Blissful (A)Ignorance: People form overly positive impressions of others based on their written messages, despite wide-scale adoption of Generative AI

As the use of Generative AI (GenAI) tools becomes more prevalent in interpersonal communication, understanding their impact on social perceptions is crucial. According to signaling theory, GenAI may undermine the credibility of social signals conveyed in writing, since it reduces the cost of writing and makes it hard to verify the authenticity of messages. Using a pre-registered large-scale online experiment (N = 647; Prolific), featuring scenarios in a range of communication contexts (personal vs. professional; close others vs. strangers), we explored how senders’ use of GenAI influenced recipients’ impressions of senders, both when GenAI use was known or uncertain. Consistent with past work, we found strong negative effects on social impressions when disclosing that a message was AI-generated, compared to when the same message was human-written. However, under the more realistic condition when potential GenAI use was not explicitly highlighted, recipients did not exhibit any skepticism towards senders, and these “uninformed” impressions were virtually indistinguishable from those of fully human-written messages. Even when we highlighted the potential (but uncertain) use of GenAI, recipients formed overly positive impressions. These results are especially striking given that 46% of our sample admitted having used such tools for writing messages, just within the past two weeks. Our findings put past work in a new light: While social judgments can be substantially affected when GenAI use is explicitly disclosed, this information may not be readily available in more realistic communication settings, making recipients blissfully ignorant about others’ potential use of GenAI.

Introduction. According to a nationally representative survey conducted in August 2024, about 39% of the U.S. population aged 18-64 is already using Generative AI, most commonly ChatGPT, with 10.6% reporting daily usage at work (Bick et al., 2024). As Bick et al. (2024) highlight, the adoption of Generative AI has been quicker than that of personal computers or the Internet. This rapid adoption has been widely documented across age groups, disciplines, and occupations, including but not limited to office workers (Humlum & Vestergaard, 2024), middle and high school students (Zhu et al., 2024), STEM researchers (Van Noorden & Perkel, 2023), humanities researchers (Dedema & Ma, 2024), medical students (Zhang et al., 2024), and medical practitioners (Blease et al., 2024).

In addition to the widespread adoption of Generative AI for personal and professional use, these technologies also have the potential to transform interpersonal interactions.

Generative AI tools empower users with the ability to generate high-quality, complex, personalized, and contextually relevant written content with minimal effort. They have the capacity to enhance the efficiency and quality of social interactions, for example, by boosting the use of positive emotional language (Hohenstein et al., 2023).

At the same time, suspected AI-use may have adverse effects on social relationships (e.g., Glikson & Asscher, 2023; Hohenstein et al., 2023; Lim et al., 2025; Weiss et al., 2022), and instead of building social connections and facilitating interactions, these emerging technologies may erect new barriers to human-to-human cooperation and coordination. Wojtowicz and DeDeo (2025) discuss these negative effects by highlighting how Generative AI may undermine “mental proofs.” Mental proofs are observable actions that allow audiences to make inferences about the communicator’s hidden mental states such as their intentions, goals, and values. As Wojtowich and DeDeo (2025) argue, outsourcing the laborious process of writing to an algorithm—thereby making social interactions more efficient and less effortful—may lead to the paradoxical consequence of undermining trust and coordination between people, because it weakens the link between what people communicate and what people actually think, feel, or want.

But are people actually becoming more skeptical towards others when judging them based on their written messages? Do people respond to the increasing prevalence of Generative AI use by evaluating written content more critically, or do they remain blissfully ignorant to the RUNNING HEAD: BLISSFUL (A)IGNORANCE possibility that messages may be AI-generated? How does (the lack of) information about AI use affect social impressions, especially in more realistic settings when there is a fundamental uncertainty about others’ use of AI? In the present paper, we seek answers to these questions and experimentally investigate the current “state of skepticism” in interpersonal communication.

There are, however, at least two crucial differences between the stylized interactions studied in this past line of research and real situations. First, people in real situations almost never know with certainty whether (and to what extent) a communicator has used Generative AI.

Assuming that communicators are aware of the potential negative consequences of disclosing their use of AI, they are motivated to conceal all AI involvement. This rational strategy of concealing AI use, combined with the difficulty of detecting and distinguishing AI-generated content (Jakesch et al., 2023; Köbis & Mossink, 2021; Kreps et al., 2022), forces audiences to rely on probabilistic inferences instead. That is, in most situations, people must form impressions of others and make decisions under the fundamental uncertainty of not knowing whether the communicator has used Generative AI, let alone, to what extent.

One goal of this paper is to study such judgments under uncertainty: when people are aware of the possibility that a message might have been generated by AI but have no way of knowing this. According to a simple decision-theoretical model, social impressions under uncertainty, IU, are the probability-weighted combination of impressions under certainty: where IH is the social impression of someone who is known to have written a message entirely on their own, IAI is the impression of someone who is known to have generated a message entirely by using AI, while pH and pAI are the corresponding probabilities (pAI= 1 −pH).1 Assuming that IH> IAI (aligned with prior empirical evidence), the above model implies that social impressions would turn increasingly more negative (positive) as people would become more (less) suspicious of others, as a function of pAI.

1 We discuss the validity and plausibility of these assumptions in Section 4.2.

Related work. 1.1. Signaling theory: writing as a source of credible social signals Written communication is a rich source of social signals, as every choice a writer makes—words, tone, and sentence structure—reflects the underlying characteristics of the individual (e.g., Mairesse et al., 2007; Pennebaker et al., 2003). These signals enable readers to assess the personal traits of the writer, such as trustworthiness, thoughtfulness, diligence, and kindness, or conversely, deceitfulness, superficiality, laziness, and antagonism. Such inferences allow people to decide whether to cooperate, compete, approach, or avoid others. For example, cover letters help employers choose which candidates to interview; dating profiles let singles decide whether they want to meet a potential date; apology letters may determine whether someone is forgiving or holding onto a grudge; and academic publications may serve as the primary avenue for researchers to showcase their intellect, creativity, and effort.

However, signals can be faked, especially when communicators have a strategic interest in deceiving or manipulating their audience. A job candidate may ask a seasoned colleague to write a cover letter on their behalf; an apology letter may be based on a template; and an academic publication may plagiarize past work. According to signaling theory (see, e.g., Connelly et al., 2011; Spence 1973, 2002), the usefulness of signals—any signal, not just in written communication—critically depends on the credibility of signaling, that is, how likely a signal is honest or fake. A core prediction of signaling theory is that as the cost of signaling increases—which can be time, effort, emotional cost, or even reputation—it becomes more credible (Spence 1973, 2002). Therefore, observers are constantly monitoring costs to gauge the credibility of signals (Gintis et al., 2001; also see “strategic vigilance”, Heintz et al., 2016). As Chaudhry & Wald (2022) highlight, three qualities can make an observable signal to be perceived as costly: 1) it is difficult-to-fake; 2) verifiable; and 3) self-sacrificing. For example, a handwritten note of apology is a more credible signal of true remorse than a text message, as hand-writing is more difficult-to-fake, easier to verify, and takes more effort and time to produce (i.e., requires more “self-sacrifice”) than texting someone.

1.2. How Generative AI may undermine the social signaling function of written messages Generative AI may undermine all three of these qualities of written signals, substantially reducing their perceived cost and credibility.

Method. 2.3. Procedure We directed participants to a Qualtrics survey titled “Social Perceptions in Interpersonal Communication” and asked them to imagine a hypothetical scenario. Crucially, we did not tell participants in the instructions (or during recruitment) that this study would be about Generative AI to avoid prompting them to think about this technology. In the scenario, we presented participants with a brief background story and then told them that another person, Alex, had sent them an email. To capture a broad range of potential situations in which someone may be judged based on their email, we implemented a stimulus sampling method (see Wells & Windschitl, 1999). We randomly assigned participants to imagine one of the following four scenarios:

1. receiving a gratitude message from a close friend (“gratitude”); 2) receiving an application from a candidate who is applying for a nanny position (“nanny”); 3) receiving a cover letter from a job applicant for a data analyst position (“cover letter”); or 4) receiving feedback on a marketing project (“feedback”). full text of these scenarios in the Appendix. Across the scenarios, we varied whether the sender, Alex, was someone who is socially close to the participant (i.e., their friend in the gratitude scenario or their colleague in the feedback scenario) or socially distant (i.e., a stranger in the nanny and cover letter scenarios), and whether the context was more personal (i.e., gratitude and nanny scenarios) as opposed to professional (i.e., cover letter and feedback scenarios).

After participants read the message sent by Alex, we implemented the main experimental manipulation: Across four conditions, we varied what participants knew about how Alex created the email. In the “no information” condition, that we designed to most closely resemble naturalistic communication settings, we didn’t provide any further information, and participants simply proceeded to evaluate Alex based on their message.

In the other three conditions, we first asked participants to read a brief description of Generative AI chatbots—what these tools are and what they can be used for (see the Appendix)—then we informed participants whether Alex used a Generative AI chatbot to create their message.

In the “human” condition, we told participants that “Alex has written their message without using a Generative AI Chatbot. Alex wrote the text on their own, word by word.”

In the “AI” condition, participants learned that “Alex has generated their entire message by using a Generative AI Chatbot. Alex didn’t change or modify the AI-generated text at all.”

Finally, in the “uncertain” condition, we asked participants to imagine that they “are uncertain: whether Alex has generated their message by using a Generative AI Chatbot, or whether Alex has written their message on their own. The text could be entirely generated by a Generative AI Chatbot, or it could be completely written by Alex, word by word.”

Each participant was randomly assigned to one of the four above conditions, thus our experiment had a full-factorial 4x4 design, with two factors (type of scenario and type of information about AI use) that we manipulated between participants, resulting in 16 unique scenario-information combinations.

Discussion. Consistent with earlier work (Glikson & Asscher, 2023; Hohenstein et al., 2023; Lim et al., 2025; Weiss et al., 2022), we found strong negative effects on social impressions of the sender when disclosing that a message was generated by AI, compared to when the same message was believed to be written by a human. We detected these effects across four distinct communication scenarios (personal vs. professional; close others vs. strangers) and in both participants’ explicit numeric ratings of the sender (i.e., overall social impression scale) and their implicit sentiments towards the sender (i.e., net valence of their open-ended impressions).

However, when the involvement of Generative AI was less than certain, participants formed overly positive impressions of the sender. In particular, when we did not even highlight the possibility that the message was generated by AI, participants’ impressions of the sender were virtually indistinguishable from the impressions they formed when they knew that the message was fully written by a human. Interestingly, we found significant differences in the explicit ratings between the uninformed and uncertain conditions, which suggests that participants, by default, did not even consider the possibility of AI involvement, but adjusted their impressions once this possibility was highlighted. Crucially, we did not tell participants any definite information about the origin of the message in either of these conditions, so the observed differences can only be explained by different levels of attention to AI involvement. Similarly to earlier work on the link between social judgments of behavior and attention to others’ vaccination status during the COVID-19 pandemic (Molnar et al., 2023), here we document that social judgments can be significantly affected by shifting people’s attention to decision-relevant information that otherwise does not readily come to mind.

At the same time, even when we prompted participants to consider the possibility that a message was fully generated by AI, the resulting social impressions were much closer to those of human-written messages, as opposed to AI-generated messages. In addition, participants seem to underestimate the prevalence of Generative AI chatbot use, especially among their close peers, even though almost half of the current study sample admitted using such tools for writing messages to others, just within the past two weeks. In the following sections, we discuss some potential alternative explanations for these results. Exploring which of these mechanisms is responsible for the main effects observed in our experiment is beyond the scope of the current paper but could be a suitable topic for future research.

RUNNING HEAD: BLISSFUL (A)IGNORANCE

Conclusion. We found that despite the relatively widespread use of Generative AI in writing or editing messages, participants did not exhibit any skepticism towards senders under naturalistic evaluation conditions, when the possibility of AI use was not explicitly highlighted. Under these realistic conditions, participants’ impressions of senders were virtually indistinguishable from impressions of senders who were known to have written their messages on their own, without using Generative AI. Even when we highlighted the potential use of Generative AI, participants formed overly positive impressions of senders, especially in their open-ended first impressions.

Notably, when participants were certain that the sender used Generative AI, their impressions were substantially more critical than in any other condition, so the overly positive judgments in the uninformed and uncertain conditions cannot be explained by participants’ overall indifference towards the use of Generative AI. Instead, our results suggest that while judgments can be substantially affected by information about Generative AI use, this information is not readily available and not top of mind in more realistic communication settings.

These findings put existing work in a new light by challenging the prevailing assumptions about the negative impact of AI on social perceptions. Previous studies have documented adverse effects of AI involvement in communication, such as reduced authenticity and sincerity in apologies and other messages. However, our research suggests that when the source of a message is uncertain, individuals may overlook the potential downsides of AI and default to more favorable evaluations. This indicates a nuanced understanding of how uncertainty and attention to AI involvement can shape social judgments, highlighting the need for further exploration into the conditions under which skepticism may arise.

Looking ahead, it is essential to consider how these dynamics may evolve over time. As awareness of Generative AI’s capabilities and its implications for communication increases, individuals may begin to adopt a more skeptical perspective.