AI for writing: unexpected utility

I know that some/many people use LLMs (ChatGPT etc) for writing (emails, docs, etc), but I’ve never found it worthwhile myself, for at least two reasons.

One, the results always look unnatural and “AI-written” (even with system prompts to make it more casual or concise etc—this post is worth reading though). My writing style may be nothing to be proud of, but it is mine, and I do want anything that comes from me to look like I wrote it.

Two, and the bigger reason, is that writing is a way of thinking. The act of writing involves rearranging my inchoate thoughts into something linear, and making explicit what was vague. Having some external system do that bypasses what was actually valuable about the process. One may think that having the AI come up with something and then reviewing it to fix its errors can be helpful (a la Cunningham’s law), but I haven’t found this to be the case in practice: trying to deal with the thoughts suggested by the AI’s writing has an effect more like what Kumārajīva said about translation:

It is like feeding another person with chewed-over rice. Not only is the flavour lost, it will cause the other person to vomit.

This has been my experience each time I’ve tried.

BUT!

A couple of weeks ago, there was an email I needed to send, and it had been a few days since I first realized I needed to send it. It was nothing big, just some purely administrative thing, though I had been putting it off.

So on a whim (and because at this point I cared more about the email getting sent than how genuine it sounded), I decided to try one of the LLMs (don’t even remember now whether it was Claude, ChatGPT or Gemini): I opened it, started typing: “Write an email to…” and explaining the situation. I reached the end, hit Enter, and without even waiting for the response, I switched back to the Gmail tab, easily wrote the email, and sent it.

So this is a case where the writing abilities of AI were genuinely useful to me: even though I didn’t look at its output, merely knowing that it was capable of the task was enough for me to initiate the sequence of actions which led to the email getting sent. In effect, the chatbot acted as a kind of rubber duck.

As long as there have been writers, there has been writer’s block. A tool whose existence can go some way towards alleviating it is very valuable. Beyond a certain baseline level of intelligence, all human problems are emotional anyway. Counterintuitively, it is possible that a great benefit from unemotional thinking machines may be in humans' emotional management.