Comment by doe_eyes

But you can already see it with Delve. Mistral uses "delve" more than baseline, because it was trained on GPT.

So it's classic positive feedback. LLM uses delve more, delve appears in training data more, LLM uses delve more...

Who knows what other semantic quirks are being amplified like this. It could be something much more subtle, like cadence or sentence structure. I already notice that GPT has a "tone" and Claude has a "tone" and they're all sort of "GPT-like." I've read comments online that stop and make me question whether they're coming from a bot, just because their word choice and structure echoes GPT. It will sink into human writing too, since everyone is learning in high school and college that the way you write is by asking GPT for a first draft and then tweaking it (or not).

Unfortunately, I think human and machine generated text are entirely miscible. There is no "baseline" outside the machines, other than from pre-2022 text. Like pre-atomic steel.

1. People don't generally use the (big, whole-web-corpus-trained) general-purpose LLM base-models to generate bot slop for the web. Paying per API call to generate that kind of stuff would be far too expensive; it'd be like paying for eStamps to send spam email. Spambot developers use smaller open-source models, trained on much smaller corpuses, sized and quantized to generate text that's "just good enough" to pass muster. This creates a sampling bias in the word-associational "knowledge" the model is working from when generating.

2. Given how LLMs work, a prompt is a bias — they're one-and-the-same. You can't ask an LLM to write you a mystery novel without it somewhat adopting the writing quirks common to the particular mystery novels it has "read." Even the writing style you use in your prompt influences this bias. (It's common advice among "AI character" chatbot authors, to write the "character card" describing a character, in the style that you want the character speaking in, for exactly this reason.) Whatever prompt the developer uses, is going to bias the bot away from the statistical norm, toward the writing-style elements that exist within whatever hypersphere of association-space contains plausible completions of the prompt.

3. Bot authors do SEO too! They take the tf-idf metrics and keyword stuffing, and turn it into training data to fine-tune models, in effect creating "automated SEO experts" that write in the SEO-compatible style by default. (And in so doing, they introduce unintentional further bias, given that the SEO-optimized training dataset likely is not an otherwise-perfect representative sampling of writing style for the target language.)

  Too deep we delved, and awoke the ancient delves.

> LLM content should just enhance and cement the status quo word frequencies.

TFA mentions this hasn't been the case.

The Inhuman Centipede