Like many people, I use AI to polish my posts. It has become almost a reflex, that small doubt creeping in, the need to double-check, to ask for a second opinion.
But when everything gets averaged — as LLMs do through statistical operations — the result isn’t necessarily better. In this constant stream of AI-smoothed content, it’s easy to lose one’s voice.
So I decided to find out through the first in a series of micro-games developed with Claude Code, putting both AI and ourselves to the test. Have we started preferring the AI version over the original voice?
I asked ChatGPT which prompts people most often use to refine their writing. At the top: rewrite and improve. Then came fix, make it better, polish, refine.
I staged a rhetorical fight — in the vein of a retro arcade Street Fighter match. The player chooses between an original quotation and its AI-generated “improved” version.
I instructed the AI: “Improve and completely rewrite the quote, staying true to the intent while making it better.”
“Improved” and “better” were deliberately left undefined — much like when we ask AI to refine our own text.
I tested figures across a wide spectrum — philosophers, comedians, politicians, cultural icons: Umberto Eco, Simone de Beauvoir, Ali Wong, Jesus, Kim Kardashian, Karl Lagerfeld, Marilyn Monroe, Barack Obama, Margaret Thatcher, and more.
It turns out I preferred the AI version over the original 34% of the time. I tended to favor the original when personality, wit, and unnecessary but colorful details were present.
After all, public figures have always been culturally assimilated and their words distorted. AI may simply be the next iteration of that process. Texts dissolve into data. Words feed a leviathan.
Perhaps the real question is not whether humans or machines win the duel I theatrically staged, but whether we can remain attentive to our own voice while participating in shared systems of knowledge that inevitably shape it.
The game was developed using Claude Artifacts, Claude Code and API (Haiku 4.5).