Understanding Reader Perception Shifts upon Disclosure of AI Authorship

As AI writing support becomes ubiquitous, the question of how disclosing its use affects reader perception remains critical and underexplored. We conducted a controlled study with 261 participants to examine how disclosing varying levels of AI involvement shifts perceptions of the author across six distinct communicative acts. Our analysis of 990 evaluations reveals that disclosure generally erodes perceived trustworthiness, caring, competence, and likability, with the most precipitous declines observed in social and interpersonal writing. A thematic analysis of participant feedback attributes these negative shifts to a perceived loss of human sincerity, diminished authorial effort, and the contextual inappropriateness of AI. Notably, however, we find that higher AI literacy mitigates these negative perceptions, leading to greater tolerance or even appreciation for AI assistance. Our results highlight the nuanced social dynamics of AI-mediated authorship and inform design implications for transparent, context-sensitive writing systems that better preserve trust and authenticity.

Introduction. Artificial intelligence (AI) systems have reached a level of writing fluency comparable to human authors leading to their widespread adoption across professional, creative, and interpersonal contexts. While these tools can enhance clarity and expression, they also introduce significant social frictions. Readers often perceive AI-generated texts as less empathetic or authentic [49], and disclosing AI authorship can degrade perceptions of authenticity and trust [22, 27, 52]. This tension arises because writing is more than mere information transfer. It is a social act that conveys an author’s intention, effort, and emotion to readers. When an AI actively contributes to the writing process, this fundamental social meaning of authorship becomes ambiguous, complicating how a text and its author are perceived. While previous work in AI-mediated communication has outlined these ethical and relational challenges [16], empirical studies have so far examined reader perceptions with limited granularity or control over AI authorship. Much of the existing research compares purely manual writing to AI-assisted text in specific, narrow contexts, often treating writing as a monolithic activity [21, 27, 52]. This overlooks the nuanced nature of writing, which serves a wide spectrum of communicative goals. As articulated in frameworks like Berge et al.’s Wheel of Writing, writing can be used to persuade, to interact socially, to reflect internally, or to imagine creatively [4]. The social expectations for an apology letter, for example, are vastly different from those for a technical manual. As AI writing support becomes deeply embedded in these varied contexts, the boundary between helpful assistance and core authorship blurs. Furthermore, readers’ attitudes toward writing co-authored by AI may vary depending on their familiarity with the technology. Previous studies have indicated that individuals who are more knowledgeable about AI tend to view its use as a pragmatic choice rather than a lack of competence [29]. This raises critical, unanswered questions: How do readers’ perceptions shift based on the degree of AI authorship across these different communication purposes? Furthermore, how do individual factors, such as readers’ AI literacy, moderate these perceptions? Addressing these questions is essential for designing AI authoring tools that are not only effective but also socially acceptable and trustworthy. To address this research gap, we conducted a controlled user study with 261 participants that examines how readers’ perceptions of a fictitious author change upon the disclosure of AI authorship across six distinct acts of writing [4]. Our analysis revealed consistent negative shifts in perceived trustworthiness, caring, competence, and likability, especially in emotional and interpersonal writing. However, we also found that participants with higher AI literacy exhibited greater tolerance toward, and even appreciation of, AI authorship. Our results highlight the complex social dynamics of AI authorship and inform the design of transparent, context-sensitive writing tools that preserve authenticity and trust. We position AI writing assistance not merely as a technical problem of automation, but as a relational design challenge. We argue that future interfaces must balance AI capabilities with the need to communicate human authorship, social context, and the perceived effort invested by the author. In summary, we make the following contributions to research on AI-mediated writing and human–AI interaction:

• Granular empirical evidence on perception shifts after AI authorship disclosure: We present findings from a controlled study (N=261) that systematically investigates how reader impressions of an author change after disclosing varied levels of AI authorship across six different acts of writing. Our results reveal consistent negative perception shifts in trust, caring, competence, and likability, with stronger effects observed in interpersonal contexts. • Insights into the social dynamics of AI-mediated writing: Our quantitative and qualitative analyses show that perceptions are shaped not only by the degree of AI authorship but also by perceived human effort, emotional context, and readers’ AI literacy. These factors provide a foundation for designing more socially aware AI authoring systems. • Design implications for future AI-mediated writing: Based on our findings, we derive four design implications for future systems: 1) context-sensitive transparency that explains the purpose of AI authorship; 2) interfaces that preserve and communicate human effort and agency; 3) writing support adaptive to emotional and social contexts; and 4) reflective interfaces that foster AI literacy and calibrated trust.

Related work. We situate our work within two primary streams of research: the application of AI across the diverse functional purposes of writing, and the growing body of literature on how readers perceive and react to the disclosure of AI authorship in communication.

Writing serves a wide spectrum of goals, from the interpersonal to the informational. Berge et al. [4] outline six fundamental purposes: persuading audiences, interacting with others, reflecting on experiences, describing facts, exploring ideas, and creating imaginative worlds. Modern AI-assisted writing tools are increasingly used across all of these distinct cognitive and social functions. Prior research has explored how AI tools support a range of writing goals. Mirowski et al. demonstrated that AI can be integrated into writers’ workflows to enhance creativity in screen- and playwriting [34]. AI has also been leveraged in contexts where writing serves an interpersonal or social function. For instance, systems that suggest, generate, or refine email responses can help users convey positivity [33], politeness, and professionalism in workplace interactions [35], while AI-assisted dating profiles can enhance self-presentation [3]. These applications demonstrate that the role of AI extends far beyond grammatical correction to shaping a writer’s tone, style, and communicative intent, thereby aiding in both persuasion and social engagement [16]. This expanding role, however, introduces social friction, particularly when AI’s contribution blurs the boundaries of authorship. Readers often question the authenticity and credibility of AI-generated content. For instance, Li et al. found that revealing the role of AI in essay writing lowered quality ratings from readers [26]. Similarly, research on AI-mediated communication showed that online profiles perceived as AI-generated are less trusted [22]. These concerns are particularly salient when writing serves a social purpose, such as persuasion or interaction, where impressions of the author are heavily shaped by perceived authenticity. Recent studies on disclosure dynamics confirm this tension: while AI assistance can improve a text’s fluency, revealing its use can lead readers to judge the author as less sincere or capable [12, 19]. AI-assisted writing thus presents a fundamental trade-off. On one hand, it empowers individuals, especially those with language barriers or weaker writing skills, to communicate more effectively [38, 39]. On the other, the social meaning of authorship complicates its adoption. When readers know or suspect AI involvement, they may perceive the writing—and by extension, the author—as less authentic or trustworthy [22].

Method. To investigate how the disclosure of AI assistance influences reader perception, we conducted a repeated-measures online study. Participants rated their impressions of a text before and after being informed that a specific percentage of the content was generated or edited by AI. Unlike prior work, our primary objective was to examine how perceptions shift across systematically varying levels of disclosed AI contribution. To achieve this, we used a deception-based design: while all texts were entirely AI-generated (with minor human proofreading), participants were told that only specific portions were created or altered by AI. This approach enabled precise control over the perceived degree of AI involvement. Our study addressed the following research questions:

RQ1 How does the communicative purpose (i.e., the act of writing) of a text shape reader perceptions upon the disclosure of AI assistance? RQ2 How does a reader’s AI literacy moderate these perception shifts? RQ3 What qualitative reasons and underlying themes explain the changes in perception following AI disclosure?

The following study protocol was approved by the Institutional Review Board of the first author’s university.

We grounded our experimental design in the six fundamental acts of writing defined by Berge et al. [4]: Convince, Interact, Reflect, Describe, Explore, and Imagine. As detailed in Table 1, these acts serve distinct communicative purposes. Convince, Interact, and Reflect are primarily person-oriented, focusing on others, social relationships, and the self, respectively. In contrast, Describe, Explore, and Imagine are object-oriented, centering on the organization, development, and creation of knowledge.

To create our experimental stimuli, we generated three scenarios for each act, yielding 18 unique texts (Table 2). We created 3–5 initial scenarios per act using GPT-4o1 by providing the model with the definition of the act. We then selected the three scenarios we found the most suitable and prompted GPT-4o to generate sentences based on their descriptions. As the study targeted Japanese participants, all texts were presented in Japanese. Pilot testing revealed that sentences generated directly in Japanese by GPT-4o lacked natural fluency. Therefore, we first generated the text in English and then translated it into Japanese. A native Japanese-speaking author reviewed all translated texts, making minor corrections to ensure linguistic quality and naturalness. The final texts averaged 20 sentences and 662 Japanese characters in length. The full English and Japanese scripts are provided as supplementary material. To manipulate the perceived amount of AI involvement, we simulated various levels of AI authorship. For each text, a random proportion of sentences was labeled as generated or edited by AI (in reality, all texts were entirely AI-authored). The distribution of these AI-labeled sentences was randomized using a seed unique to each condition. The disclosed proportion varied from 0% to 100% in 10% increments across conditions, corresponding to approximately two sentences per increment (Figure 1). This design allowed us to model situations where a reader discovers that some portions of a text, from minor local edits to the entire document, were created with AI assistance.

The experimental workflow consisted of three main parts for each text evaluated by participants.

3.2.1 Pre-disclosure Evaluation. After providing informed consent, participants were each assigned four of the 18 texts, representing different acts of writing, and were asked to read them carefully. For each text, they answered a questionnaire assessing their initial impressions of the (fictitious) author (Figure 1a). The questionnaire assessed five perception dimensions using 7-point Likert scales adapted from established instruments (27 statements in total). Trustworthiness (6 statements), Caring/Goodwill (6 statements), and Competence (6 statements) were adapted from McCroskey and Teven [32], and responses ranged from -3 (very negative) to +3 (very positive).

Discussion. Our findings provide a nuanced understanding of how readers’ perceptions of an author shift upon the disclosure of AI authorship. We distill our results through the lens of our three research questions, discussing the significance of how communicative purpose, AI literacy, and perceived authorial effort shape reader judgments.

Our results reveal a sharp distinction in how readers perceive AI authorship depending on the communicative purpose of the text (Table 3). In acts of writing that are object-oriented, such as persuading (Convince), creating narratives (Imagine), or developing knowledge (Explore), AI involvement was viewed more favorably. This task-dependent favoritism aligns with insights from Castelo et al. [10], who found that objective tasks, such as predicting stock market outcomes, were more readily accepted than subjective tasks like romantic partner recommendation. However, AI generation may still be discouraged in objective tasks where accuracy is paramount, particularly in forecasting contexts [11]. The positive association with perceived competence in Convince and Imagine suggests that readers may appreciate AI as a tool that enhances the quality of argumentation or creative expression. This is supported by prior work suggesting that readers value AI’s contribution when it serves to consolidate or develop an author’s ideas [19]. In stark contrast, the use of AI in the person-oriented act of Interact elicited strong negative reactions. Our qualitative themes confirm that in such contexts, AI-authored text was perceived as “cold”, “impersonal” (Theme N1), and even “insincere” (Theme N4). This negative perception is consistent with prior work showing that AI disclosure can signal a lack of author engagement and thereby erode credibility [21, 22, 52]. These findings underscore the context sensitivity of algorithm aversion. As noted by Morewedge [36], algorithm aversion is often triggered by tasks lacking established evaluation criteria (e.g., creative work) or those dependent on individual subjectivity (e.g., gift recommendations). The Interact act embodies both characteristics, and our findings confirm it to be the act of writing most sensitive to AI disclosure. This also highlights the importance of the “communicative relationship” in interpersonal communication [8]. Because AI is perceived as incapable of genuine empathetic engagement [27, 49], its use in relational contexts can be interpreted as a violation of social expectations, making human attribution critical. Interestingly, in the Convince and Interact act, AI use boosted perceived competence, suggesting that when AI’s role is to strengthen an argument and leverage its perceived objectivity [40], its use can be seen as strategic rather than socially detrimental.

Two factors emerged as key determinants of the magnitude of perception shifts: the disclosed level of AI contribution and the reader’s AI literacy. As confirmed by our regression analysis, a higher disclosed AIRatio consistently led to more negative perceptions of the author. Our qualitative themes provide the reasoning behind this trend: as AI’s role grew, participants were more likely to perceive a loss of human touch (Theme N1), question the author’s expertise (Theme N3), and infer a lack of effort or agency (Theme N5). This echoes recent work suggesting that dominant AI authorship can lower an author’s perceived credibility and ownership [12, 23]. Our results also confirm that a reader’s AI literacy can significantly mitigate these negative effects. Participants with higher self-reported AI literacy exhibited smaller negative perception shifts and, in some cases, expressed positive attitudes toward AI’s capabilities (Theme P8).

Conclusion. We examined how the disclosure of AI authorship in writing influences readers’ perceptions across different acts of writing. Our study with 261 participants revealed that disclosure generally erodes perceptions of trustworthiness, caring, competence, and likability, particularly within social and emotional contexts. However, these negative effects are moderated by the reader’s AI literacy and the degree to which human effort and agency remain visible. Our findings demonstrate that the social acceptability of AI-mediated writing hinges not merely on textual quality, but on how human intentionality is communicated and perceived. Based on these insights, we propose design implications for AI writing tools that prioritize context-sensitive transparency, the preservation of human effort, and the cultivation of AI literacy to sustain authenticity. Our work advocates for re-imagining AI writing assistance as a co-creative partnership and relational process where technology amplifies, rather than replaces, human expression and empathy. As AI becomes increasingly embedded in our communicative practices, future research must continue to explore these dynamics across diverse cultural contexts and longitudinal timelines.

Limitations. Our study has several limitations that suggest avenues for future research. First, we manipulated the perceived level of AI authorship using a quantitative ratio of sentences. While systematic, this does not account for the semantic weight of those sentences. For instance, a sentence articulating a core thesis carries more weight than one providing a minor detail; consequently, the perceived influence of AI likely varied depending on which specific sentences were highlighted. Future work should explore how reader perceptions are affected when AI contributes to different functional parts of a text, such as core arguments versus stylistic refinements. Second, our simulation of AI disclosure simplifies the complexity of real-world human-AI writing collaborations.