An essay on AI writing reached 93 points on Hacker News

An essay posted to a personal blog is sitting at 93 points and 61 comments on Hacker News. The post is "On AI writing" by Dima Bolkunov, and the argument it makes is the one most teams putting prose through an AI pipeline this year have not stopped to answer. The author's complaint, in one line he writes himself: he reads AI-generated text "at least a third slower" than human-written text of the same length.

Bolkunov writes from a specific angle. His bachelor's thesis was about telling authors apart by the statistical shape of their word use, which is a real body of work that predates the current writing-assistant market, so his claim about losing style when a model averages it carries more than opinion. The HN discussion is in agreement with the direction of his argument, but splits on whether he overshoots.

What Bolkunov argues

He makes four claims in sequence. The first is about style: a model trained on an average of writers averages out the writer, which hurts a reader's ability to see the person behind the text and to extend them any trust. The second is about precision: a model has to make assumptions to stitch a draft together, and those assumptions push the writer toward safer, more ambiguous language than they would have reached on their own. The third is about thinking: putting words on a page is how people find the gaps in their own argument, and an AI-written first draft takes that step away.

The fourth is the one with a number against it. He reports reading AI-written text "at least a third slower" at the same length, because he has to translate the ambiguous phrasing back into something firmer as he reads, and he has less of the author's context to help. He calls this a cost transfer: the writer saves time, every reader loses it, and across a large audience the sum is negative.

Where the Hacker News thread agrees and disagrees

The thread's own top comment reads a style tell: an attentive reader catches a certain shape of AI prose, which one commenter calls "a cross between an exec assistant who just started using Grammarly and a junior beat journalist". A separate thread flags the authenticity cost on code review: one commenter writes that when an AI-written pull request has somebody's name on it, "I can't bring myself to see it as yours", so the attribution carries less weight than the writer intended.

Pushback comes from a comparison to past standardisation. A commenter writes that the same arguments were made against dictionaries and spelling rules, and that making writing more interoperable did not kill writing. They add that language models can be used in the opposite direction too, as critical readers that interrogate a claim, check evidence and expose a gap in an argument, and that calling AI writing a loss of intellectual work is too broad.

A team shipping prose through a model has a design choice sitting under all of this. The choice is whether the writer's one-third saving is bigger than every reader's one-third loss, over the audience the piece actually reaches. For an internal memo of five readers, the sum is probably positive. For a public blog or a product page, Bolkunov's maths tilts the other way.

Source

• On AI writing by Dima Bolkunov
• Hacker News discussion, 93 points, 61 comments