Workslop: Examining the prevalence, antecedents and consequences of low-quality AI-generated content at work

Source: PsyArXiv (Liebscher, Lee, Rapuano, Kellerman, Niederhoffer, Hancock; BetterUp Labs / Stanford Social Media Lab) · 2026-02-05

Abstract (218 words) Artificial intelligence (AI) adoption in the workplace has skyrocketed but produced ambiguous effects, raising questions of when AI use may be beneficial or harmful to productivity and collaboration. To understand the conditions under which AI workplace use is harmful, we examined workers’ experiences with AI-generated workslop: low-quality, predominantly AI-generated content that appears to fulfill a given task, but lacks the substance to meaningfully advance the work at hand. We conducted a representative survey of American full-time desk-workers (n = 962) that found that 52.7% reported sending some amount of AI-generated workslop to others. This behavior was more common among people who had high trust in AI, low agency over AI use, organizational encouragement to use AI, and low psychological safety at work. A substantial proportion (38%) also reported receiving work from their colleagues or teammates that they would consider to be workslop, and that the effort involved in processing, discussing, and revising - or even redoing - the work cost them around 3.4 hours per month on average. In addition to consuming workers’ time, receiving workslop negatively affected their mood and harmed their perceptions of their colleagues as trustworthy, reliable collaborators. Together, our findings paint a cautionary tale regarding AI integration into the workplace by highlighting how using AI under certain circumstances can be counterproductive and undermine interpersonal relationships.

Introduction. Advances in artificial intelligence (AI) appear poised to reshape the future of work (Selenko et al., 2022; Anthony et al., 2023; Brynjolfsson et al., 2025). Workers across nearly every industry increasingly turn to tools powered by large-language models (LLMs) like ChatGPT and Gemini to write emails, generate code, process data, and make decisions (Maslej et al., 2025). A recent national survey found that 28% of employed American adults report using ChatGPT for work (Sidoti & McClain, 2025), with a global study finding that 30% of consumer ChatGPT use is for work-related inquiries (Chatterji et al., 2025). Executives at major companies, schools, and healthcare institutions have also moved quickly to encourage their organizations to broadly adopt AI tools in hopes of augmenting worker productivity (Box, 2025; Oliver & Dabney, 2025; Lenharo, 2023).

Continued research on the effects of AI adoption, however, has tempered this initial excitement by demonstrating how workplace AI use can be both beneficial and harmful. While some AI tools have been shown to improve work outcomes by automating routine tasks or polishing written work (Noy & Zhang, 2023; Wu et al., 2025; Brynjolfsson et al., 2025), many fail to help people make meaningful progress on the tasks at hand. For instance, the MIT Media Lab found that fewer than 5% of public AI initiatives produced a “marked and sustained” impact on productivity and profitability (Challapally et al., 2025). Similarly, a longitudinal study of over 7,000 workplaces in Denmark found that implementing AI chatbots did not affect organizational outcomes (Humlum & Vestergaard, 2025). At the same time, there are growing complaints among everyday workers who are frustrated with receiving poor-quality AI-generated work from their colleagues (Hamilton, 2025).

Understanding the discrepancy between the anticipated benefits of AI and the observed mixed results in the workplace is imperative to informing decisions about engaging with AI tools at both the individual and organizational level. Previous studies have investigated this tension by looking at differences in AI adoption across industries (Maslej et al., 2025), shifting perceptions of AI among the public (Kennedy et al., 2023; Cheng et al., 2025), and frontier models’ performance on common work tasks (Doshi & Hauser, 2024; Chang et al., 2025). However, growing research on human-centered AI emphasizes how the effects of AI use are shaped not only by the tools’ technical abilities, but also by how people choose to use them in the context of their own lives (Shneiderman, 2022; Capel & Brereton, 2023; Hancock et al., 2020).

Specifically, workplace AI use may fail to improve productivity when people use these tools to take shortcuts. While unmanageable workloads and low motivation have long caused people to cut corners at work (Beck et al., 2017; Jonason & O’Connor, 2017; Ansari et al., 2024), the specific affordances of AI now allow people to produce plausible-looking work with less effort (i.e. a single prompt) - though the final product may not contain the substance or context to achieve the goals of the task (Cowgill et al., 2024; Herbold et al., 2023). Colloquially, frustrations with AI-generated products are known as slop, which references “the flood of lazy, derivative, endlessly recycled outputs” produced by LLMs (Madsen & Puyt, 2025). Although often used to describe mild annoyances - such as redundant email templates, sycophantic text, or poorly rendered images - such content could have serious consequences when shared in the workplace.

We propose that one reason why workplace AI use may fail to improve productivity is because people sometimes use these tools to generate workslop: predominantly AI-generated content that fulfills the appearance of completing a task at work, but is perceived by the recipient as low-effort, low-quality, and/or not meaningfully advancing the task. We conducted an online survey with a sample of American full-time desk-workers across diverse industries including technology, education, health, and marketing (N = 962). The survey examined the prevalence of encountering AI-generated workslop, the dispositional and situational factors that were associated with sending workslop, and the effects of receiving workslop on their affect, productivity, and perceptions of their colleagues. Furthermore, we elicited and analyzed 362 in-depth descriptions of specific instances where individuals received work from their coworkers that they believed to be workslop. We addressed the following research questions:

●​ RQ1: What is the prevalence of American desk-workers who send and receive AI-generated workslop? ●​ RQ2: What makes people perceive work to be AI-generated workslop? ●​ RQ3: What are the psychological and structural predictors of individuals a) sending AI-generated workslop to their colleagues? and b) of sending greater amounts of workslop to their colleagues?

Related work. The introduction of new technologies like email, instant messaging, and video-conferencing have repeatedly transformed how people communicate and collaborate in the workplace (Skovholt & Svennevig, 2006; Isaacs et al., 2002; Fauville et al., 2023; Colbert et al., 2016). While these tools enabled more flexible work across time and space, they also introduced many new challenges. Recent research on human-AI interaction suggests that the adoption of AI tools in the workplace is likely to follow a similar trajectory of producing both new efficiencies and new sources of friction (Brynjolfsson et al., 2025).

The use of AI-based tools such as ChatGPT to generate, alter, and share work can be understood through the framework of AI-mediated communication, which examines how the dynamics of information exchange are altered when AI systems intervene in human-to-human interaction (Hancock et al., 2020). This framework proposes that AI systems function as interlocutors that “modify, augment, or even generate” messages on behalf of a sender, to be interpreted by receivers who may or may not be aware of the AI’s involvement. This mediation of the traditional communication process alters both the characteristics of the content created and the social dynamics of the people involved (Jakesch et al., 2019; Hohenstein & Jung, 2020). For example, past work has found that “smart replies” that provide personalized recommendations for emails are often biased towards short, positive responses (“That’s great!”). When people rely on these smart replies instead of writing their own, they are perceived as being more cooperative by others, unless they are suspected as using AI, in which case they are evaluated as less trustworthy (Hohenstein & Jung, 2020). Similarly, more recent studies have found that content produced by LLM-based chatbots has some shared, distinctive characteristics. Content written by models like ChatGPT, Gemini, and Claude often contains certain punctuation (e.g., em-dashes), words and phrasings (e.g., “delve”, “not only X, but also Y”; see Juzek & Ward, 2025), and tones that are often perceived as sycophantic or robotic (Chen et al., 2025; Cheng et al., 2025).

Workslop as a harmful form of AI-mediated communication in the workplace. To date, hundreds of studies have examined how co-producing content with AI can affect psychological, interpersonal, and organizational outcomes. A clear theme across this literature is heterogeneity: the effects of AI depend on how people take up these systems within the practical constraints of their everyday tasks (Shneiderman, 2022; Capel & Brereton, 2023; Hancock et al., 2020; Markowitz, 2024).

Method. Pilot study ​ A pilot study invited full-time American workers from the survey platform CloudResearch’s Connect platform (N = 150) to describe their thoughts about workplace AI use (see Appendix 1 for details). After being presented with a brief definition of workslop, 40.3% of participants believed they had This paper has not yet been peer-reviewed.

In September 2025, an additional 1,015 full-time American workers were recruited from CloudResearch’s Connect platform. Sample size was determined using the pilot study to achieve a margin of error of ±5 percentage points for our primary prevalence estimate at 95% confidence, requiring approximately 370 participants. We oversampled for additional precision. To be eligible, participants were required to be over 18 years old, currently employed in a full-time role, fluent in English, and located in the United States. Participants had to have a CloudResearch task approval rating of over 90% on more than 100 tasks. To represent a population of desk workers, participants were only eligible if they classified their work arrangement as white collar, professional, creative, IT, or public sector or government. An equal balance of men and women was requested. All procedures were approved by the Salus Institutional Review Board and participants were compensated with $3.00 upon completion ($12/hr). Upon consenting, participants were asked to answer some questions about themselves and their workplace. For a full list of all survey measures used, please see Appendix 2. Data and code are available at https://osf.io/xh2wn/overview?view_only=675432f416474ce592257dfd684343e3. Participants Eleven participants were excluded for failing one or more of three attention checks. A manual review of open-ended responses further excluded 42 participants for misunderstanding the primary workslop description question (i.e., wrote nonsense, described an instance of high-quality AI-generated work). The final analytical sample included 962 participants. There were approximately equal numbers of men (n = 480, 49.9%) and women (n = 482, 50.1%). The majority were White (n = 737, 76.6%), 7 were American Indian or Alaskan Native (0.7%), 72 were Asian (7.5%), 122 were Black or African American (12.7%), and 24 were another ethnicity or preferred not to say (2.5%). The mean age was 40 years old (SD = 10.3 years). A subset of 71 were Hispanic or Latino (7.4%), with the majority not being Hispanic or Latino (n = 891, 92.6%). The sample was relatively highly educated, with 756 indicating they held a Bachelor’s degree or higher (78.6%). ​ Approximately half of the participants (55.6%) were individual contributors (n = 535), with 337 who were frontline managers (35.0%) and 90 who were managers-of-managers (9.4%). For the analyses below, these two categories of managers were aggregated into one. Mean work experience was 18.4 years (SD = 10.2 years) and participants worked regularly with a mean of 12 other people (median = 8, SD = 19.2). Participants varied in the degree in which they worked in-person vs. remotely, with 300 working in-person all of the time (31.2%), 201 working fully remotely (20.9%), and 461 working some degree of hybrid (47.9%). The mean individual income was $74,655 (median = $50,000, SD = $48,633, range = $10,000 – $300,000). Participants came from a variety of industries, with the top three being Technology (18.2%), Government / Public Sector (17.9%), and Healthcare (16.2%) (see Appendix 3 for a full table). Measures See Appendix 2 for detailed descriptions of all measures administered, including response options and the full set of items. Validated scales were used whenever possible and Cronbach’s α’s were used to assess reliability. The term “workslop” was not used anywhere in the survey.

Individual differences in psychological traits. Participants completed the Ten Item Personality Inventory (Gosling et al., 2003), a 10-item scale assessing openness to new experiences (Cronbach’s α = .53), conscientiousness (α = .64), extraversion (α = .77), agreeableness (α = .48), and neuroticism (α = .76). Responses were rated on a 7-point Likert scale (1 = strongly disagree, 7 = strongly agree).

Discussion. As organizations rapidly integrate AI into their workforces, there is an urgent need to understand the conditions under which workplace AI use supports - or undermines - productive collaboration. Answering this question requires going beyond analyses of the technical capabilities of frontier models, to consider the perspectives of the everyday workers who are on the frontlines of technological change. To meet this need, we surveyed full-time American desk-workers (N = 962) to understand how they contended with the day-to-day implications of sending and receiving AI-generated work.

Examination of these experiences revealed that many workers have encountered a specific, problematic form of workplace AI use, which we term workslop: low-quality, predominantly AI-generated content that appears to fulfill a task but lacks the substance needed to meaningfully advance the work at hand. We advance efforts to understand the effects of AI on organizations by (1) identifying workslop as a widespread but counterproductive application of AI, (2) characterizing the individual and organizational conditions that make sending workslop more likely, and (3) demonstrating that receiving workslop not only impedes productivity but also undermines workplace relationships. Conceptualizing AI-generated workslop Relative to past studies of workplace AI use that focus on documenting the technical limitations of AI (i.e., producing inaccurate information like hallucinations; Farquhar et al., 2024), workslop can be understood as a failure mode of AI-mediated communication (Hancock et al., 2020) because it emerges from how the AI mediates the sender-receiver relationship: senders offload cognitive work to AI systems that lack full context, creating work with manipulated quality signals that impose disproportionate burdens on the recipient. Analysis of workers’ experiences with workslop indicates that it comes in many forms, ranging from interpersonal communications (e.g., emails, Slack messages) and internal drafts to public-facing work and presentations. Below, we discuss several distinctive features that characterize workslop across modalities.

First, it is important to note that workslop is an inherently perceptual phenomenon. Whether a given artifact is experienced as workslop depends on the recipient’s interpretation of whether it is low-quality, the product of low-effort, or largely AI-generated. These judgments are important because workers generally have limited visibility into the steps their colleagues take when writing emails, developing slides, or producing other work artifacts. Similarly, they often lack information about if, when, or how AI tools were used unless it is explicitly disclosed or accidentally revealed (i.e., leaving in text that says “Claude can make mistakes”). As a result, even work produced with substantial prompting and thoughtful revision may therefore still be perceived as workslop if that effort is not legible to the recipient. Our results suggest that the mere perception of AI being used to cut corners at work can carry significant consequences. This aligns with past work finding that even when individuals could not confirm if AI was used or not, they rated those suspected of using AI as significantly less trustworthy (Schilke & Reimann, 2025). In this context, the perception of AI-enabled shortcutting - rather than confirmed AI use or technical error - can be sufficient to trigger negative interpersonal judgments.

Second, a feature that differentiates workslop from merely low-quality work is that it often appears legitimate on the surface, but lacks the substance needed to meaningfully advance the work at hand.

Conclusion. Many American workers in diverse occupations report both sending and receiving AI-generated workslop: work that they believe to be low-quality and unable to meaningfully progress their projects and initiatives. We conceptualize this phenomenon within the theoretical framework of AI-mediated communication, quantify its prevalence, and describe characteristics about its form and signals used for detection. We also identify specific individual-level and organization-level predictors of sending workslop. Finally, we describe the negative impact of receiving workslop on productivity and interpersonal relationships. While it is important to recognize that not all AI-generated work is workslop, and can often be beneficial, our results offer a counterpoint by highlighting a specific case in which workplace AI use at best is ignored, or worse, harms productivity, collaboration, and trust in teammates. Understanding the nuances of how specific people make decisions about if and how they use AI in their jobs can help us identify the contexts in which AI use may be most beneficial, or most detrimental.

Code and Data Availability ​ All code and data may be found at https://osf.io/xh2wn/overview?view_only=675432f416474ce592257dfd684343e3