Having recently read Paul Graham’s essay on the beginning divide between the “writes” and the “write-nots” in the wake of AI-based language models taking over many of our writing tasks for us, I have started to contemplate writing more.

While there’s no denying the appeal of language models and the efficiency they provide, especially for more mundane tasks like routine correspondence or boilerplate text generation, they have big drawbacks at this point in time. Even setting aside content issues like LLM hallucinations to focus on the writing itself, one of these drawbacks is that LLM-generated prose often feels boilerplate, even when written for more complex tasks. After seeing a few pieces of AI-generated text, it becomes increasingly apparent where the patterns lie, and as such these texts quickly stand out. In the short term, this impedes the reader’s flow and, to an extent, motivation to interact with the piece of writing at hand. If a writer put in only a half-hearted effort in creating a (meaningful) piece of writing, a reader could be forgiven for putting in a half-hearted attempt at processing it.

In the long run, this unfortunately also results in a sort of siloing of language. As more and more of the content that AI is trained on is itself AI-created, the sometimes bumpy turns of phrase, the unnatural-sounding language it uses, start to compound and become more commonplace in its output. Even employing ways to regularize this effect, chances are this compunding effect will over time contribute to the ever-present shift that languages undergo. We already know that use of technology can have a noticeable impact on language learning. I would love to see a more linguistic investigation into the effects of AI writing tools on language development, both personal and popular. Would it lead to more homogenous phrasing and slow disappearance of words that have low prediction rates, even in cases where they would be an ideal semantic match? It’s hard to say at this point.

While most of my writing these days is in areas where AI would only be marginally helpful anyway, I support Paul Graham’s point that the only way to keep personal writing skills honed is to write. At the same time, I have noticed that my verbal fluency and command of language in general have been affected by a traumatic brain injury I incurred 2 years ago in a traffic accident. Maybe there will be a benefit to writing more often, and about a larger variety of topics, in terms of re-developing some of it. As such I won’t limit myself to any one genre or type of content in writing this blog, but will instead roam around somewhere between technology, tutorials, politics, personal experiences and all sorts of weekly-changing interests. Let’s see where this goes.