Notes on LLM Assisted Communication

When asked to send messages in professional settings, I found that a lot of people default to "oh, I'll let claude|gemini|chatgpt write this for me", without really considering the consequences. I hear people say that they feel the LLM does a better job at writing the text than them. Other people think that this communication might not be worth it for them to spend time on writing everything out formally. I want to make the case here that for a lot of applications, writing text yourself is paramount to getting the right context across.

Having been on both the sending and receiving end of LLM text in various circumstances, I have gained a lot of appreciation for what text can be. I found that almost all LLM generated text is generally lacking in precision and purpose. This is not necessarily due to LLMs being bad, and has more to do with the process of how these properties appear in writing. I would define these two properties roughly as:

• Precision: Wording the text such that the exact nature of the issue, and the actions you want them to take, are obvious to the reader.
• Purpose: Restricting the points and arguments you make to just the things you want the other person to get out of the message, so no wandering sentences about things that don't actually matter to the issue at hand-

I found that LLM messages tend to (1) need a lot of prompting in order to be able to understand the nuance of the problem and (2) are much more bloated than hand-written text. The first point means that low-context prompts result in imprecise messages. If I want to get a high-precision message out of an LLM, I need to write a detailed prompt. Now, the LLM will generate text from this prompt that changes some words into ones that may mean something slightly different, while also adding a lot of filler in between. In the end, I feel that the prompt I crafted for the LLM would make for a better message than the one the LLM produced. So I cut out the middleman and just write an e-mail from the start.

In the end, I feel that effective communication (especially in professional settings) requires purposefully crafted messages with precise wording. The context needed to make an LLM generate such messages often exceeds the work required to write such messages from scratch. Hence, using an LLM for such tasks will either result in bad communication, or in more effort than doing it "manually".