LLM text makes human interactions less fun

All humans are unique. We look different, have different voices, different mannerisms, different fashion preferences, different ways of speaking, and different personalities. Part of what makes life fun, is getting to know those other humans, and building deeper connections with them.

LLMs are, in one sense, all the same. Everybody uses the same foundation models, and the handful that are there, are trained in a way that makes their default writing styles very similar.

I have a remote job in tech. This means that a large part of my interactions with other humans happens online, through text. In text, the things that set us apart are stripped away. You don’t hear voices or intonation, you don’t see faces and body language. We may get to pick a tiny picture as avatar, but even the fonts are uniform. The one thing that remains in text — the last trace of personality that cannot be stripped away — is writing style.

Learning to recognize somebodies writing style is part of getting to know that person, just like learning what their taste in food or music is. With people I know well, I can often identify the author by the writing style in natural languages that I’m fluent in, and sometimes even in code. That we express ourselves in different ways, is what makes interacting with other people fun!

When people start using LLMs to inflate text for them, or to rewrite text to sound more professional or eloquent, that last trace of personality is lost. I can often spot LLM use by the writing style, and it makes me sad when real humans do that. It’s the equivalent of turning off your camera in a video call: hiding behind uniformity, erasing personality.

When LLMs became popular, I initially thought they could bring productivity gains, and help less advanced speakers express themselves. But as LLMs get integrated into our daily lives, LLM-generated text starts to feel more like a double insult: it says “I couldn’t be bothered to spend time writing this myself,” yet it makes me waste time reading through the generated fluff! It would be more respectful of each other’s time to just share the prompt directly. As for helping less advanced speakers, I now realize that I much prefer to read broken English, over reading something that sounds like AI slop. Because then, at least I feel like I’m interacting with a person!

All of this is not to say that LLMs are not useful, or that they can’t impersonate different writing styles when prompted. It’s just that when I’m interacting with other people, I like to interact with those people directly, and not through an LLM filter that strips away personality, and removes the joy from human interactions.