"If You're Very Clever, No One Knows You've Used It": The Social Dynamics of Developing Generative AI Literacy in the Workplace

Abstract Generative AI (GenAI) tools are rapidly transforming knowledge work, making AI literacy a critical priority for organizations. However, research on AI literacy lacks empirical insight into how knowledge workers’ beliefs around GenAI literacy are shaped by the social dynamics of the workplace, and how workers learn to apply GenAI tools in these environments. To address this gap, we conducted in-depth interviews with 19 knowledge workers across multiple sectors to examine how they develop GenAI competencies in real-world professional contexts. We found that, while knowledge sharing from colleagues supported learning, the ability to remove cues indicating GenAI use was perceived as validation of domain expertise. These behaviours ultimately reduced opportunities for learning via knowledge sharing and undermined transparency. To advance workplace AI literacy, we argue for fostering open dialogue, increasing visibility of user-generated knowledge, and greater emphasis on the benefits of collaborative learning for navigating rapid technological developments.

Introduction. Generative Artificial Intelligence (GenAI) tools are rapidly redefining modern knowledge work [23, 36, 37, 45]. Adoption rates have increased annually since 2023 [105, 106]. Around 78% of 31,000 knowledge workers surveyed in Microsoft and LinkedIn’s 2024 Annual Report of Work Trends Index reported bringing their own AI tools to work [72]. As a result, there has been increased interest in reskilling existing employees to newly automated workflows, and ensuring responsible usage by increasing users’ awareness of the social and ethical implications of GenAI use [69, 75, 105, 106, 139]. Establishing AI literacy is now essential. AI literacy describes the competencies necessary for individuals to understand, interact with, and critically evaluate AI technologies [6, 66, 86]. Important competencies often include understanding the technical underpinnings of AI [18, 66]; interaction techniques (e.g., prompt design [55]); variations of learning, critical thinking, reflection, and metacognitive skills [28, 58, 68, 74, 86]; and socio-technical competencies, such as understanding the role of human actors in AI use through the lens of social and ethical impacts (e.g., [8, 18, 86]). Lower levels of AI literacy are consistently linked with a range of problems, including a tendency to over-rely on AI tools [88, 117, 136], be more susceptible to hallucinations and biases [83], lower job performance compared to peers with higher AI literacy [65], and lower employability [89]. Significant effort has been directed towards defining and designing for student AI literacy in formal educational environments [22, 74, 131]. However, professional users often rely on more unstructured methods of learning (e.g., informal knowledge sharing [51, 59, 99]) to develop highly personalised understandings of how and when different tools should be applied in their work, that can differ from those emphasised in existing AI literacy frameworks. Indeed, current GenAI adoption among employees often occurs independently of organisational guidance and strategy: surveys indicate that 53% of employees do not disclose their GenAI use to their employers [72], while McKinsey & Company’s 2025 analysis showed that C-suite leaders significantly underestimated the extent of such adoption in day-to-day work [69]. Many professionals are therefore constructing their own understanding of GenAI competencies and learning strategies, shaped less by organisational guidance or formal literacy frameworks, and more by personal priorities and their wider social context [46].

As the attitudes and beliefs of today’s workforce will shape future norms of interaction and learning, it is vital to understand how knowledge workers develop GenAI competencies in practice. Specifically, we explore perspectives that elaborate and contextualise workers’ beliefs and learning strategies for GenAI competencies, focusing in particular on extending understanding of overlooked socio-technical competencies (see Table 1). We aim to address two main Research Questions (RQs):

RQ1: What competencies do knowledge workers value when using GenAI tools in the workplace? RQ2: Which strategies do knowledge workers consider essential to developing these competencies?

In answering these questions, we make three main contributions:

(1) We conducted semi-structured interviews with 19 knowledge workers from diverse professions, allowing us to develop an empirically grounded account of how GenAI literacy is practised and negotiated within our sample. (2) We showed that, while workers benefitted from colleagues sharing knowledge, the act of critiquing and obscuring their own GenAI usage was also valuable to them. Importantly, while prior work has interpreted acts of GenAI hiding as rooted in shame [137], or a desire to elevate one’s status [98], we found that users can also view successful erasure of telltale signs of GenAI outputs as a positive indication of one’s professional expertise. However, these behaviours have negative impacts on further opportunities to learn via knowledge sharing, and promote cultures that lack transparency (Section 4). (3) We extend existing AI and GenAI literacy frameworks by highlighting the need to consider human stakeholders both as a resource for learning and as a source of social influence on usage. Building on this perspective, we highlight implications for HCI design to support open dialogue, disclosure, and sustainable knowledge sharing in the workplace (Section 5).

Related work. Our work builds on several lines of research that may be grouped into three areas: (1) GenAI in knowledge work (Section 2.1); (2) developing digital skills in the workplace (Section 2.2); and (3) defining AI and GenAI literacy (Section 2.3).

Knowledge work describes the complex and non-routine labour that requires tailored problem-solving skills and high levels of expertise [31, 34, 113]. Knowledge workers in various sectors (e.g., researchers, consultants, creatives, IT professionals [32, 113]) comprise a significant part of the world economy. In recent years, there has been an increase in interest and uptake in the use of GenAI tools to knowledge work [23, 36, 37, 45]. We define GenAI as end-user tools that draw on deep learning approaches and large amounts of training data to create outputs in response to user prompts [11, 23, 97]. The integration of GenAI tools into knowledge work is expected to bring benefits to productivity, but also challenges [103]. GenAI tools can conduct many core components of knowledge work tasks [13, 93, 116], such as searching and collating information [107, 132, 138], communication [3, 112], or generating tailored solutions in problem-solving [12, 94, 120]. However, recent work suggests that the productivity gains from automating these tasks may be offset by demands for new skill-sets which arise from using GenAI [26]. Multiple studies have highlighted that the use of GenAI tools presents a shift in users’ labour from production to supervision [57, 104, 126]. This shift introduces new tasks to preexisting workflows, such as refining prompts or adapting outputs (i.e., task stewardship) [30, 55, 57, 135], but also require users to make informed decisions as to when and what types of tasks could be delegated to GenAI [48, 54], termed ‘critical integration’ [95]. However, this introduces new challenges. Tankelevitch et al. [115] highlight how adoption of GenAI introduces significant metacognitive effort by requiring users to clearly articulate and refine their goals when prompting [55, 135]. Reliance on low quality generated GenAI outputs can also threaten the need for human expertise and risk deskilling industries, negatively impact critical thinking and cognition [57], or promote practice that lose the human characteristics (e.g., empathy) for interpersonal work [54, 126]. As GenAI reshapes the skills and cognitive demands placed on knowledge workers, it becomes increasingly important to re-evaluate which competencies and learning strategies users prioritise in these evolving work environments, and their motivations for doing so.

Method. 3.1 Participants During participant recruitment, we first identified sectors, which, based on previous economic analyses, were adopting generative AI tools at rapid rates [23, 37, 69, 106]. We then identified three core characteristics of ‘knowledge work’ and recruited participants belonging to sectors that met those characteristics. We recruited from sectors where expertise is developed primarily through formal education or training [34] (i.e., IT, Data Science, Law), sectors where tasks involved non-routine or creative problem-solving that require contextual sensitivity to meet stakeholders’ needs [31] (i.e., Management and Human Resources, Creative industries), and sectors which specialised in knowledge creation and dissemination [32] (i.e., Research and Development). Incorporating diverse perspectives from different sectors is important for ensuring rigour in qualitative research, as our goal is not to produce a generalisable account of human behaviour, but to ensure that we extend existing interpretations of the research subject [15].

In order to better access participants from these different occupational roles, we used the online recruitment platform Prolific. Prior evaluations has demonstrated that data quality from participants on Prolific, evaluated through user attention, comprehension, honesty, and care with answering questions, was consistently higher than comparable platforms [4, 33, 82]. One of the advantages of using Prolific was that it allowed finer control over participants’ demographics and attributes, allowing recruitment to be targeted. We used both Prolific’s own in-built filters, based on participants’ self-reported demographics when registering for a Prolific account, as well as our own screening survey to ensure participants met our recruitment criteria, namely, participants should: have occupations belonging to one of the sectors identified above; use GenAI tools at least once a week and primarily for work-related (as opposed to leisure or personal) purposes; be employed in a full- or part-time capacity in their specified sector and role; have more than 1 year of experience in their current role; and are not working in organisations which explicitly prohibited the use of GenAI tools in their work, as this may constrain the types of learning resources they can access, which was an important aspect of the research question. However, we allowed participants to participate if their organisation had no clear GenAI policies, or if participants themselves were uncertain of the policies, as this allowed us to explore how our participants navigated and developed their own understanding of GenAI competencies independent of top-down organisational guidance. These criteria ensured that our participants had sufficient familiarity with their specific work environment, domain, and had the opportunity to access the relevant GenAI tools and learning resources to be able to discuss their experiences, perceptions, and learning strategies in the interview. We initially identified 25 eligible participants for interviews. Six participants (not included in the table) were excluded. Four dropped out of the study before the interview or experienced technical issues during the interview, and two were excluded post-interview because they used tools which we could not reliably verify as using GenAI models. Our final data-set consisted of semi-structured interviews with 19 knowledge workers collected from December 4th 2024 to January 17th 2025, which meets the standard sample size expected for the field [16].

3.2 Interview Protocol and Data Collection We developed a semi-structured interview protocol to facilitate data collection. As our participant sample spanned a wide range of sectors, we ensured mutual understanding by establishing at the beginning of interviews: (1) the occupation and professional responsibilities of the participant; 2) their understanding of the term ‘generative AI’; and 3) how participants currently use GenAI tools in their work.

Discussion. The goal of our study was to extend existing interpretations of GenAI literacy by providing detailed insight into the beliefs, learning strategies, and in particular extending understanding of how users interpret socio-technical GenAI competencies within a sample of 19 knowledge workers. We found that, while participants valued learning from their colleagues’ experiences in order to expand their own awareness and knowledge, the ability to remove cues indicating their own GenAI usage was also important. This activity inevitably reduces others’ ability to detect GenAI usage, and therefore to develop awareness and access to others’ experiences. This promoted a lack of transparency around GenAI use which was exacerbated in organisations with unclear policies. However, we also found that the ability to remove obvious GenAI ‘tells’ was not always motivated by stigma and fear of social judgement (though this was an important influence), but could also be interpreted as a positive indication of users’ domain expertise. These findings together surface a tension between individual identity work and collective learning that has meaningful implications for how we design for and support GenAI literacy in practice.

Our study contributes and extends existing GenAI literacy frameworks by encouraging researchers to reflect more deeply on how competencies for interacting with GenAI manifest for users in context. In the following section, we reflect on and explore the potential implications of our findings.

5.1 Contrasts with Prior GenAI Literacy Frameworks Our findings challenge current interpretations about the competency needed to detect GenAI outputs. The most recent GenAI literacy framework proposed by Annapureddy et al. [8] positions the ability to identify cues of GenAI-generated outputs as a way to shield individuals from being misled by bias, errors, and misinformation perpetrated by other GenAI users. However, our findings in Section 4.3 demonstrate that users apply their knowledge of GenAI cues to remove evidence of GenAI usage in their own work as well. While previous literature have predominantly focused on AI hiding behaviours as an act of shame [98, 137], and indeed, we also observed social stigma as an important motivator for concealing GenAI use [90, 98, 101], our study highlights how the ability to remove cues of GenAI use can both serve to improve the perceived quality of work, and help users to validate their own domain expertise over those who ‘fail’ to hide their use of GenAI tools. Indeed, GenAI hiding could even be applied to accumulate personal reputational credit for participants, as we saw with P14, who utilised GenAI to solve a colleague’s problem, but took personal credit for developing the solution.

This suggests a potential need for researchers themselves to reframe the behaviours around what is currently perceived as GenAI hiding or GenAI shaming, to consider how such behaviours may in fact benefit other user needs. This mirrors similar research into GenAI non-use, where, rather than viewing non-use as an unambiguously undesirable rejection of a ‘useful’ tool, researchers have instead highlighted how non-use can instead reflcet strategic acts of identity protection on behalf of the user, to preserve essential skills and knowledge rather than delegating to the user [20, 126].

Conclusion. This paper extends current discussions on GenAI literacy through an empirical exploration into how knowledge workers outside of formal education define key GenAI competencies and strategies for learning. We found that users value broad awareness of GenAI’s capabilities due to the complexities of the constantly evolving field and to enhance their own competitiveness. Furthermore, learning from the experiences of others is crucial for maintaining this awareness. However, concealing ‘obvious’ cues of GenAI usage is also valued as a competency, not only to avoid stigma, but as a way of re-affirming users’ own domain expertise. We contribute to existing knowledge on GenAI hiding behaviours by highlighting why users may be motivated to engage in hiding, and that these actions can promote a lack of transparency and critical reflection on GenAI use. To mitigate hiding behaviours, we suggest that designers enhance the visibility of user-generated knowledge when using GenAI tools, and that organisations can better emphasise the benefits of maintaining collaborative learning and open discussions to support responsible, transparent use.

Limitations. 5.5 Limitations and Future Work While our work provides important insights into how current adopters of GenAI negotiate with the concept of GenAI literacy in the workplace, there still remain several open questions for future works to explore. First, our work identifies the important role of other users in supporting discoverability of GenAI functionalities and applications to work, but it is not known to what extent which GenAI technologies could be applied to fulfill or replace this role, and what the consequences might be. Many current commercial GenAI tools do not yet incorporate clear, explainable AI principles to comprehensively support user understanding of how they may be used in context. Though explainable AI principles can be embedded into design to support the goal of learning, we suggest that the social influence of others on the practice of GenAI usage is likely to persist regardless. Our findings suggest that, regardless of whether or not people actually learn from their colleagues, the act of sharing one’s software knowledge can still be an important act for accumulating social credit and signalling domain expertise, and people’s understanding of which GenAI competencies matter (if they matter at all), is still influenced by others’ attitudes.