As artificial intelligence begins to write for people, an old question has returned with new urgency: who is the author of a text? Is it the person who directly typed the sentences? The person who provided the idea? The person who instructed the AI? Or the AI system that actually generated the words? Draxler and colleagues’ paper, The AI Ghostwriter Effect, examines this question experimentally from the perspective of human-AI interaction.
The paper’s central finding is clear. Users do not strongly feel that they are the true owners or authors of AI-generated text. Yet when they publish that text, they tend not to identify AI as the author and instead put their own names on it. The researchers call this mismatch the “AI ghostwriter effect.” In simple terms, AI writes like a ghostwriter, but users do not publicly reveal the ghostwriter’s role.
The importance of the paper lies in the fact that it does not confine the debate over AI writing to legal or ethical norms. Instead, it empirically shows how users actually perceive and behave with AI-generated text. Until now, debates about generative AI authorship have largely focused on questions such as whether AI can be recognized as an author, how AI-generated content should be labeled, and who owns the copyright. This paper asks a slightly different question: how do people actually feel about texts written with AI, and how do they present them publicly?
To explain the issue, the researchers distinguish between ownership perception and authorship attribution. Ownership perception is a subjective judgment. It concerns whether users feel that a text belongs to them or to the AI. Authorship attribution, by contrast, is a public act. It concerns whose name appears in the author field of a blog post or publication. The paper shows that these two layers do not necessarily align.
Ghostwriting existed long before the age of AI. Political speeches, celebrity memoirs, business columns, and even some medical papers have long involved cases where the actual writer and the publicly credited author were different. Ghostwriting is the practice in which one party writes a text while another person’s name appears on the final published work. This paper focuses on the possibility that generative AI may popularize that structure. In the past, professional ghostwriters were needed. Now, anyone can ask AI to write for them.
The researchers conducted two empirical studies. The first study involved 30 participants. Participants were given a scenario in which they were ending a trip to New York and writing a postcard to a friend. The researchers presented four different interaction modes. First, participants wrote the text entirely by themselves. Second, they edited a text generated by AI. Third, they selected one of three AI-generated texts. Fourth, they received a single AI-generated text without modification.
These four modes were designed to represent different levels of user influence. Writing the text oneself represented full user control, while receiving the text as-is represented near-total AI control. Editing and selecting fell between those two extremes. For each condition, the researchers measured who participants felt owned the text, whom they listed as the author when publishing it, and how much control and agency they felt they had.
The results of the first study were striking. Participants tended to feel that AI had more ownership than they did over texts involving AI. This was especially true when AI generated the text and users made little or no modification. In that condition, participants strongly perceived the text as belonging to the AI. But the public attribution stage told a different story. When posting the postcard on a blog-style website, most participants still listed their own names as the author. Only about six or seven participants in each AI-related condition mentioned AI as the author.
This mismatch is the AI ghostwriter effect. Internally, users may feel, “I did not really write this.” Externally, however, they present the text as if it were their own. The significance of the paper lies precisely here. The central issue in AI writing is not only how people evaluate AI. It is also how people construct a gap between their internal judgment and their public behavior.
The researchers also examined whether personalized AI changed this effect. Some participants were given AI that was actually adapted to reflect their personal writing style. Others were told that the AI had been personalized, even though it had not actually been adapted. The results showed that personalization did not create a major difference in ownership perception or authorship attribution. In other words, the AI ghostwriter effect had less to do with how personalized the AI actually was and more to do with whether users could connect the AI-generated text to their own intentions.
Control, however, mattered. The more users felt that they had influenced the text, the stronger their sense of ownership became. Compared with simply receiving an AI-written text, editing or selecting the text gave users a stronger sense of control and agency. This connects directly with psychological ownership theory. People are more likely to accept something as their own when they feel that their effort, choices, control, and judgment have been invested in it.
The second study involved 96 participants. Here, the researchers compared an AI ghostwriter with a human ghostwriter. Participants received the same postcard text, but in some cases they were told the text had been written by AI, while in other cases they were told it had been written by a human ghostwriter named “Sasha.” The researchers then examined how people treated AI and human ghostwriters differently.
The results were clear. Participants felt greater ownership over texts written by AI than over texts written by a human ghostwriter. In other words, when a human had written the text on their behalf, participants were more likely to feel, “This is not mine.” But when AI had written the text, they were more likely to treat it as something they could claim as their own. This can be interpreted as evidence that people perceive AI less as an independent author and more as a tool.
Authorship attribution also differed. Participants credited the human ghostwriter more often than the AI. In the second study, the researchers provided explicit authorship options — including AI, GPT-3, the human ghostwriter, and the participant’s own name — so mentions of AI increased compared with the first study. Even so, the human ghostwriter was acknowledged more frequently than AI. People appeared to feel that humans have emotions, effort, intellectual property, and a right to recognition, while AI is closer to a tool such as a word processor or autocomplete system.
This result has important implications for debates over authorship in the age of generative AI. Even when people know that AI generated the actual sentences, they do not treat AI in the same way they treat a human author. AI has contributed, but its contribution is interpreted more as the function of a tool than the contribution of an author. As a result, users may not feel complete ownership over AI-written texts, yet may still feel little resistance to publishing them under their own names.
The paper also connects this issue to the problem of “algorithmic exploitation.” People feel they can use AI more easily than they can use humans. Hiding a human ghostwriter carries an ethical burden, but hiding AI feels less burdensome. This is because people assume AI has no emotions, does not demand copyright, and does not require social recognition. But this attitude creates a new ethical gap. Society has not yet established clear standards for when hiding AI’s contribution is acceptable and when it becomes deception.
The paper also highlights the importance of interface design. In the first study, participants were asked to freely enter the author’s name, and relatively few mentioned AI. In the second study, when options such as AI, GPT-3, the human ghostwriter, and the participant’s own name were explicitly provided, the rate of AI disclosure increased significantly. This means authorship attribution is not determined only by the user’s ethical awareness. The choices offered by the platform, and how easy it is to disclose AI contribution, can shape actual publication behavior.
AI writing tools and publishing platforms should therefore treat authorship disclosure as a design issue. For example, platforms could provide more nuanced labels such as “written with AI assistance,” “AI-generated draft edited by the user,” or “AI-generated based on user input.” A simple binary of either disclosing or hiding AI use does not adequately reflect the complexity of human-AI co-writing.
The paper also suggests that contribution classification systems such as the CRediT taxonomy may need to be expanded for the AI era. In academic publishing, it has become more common to identify different types of contribution, such as conceptualization, data analysis, drafting, review, and supervision. When generative AI enters the writing process, AI’s role could be similarly classified. Did AI suggest ideas? Generate sentences? Change the style? Summarize? Translate? Participate in final editing? These distinctions matter.
The paper is deeply connected to research on psychological ownership. People feel stronger ownership over things they have created, controlled, modified, and selected. Even if AI writes a text, users may feel some degree of ownership if they provided prompts, selected outputs, revised the text, and gave final approval. Conversely, when AI provides a complete text with little user involvement, the user’s sense of ownership decreases. This is a crucial variable in AI co-creation research. The key is not merely whether AI was used, but how substantially the user was involved in the process.
The study also offers a realistic balance for AI ethics discussions. Many institutional debates tend to move toward normative judgments such as “AI-generated content must always be disclosed” or “AI cannot be an author.” But users’ actual judgments are more complex. Users may see AI as a tool while still not feeling that they fully wrote the text themselves. They may also believe AI contribution should be disclosed, but fail to disclose it in specific situations. Ethical rules should be designed with an understanding of these psychological contradictions.
The paper has limitations. First, the writing task was a relatively light personal activity: writing a postcard. Results may differ in higher-stakes contexts such as academic papers, journalism, job applications, business reports, or literary works. Second, the first study had a small sample of 30 participants. The second study expanded the sample to 96, but cultural and linguistic contexts remained limited. Third, the quality of the personalized AI used in the study may differ from today’s state-of-the-art generative AI systems. A GPT-3-based personalization experience in a 2022 experimental setting may not feel the same as current conversational AI tools.
Even with these limitations, the value of the paper is substantial. It shows the authorship problem of AI-generated text not as an abstract debate, but as actual user behavior. Users feel that AI wrote the text, yet they often do not disclose AI. They treat AI less as an independent author and more as a tool. Their sense of ownership increases when they feel greater control and agency. And authorship attribution interfaces can change whether AI contribution is disclosed. These four findings have direct implications for writing policy and service design in the age of generative AI.
The paper’s message is also important for journalism, publishing, academia, and education. Going forward, simply banning or allowing AI use will not be enough. Institutions will need to distinguish at what stage AI was used, how much the user contributed, who bears final responsibility, and what information readers or evaluators have a right to know. Transparency in AI writing cannot be reduced to a single line saying “AI was used.”
The paper raises especially important questions for education. Is it academic misconduct for a student to submit AI-written text under their own name? If the student selected the topic, provided sources, edited the AI draft, and constructed the final argument, how much of the text belongs to the student? Conversely, if AI wrote almost everything and the student merely submitted it, is that ghostwriting? To answer these questions, educators must examine not only the final text, but also the process and the level of control.
The implications for AI co-creation research are equally significant. In human-AI co-created works, self-attribution is not determined simply by the quality of the final product. What matters is how much the user controlled the process, how much they edited, and how strongly they felt their intention was reflected in the outcome. AI creative tools should therefore allow users to intervene, choose, revise, and take responsibility throughout the process. Only then are users more likely to feel that the final output is truly their own.
Ultimately, the core message of the paper is clear. Authorship in the age of generative AI is not only a matter of legal rights. It is a problem involving psychological ownership, social recognition, public attribution, and interface design. People know that AI wrote the text, but they do not treat AI like a human author. In that gap, AI becomes a new kind of ghostwriter.
The task ahead is not simply to decide whether AI should be listed as an author. The more important challenge is to design systems that clearly indicate how AI contributed, what control and responsibility the human had, and what readers need to know. The future of writing is moving toward collaboration between humans and AI. In that future, the question “Who wrote this?” can no longer be answered by a single name.
