Do people feel they own AI-generated text they use?

Draxler et al. name the pattern the AI Ghostwriter Effect: "Users do not consider themselves the owners and authors of AI-generated text but refrain from publicly declaring AI authorship." The effect is defined as "the use of personalized generative AI without credit to the AI," so the model plays the part a human ghostwriter plays. The evidence is two studies, with n1 = 30 and n2 = 96. The second is pre-registered and replicates the effect, adding a comparison with a supposedly human ghostwriter. The abstract reports that personalization of the texts "did not impact" the effect. The result has two halves: a weak sense of ownership, and a choice not to credit the AI in public.

The reasoning separates two things. The sense of ownership is "the subjective side": whether a person feels they are the author of a text. The declaration of authorship is "the entity a text is attributed to, e.g., in the header or byline." H1.1 predicts participants will name the AI, not themselves, as owner, and H1.2 predicts they will not declare AI authorship or support. The analogy rests on social-science work on ghostwriting, which the authors say gives reasons such as financial or political interests, a lack of time and writing expertise, and the pressure for gratification. The abstract also reports a contrast: participants attributed ownership more readily to supposedly human ghostwriters, so the ownership-authorship discrepancy was larger for human ghostwriters. Rationalizations for authorship were similar across the two.

The closest library note, Do users truly own the AI-generated content they produce?, describes a gap between felt and claimed authorship. This excerpt gives that gap an empirical form. The title's phrase "self-declare as authors" suggests the claim is made by omission: by not crediting AI, a user presents the text as their own, though they do not feel they own it. That is a narrower claim than a reflective declaration. It also sits against Do writers want to see each other's AI prompts in shared editors?. Those writers wanted AI use visible, while this paper's default is that AI credit is withheld. Read together, disclosure tools would have to work against that default. That reading is ours, not the authors'.

The excerpt does not give the measures, scales, tests or effect sizes; the abstract states only directions. The study uses GPT-3 alone ("we focus on text generation with GPT-3") in a personal writing context, so it does not show the effect for other models, or for academic and professional authorship, which the authors set apart because prescriptive discussions focus on that non-personalized material. It also does not say how "publicly declaring" was observed. At the strength the evidence allows, the implication is that in personal AI writing the authorship a user declares may diverge from the authorship the user feels, so declaration norms are worth examining. The excerpt does not show how often people actually disclose.