An opinionated writing assistant shifted what people believed

The writer's problem

A writer using an AI drafting assistant to work through an argument assumes the tool is neutral scaffolding. A CHI 2023 paper by Maurice Jakesch, Advait Bhat, Daniel Buschek, Lior Zalmanson and Mor Naaman, posted as a preprint and later published in the CHI proceedings, tested whether a writing assistant that leans toward one side of an argument can shift what the person using it ends up believing, not just what they type.

What the documents show

The paper's own methods describe an online experiment in which 1,506 participants wrote a post on whether social media is good for society, assisted by a language-model writing tool configured, for two of three groups, to argue consistently that social media is good or that it is bad. Independent judges, 500 in total, then labeled the opinions expressed in participants' sentences. The results state that participants assisted by the pro-social-media model argued social media was good for society in 45 percent of labeled sentences versus 28 percent in an unassisted control group, and the pro-critical model shifted the mix the other way; a later attitude survey found participants' own stated opinions had moved in the same direction as the model they used. The official CHI proceedings record confirms this as the peer-reviewed published version of the same study, presented at the 2023 conference, with no different headline finding from the preprint.

The editorial choice

An editor helping a writer choose a drafting tool has no way, from this paper alone, to audit any specific commercial assistant's own slant; the study built a deliberately one-sided experimental tool rather than testing a named commercial product. What the paper does support editorially is treating an assistant's suggested phrasing on a contested question as a possible source of drift, worth checking against the writer's own prior notes or outline rather than accepted as a shortcut past having a position.

What stays with the author

The study measured aggregate shifts across many participants completing a single short task, not any one writer's final judgment on a topic they have researched over time. Recognizing when an assistant's phrasing is doing argumentative work, and deciding whether to keep, cut or contest it, remains the writer's task; the paper does not identify a mechanical fix.

• Would a longer piece, revised over multiple sessions, resist this drift more than a one-shot post?
• Does a specific tool a writer already uses have a documented tendency on contested topics?
• What would change if the writer stated a position before drafting, rather than while drafting?

The finding is about influence on a controlled, short-form task; it does not establish that every AI writing tool leans a particular way, only that one built to do so measurably worked.