to conceal your identity and if you suddenly sound like someone else or it’s just really obvious that you ran your text through an LLM then it kinda does the job no? As long as it’s not introducing biases that are connected to your own biases it shouldn’t be an issue.
Right! It’s definitely fulfilling the purpose OP stated here in this post, as long as that’s what you’re using it for. I’m just pointing out that it doesn’t do the other things it claims to do in the readme for the repo, so that’s something to be aware of.
