One of the least controversial opinions one may hold is that putting thoughts into writing requires discipline and a certain level of clarity. At least, that was the case pre-AI, when creating a paragraph of written language required significant effort.
In the last couple of years, that proof-of-work mechanism has arguably given way to a tsunami of generated work of variable quality. Depending on the platform and the medium, you may notice more or less of it.
During a July discussion around AI in the literary market, I liked a note from Byrne Hobart in The Diff most. It referred back to Hachette's decision to cancel publication of Shy Girl over evidence of AI use:
“It only gets easier to detect LLM use over time, so we'll all be judged based on the norms and evidence-gathering techniques of the future.”
Byrne Hobart, The Diff
That sounds very rational and true to me, at least for past and current generations of models.
Looking back at my own posting experience over the last few years, the usefulness of every new model release for a non-native speaker was also clear: first punctuation and structure improved, then clarity and sourcing, and later the analysis itself for essay-style pieces. In sync with the capabilities—writing, coding, and the rest—the posts became more substantive.
But they also became, if not voiceless, certainly less distinctive. At some point, the authorship provenance became clearer, especially with the Claude/Codex family of models.
My takeaway is that it can still be useful to look at less-discussed topics, underexplored data, or similar things, and employ the latest battery of tools to work through them. But the touch of humanness is still a kind of agency that agents will miss, as well as the curiosity, care, and randomness that each of us develops over a lifetime.
And to keep that randomness in place, I want future posts here to contain my writing, and not my models' writing.