Hacker News Reader

Comment by burkaman | Hacker News Reader

Obviously the AI version is bland and terrible, but arguably more importantly it has also completely changed the meaning of the message. The AI version:

- apologizes

- implies the recipient was "promised" this email as a "response" to something

- blames a hectic schedule

- invites questions

None of this was in or was even implied in the original. This is not a "polished" version, it's just a straight-up different email. I thought that style transfer while maintaining meaning was one of the few things LLMs can be good at, but this example fails even that low bar.

View on HN

The AI has some ... "ideas" ... of its own on what workplace relationships apparently need to be like.

A lot of people who want to replace most human interactions with LLMs assume that there is some objective set of cultural values true in all contexts, and that it is good and easy to encode these as axioms into an AI.

Yes. As an Old now (GenX), I feel like moving all interactions to text and now having AI as a man in the middle is just reinventing ways to get in a situation where a decade down the line you reconnect with someone who used to be a friend and both discover "Hey, that wasn't what I meant at all!"

As ever, T.S. Eliot was right: "It is impossible to say just what I mean!"

And those objective set of cultural values are, apparently... a sort of parody of 90s corporate culture, a sort of polite version of Michael Scott. Like, no-one ever _actually_ wrote like LLMs tend to write; it reads as a parody of a now slightly obsolete corporate-speak.

Have you read the papers on how they optimize these LLMs for demeanor?

AI exists in a Matrix where toxic positivity is enforced with electric shocks.

It's the California Ideology, now written into a program.

When I was at Google this tone was so pervasive that we had "CongratBot" to make fun of this stuff, and Memegen to somewhat counteract it.

But Silly Valley is the poster-child for "hey don't be a downer" communications vibes.

As the child of a super-critical German father, it always seems insincere to me.

And those ideas seem far more in line with millenial Silicon Valley culture. It's weird when they expect Germans to fake that sort of overly formal, overly cheery tone. People just don't talk like that.

Yeah, my father is German, and I'm Canadian, but of a .. grouchy ... variety. This kind of business culture has never really fit with me. But it's how the people with the $$ speak.

Correct. This is called the production of subjectivity.

(tyrna be funny not patronizing. but the machinery of subjectivity production is ofc very real)

this is like when my manager once yelled at me for not writing in corp speak enough

“OK Fine. But could you at least yell at me in corp speak?”

It's no surprise LLMs are using corp speak and vapid marketing prose as a template. There is so much of it out there.

This is from that Autodesk post last week where they admitted their mistake and… Nope it's corp speak:

“We are excited to share some important updates regarding Archiving and our Idea Boards and Forums that aim to enhance your experience and ensure valuable content remains accessible. Please read the details below to understand how these changes might impact you.”

Barf. But to an LLM this looks like a human communicating in a meaningful way.

> Look, we need to align on language here. If you’re not speaking in scalable, results-driven terminology, you’re slowing down the team. We don’t “talk about things”—we sync and strategize. We don’t “try something new”—we leverage data-driven insights to drive innovation.

It's just shitty prompt design

Well... that's a very 2025 sentence...

No, it's because the AI makes shit up. No amount of prompting will fix this.

Reply on news.ycombinator.com