Post

we really do need a discussion on how we as a society conceptualize "when you write a post and run it through a LLM, how much of that is YOU and your ACTUAL VOICE AND MEANING, and how much of that is the AIs"

Andrew Rothstein@asrothstein.bsky.social

Facebook asks you if you want to do that on every post. Gmail asks you if you want to do that on every email. It’s mind numbingly aggravating but increasingly acceptable.

10 reposts109 likes

I've begun to see small typos and swapped letters etc as evidence someone actually wrote something themselves. I see it in work chat a lot, and I hate it. Writing pre-AI, even just boring work stuff, was an opportunity for little bits of folks' personalities to squeak out, esp in remote teams

personally i loathe the practice and would never do it - to me, it's like how some cultures believe that a piece of your soul is taken each time you appear in a photograph.

still, it is increasingly common, doesn't seem likely to go away, and I also don't feel all that comfortable viewing these augmented but apparently actual-human posts the same way as I do totally prompt-generated text.

To me, the important thing is not letting the LLM be a means of evading responsibility. Want to punch up your email with Claude? Sure, whatever. But if it sucks I’m blaming you, not Claude.

But also, if I don’t already trust that your writing is worth reading, a clear AI Smell makes me much less likely to bother

I constantly call out this behavior on technical subreddits and as a writer, and more importantly a reader of a lot of content, I fucking hate it. I don't care if your English or grammar is bad, I want to read your thoughts, with your opinions.

I'll rarely use AI for a title idea or something, but if you are reading something I posted, I fucking wrote it. You can probably tell because I said fucking.

Rovo writes 90% of my JIRA tickets from run on sentence mind dumps I feed it. 1) I voice type a ton of information in a stream of consciousness 2) I tell rovo to make it into a ticket 3) I edit its output 4) I’ve now deleted 5-15min of low value typing work from my day

I know when I write I sometimes think VERY carefully about exactly which word I want to use. I couldn't ever be sure that a machine was conveying the right meaning if I let it edit what I wrote.

I discourage LLM use and I've seen it hollow out students' education. But I am not gonna tell another grown adult what to do, and the statement in question is in any case clear and forceful, whereas the main LLM problem is wishy-washy filler, mush.

If you let it do that then you don’t actually know what you agreeing to in emails and then you need to read ai summaries of the email thread (which are sososo shitty) to eventually know what you’re actually meant to do.

I think if people used AI to generate like a template and then just tweaked and used that template I'd still think it was silly but like it would be one prompt used multiple times. It doesn't seem like that's what is happening tho. Which makes it worse.

I am a huge fan of automation, snippets and templates. Why constantly work from scratch when semi I have serviceable shit I already wrote? I am profoundly efficient when it comes to rote and boring shit, but so far AI just seems to shit all over efficiency gains to worse ends!