Preserving the Authorial Voice in an Age of Synthetic Prose

Every scholarly author who has run a draft through an AI writing assistant and accepted its suggestions has, in some small measure, handed the argument to someone else. That is a plain description of what happens when you let a language model rewrite your sentences.

The model does not know what you are trying to say. It knows what sentences that look like yours tend to be followed by. It optimizes for coherence at the surface level, for prose that reads smoothly and moves without friction. What it cannot do is tell apart two kinds of rough sentence: the one that reads a little rough because you haven't found the right words yet, and the one that reads a little rough because you are doing something difficult, something the available smooth formulations would fail to capture.

Authorial voice is the thinking itself — the particular way a trained mind in a particular discipline approaches a problem, qualifies a claim, and positions an argument in relation to the literature that came before it. Two authors working on identical data sets will write different dissertations. The difference goes deeper than style. It is intellectual, the record of two different minds making sense of the same material.

An automated tool that normalizes that voice does specific damage. It smooths the rough edge that was doing intellectual work. It replaces the precise but unfamiliar term with the common but approximate one. And it breaks the long sentence that was holding a complex qualification together into two shorter sentences that lose the relationship between the clauses. The writing has been replaced.

Preserving voice in the editing process calls for a different kind of attention from correcting errors. It requires an editor who reads the whole manuscript before touching a sentence and understands the argument well enough to know which roughnesses are problems and which are features. That judgment has two sources. One is wide reading in the field; the other is having worked with enough scholarly authors to know the difference between a sentence that needs to be fixed and a sentence that needs to be left alone.

The stakes are higher in scholarly writing than in almost any other genre, because scholarly writing is evaluated by readers who are specifically trained to detect this kind of substitution. A dissertation committee member who has spent thirty years reading work in her field knows immediately when a passage sounds like the author and when it sounds like something else. She may be unable to name what happened. She will feel it anyway — the absence of a mind behind the prose, the slight uncanniness of writing that is correct and not quite inhabited.