The long and the tall of it

There’s an episode of Midsomer Murders where Barnaby cracks the case when he realizes two people who supposedly barely know each other are actually mother and daughter—because they both use the same mangled idiom (“the long and the tall of it” rather than “the long and the short of it”). It’s a clever little twist, and it has to be, since it’s probably been close to two decades since I’ve seen the episode.

Little linguistic habits, quirks, and mistakes like this can be inherited, imitated, or otherwise socially transmitted. It’s what makes language so distinct from region-to-region, city-to-city, even household-to-household. But it got me thinking about the AI writing style. Or more specifically, the AI editing style.

I often use AI to perform an editing pass on these blog posts, looking for spelling/grammar mistakes, tortured or overlong phrases, or factual or logical weaknesses (side note: do I need an explicit AI use policy for this blog? Probably!). It’s a genuinely useful tool that I believe improves my writing.

However, AI will often also make stylistic suggestions that amount to removing any personality, liveliness, or hyperbole my writing may have. Taking out my voice, basically. Sanding off all difference, all distinction, until nothing is left but the sacred median: the entire universe compacted down to the next most probable token.

The mechanism at work here is simple: AI has a strong pull toward patterns that are common in its training data. This often produces an identifiable style that is quite easy to spot for anyone who works frequently with AI-generated text. And AI writing is everywhere now, choking the discourse on social media and the broader internet (side note: it was ChatGPT that told me that my previous use of “Internet” with a capital “I” was anachronistic, and this is one suggestion I’ve largely chosen to adopt). Just as important, it makes writing harder to trace, since many will use AI editing to remove their distinctive styles and linguistic habits. Imagine trying to do stylometric analysis in the future!

What happens as a larger and larger percentage of the text we consume becomes partly or wholly AI-generated? At minimum, I suspect it means less linguistic diversity: fewer odd constructions, fewer distinctive turns of phrase, more of us flattened into the same indistinct grey paste.

The larger possibility is that this eventually affects not just how things are said, but what gets said (and thought). We’ve discussed on this blog before the AI tendency to hedge every argument until you end up with a dispassionate discursive writing exercise rather than an unpolished but full-blooded conviction rooted in the world you actually want to live in.

There’s a related phenomenon in finance, allegedly. Investor Gavin Baker recently claimed in an interview:

Everyone I know in the public equity investment business, whether retail or institutional, everything immediately, every piece of news gets fed into Claude […]. Claude is probabilistic. There’s probably not that much variation in the way it’s interpreting this news. People talk about the fragmentation of media and how it used to be like Walter Cronkite was the only voice of truth, and now we don’t have that anymore. It’s like Claude is Walter Cronkite for the stock market, and everybody just believes whatever it says.

The risk is modelthink: groupthink for the AI era. If Claude is wrong about something, then everyone is wrong about something. In markets, this creates correlated risk. In the broader culture, the risk is a narrower range of ideas that feel natural, reasonable, or even available to us.

There are ways this could be wrong, of course, such as if AI lowers the bar of entry for writing and encourages more and different kinds of people to share their viewpoints. The clearest example of this would be people writing in a non-native language, particularly English. Certainly we have seen this happen with programming languages, with nearly anyone able to create a landing page or generic product demo with minimal effort and guidance. AI could increase the diversity of people contributing to the discourse even as it reduces the diversity of how they express themselves.

Maybe that trade-off is worth it, but I’d rather keep the mistakes.

AI would have never accepted “the long and the tall of it”.

For transparency, my AI editor suggested removing the hyphens from the x-to-x terms and using and rather than slashes, as in “spelling/grammar mistakes”. I will not! (AI also strongly prefers em dashes for parentheticals, whereas I prefer literal parentheses.)