Invisible Competencies in Practice
- Invisible competencies are hidden, tacit skills that drive effective integration of AI tools and human judgment in sociotechnical systems.
- They encompass both technical workaround strategies and non-technical abilities such as communication, ethical reasoning, and contextual diagnosis.
- Research shows that these competencies are crucial in agile environments, AI literacy, and institutional governance despite being under-measured.
Searching arXiv for the cited papers and closely related work on invisible competencies, AI literacy, and sociotechnical labor. Invisible competencies are hidden, tacit, under-observed, or socially obscured capacities that make technical and organizational systems function in practice. Across recent work on generative AI, AI literacy, software engineering, requirements engineering, crowd work, and AI system design, they are treated not as marginal supplements to formal procedures, but as central capabilities that are often absent from official metrics, managerial narratives, or visible outputs. The literature presents them in several overlapping forms: as “articulation work” and “invisible labour” required for sociotechnical integration, as ethical and judgment-based dimensions of AI literacy, as separable skills for reasoning with generative models, as latent system competency distinct from nominal capability, and as non-technical or contextual abilities that are indispensable in collaborative technical work (Lee et al., 24 Dec 2025, Apartsin et al., 4 Jun 2026, Karlapalem, 2023, Qing et al., 1 Feb 2026).
1. Conceptual range and terminology
The term encompasses multiple but related phenomena. In one strand, invisible competencies denote the hidden work required to make an “unfinished” system operational within a real institutional setting. In another, they denote competencies that are real and consequential but latent, inferred only through success, failure, audits, or scenario-based testing. In a third, they denote non-technical capacities—communication, teamwork, contextual diagnosis, ethical reasoning, or judgment—that are overshadowed by technical discourse despite being structurally necessary to competent performance. A further strand shows that competencies can become socially invisible when workers intentionally remove cues of AI use, so that AI-assisted output appears as unaided human expertise (Lee et al., 24 Dec 2025, Karlapalem, 2023, Lima et al., 30 Apr 2026, Groeneveld et al., 2019, Qing et al., 1 Feb 2026).
A recurring pattern is that invisibility does not imply triviality. The studies repeatedly position these capacities as central to implementation, reliability, assessment, or professional excellence. What remains hidden is often the labor of integration, the judgment that governs tool use, the context-sensitive adaptation of practice, or the unpaid overhead required to convert formal capability into actual performance (Lee et al., 24 Dec 2025, Jantunen et al., 2019, Toxtli et al., 2021).
| Literature | Framing of invisible competencies | Principal emphasis |
|---|---|---|
| (Lee et al., 24 Dec 2025) | “articulation work” as “invisible labour” | sociotechnical integration |
| (Apartsin et al., 4 Jun 2026) | Framing, Judging, Steering | assessable AI-assisted reasoning |
| (Karlapalem, 2023) | micro-competency / macro-competency | capability vs competency |
| (Khan, 3 Nov 2025) | cognitive, behavioral, normative competencies | AI literacy and ethical judgment |
| (Qing et al., 1 Feb 2026) | technological, learning and reflective, socio-technical competencies | visibility, disclosure, and concealment |
| (Lima et al., 30 Apr 2026, Groeneveld et al., 2019) | soft skills / non-cognitive abilities | collaborative and interpersonal performance |
| (Jantunen et al., 2019) | contextual intelligence | diagnosis and adaptive behavior |
This range suggests that “invisible competencies” is best understood as a family resemblance concept rather than a single formal construct. The family resemblance lies in three features that recur across the literature: the competencies are consequential, they are often poorly represented by standard output-based evaluation, and they become visible most clearly when systems fail, when coordination breaks down, or when learning and governance mechanisms prove inadequate.
2. Articulation work, invisible labour, and sociotechnical integration
A detailed higher-education case shows invisible competencies as the labor required to convert a technically limited GenAI system into an institutional workflow. The original goal was to let non-technical staff query a large transactional Staff Development Fund log conversationally. This ran into explicit technical constraints: the full annual log for one faculty was projected to exceed 35,000 rows, this would exceed the model’s context window, LLM performance can degrade when relevant information is buried in the middle of long contexts, and large tabular data are token-inefficient and noisy for LLMs. The workaround was a two-part human-in-the-loop process: a Python script to join and summarise raw transactional Staff Development Fund (SDF) data, followed by a custom Microsoft Copilot conversational interface that queried the processed output rather than the full log. The script performed the “Serialization” and “Table Manipulations” needed to make the table usable for the LLM (Lee et al., 24 Dec 2025).
The same case makes clear that the workaround was not merely technical. It was also a response to concerns about job redundancy, bureaucratic and approval-related obstacles, perceived territoriality, positional power, and the risk that the project would be trapped in centrally managed bureaucracy. Using Alter’s theory of workarounds, the paper defines a workaround as a user-driven solution to perceived system limitations and maps the case onto a process of conflicting intentions / goals / interests, perceived need, identification, selection, development, and execution. Its central theoretical move is to interpret articulation work—“the meshing of the often numerous tasks, clusters of tasks, and segments of the total arc,” and work that “serves to establish, maintain, or break the coordinated intersection of task chains”—as a form of invisible labour. In this account, the more significant hidden work was not the script itself but “navigating the project’s human and political friction,” including “hidden emotional labour” required to traverse the “political and bureaucratic landscape” (Lee et al., 24 Dec 2025).
A different form of invisible labor appears in crowd work, where invisibility is measured economically rather than ethnographically. A field study of 100 workers on Amazon Mechanical Turk, instrumented through a plugin across 40,903 tasks, found that workers spent 33% of their time daily on invisible labor. If invisible labor is ignored, workers’ median hourly wage was \$3.76**; including it reduces the wage to **\$2.83. The largest categories were Payments and Hypervigilance. Category medians were 13 minutes/day for Payments, 11 minutes/day for Hypervigilance, 6 minutes/day for Lack of Guidance, 3 minutes/day for Breaks, and 1 minute/day for General Logistics. The paper formalizes hourly wage as
and
with unpaid-rate defined as
Its findings show that invisible competencies in platform labor include market scanning, requester evaluation, queue management, payment monitoring, and strategic vigilance, even though those capacities are unpaid and often excluded from wage calculations (Toxtli et al., 2021).
Taken together, these studies show two complementary meanings of invisible competencies. One is coordinative and political: the work of making systems fit organizations. The other is infrastructural and economic: the labor of making work itself obtainable and sustainable. In both cases, visible output systematically understates the competencies required to produce it.
3. Assessment, decomposition, and the formalization of hidden competence
One major research direction is to turn invisible competencies into assessable units. The CoRe-3, or Co-Reasoning, model argues that productive AI use depends on three separable competencies: Framing, Judging, and Steering, abbreviated FJS. Framing is pre-generation task specification: surfacing unstated assumptions, fixing constraints and scope, and stating what counts as an adequate solution. Judging is post-generation evaluation of the output for factual errors, logical errors, omissions, missing edge cases, unstated assumptions, and risks. Steering is iterative post-generation control through targeted corrections that redirect the model toward a better answer. The model’s distinguishing claim is the split between pre-generation Framing and post-generation Steering, with Judging as the gate between them (Apartsin et al., 4 Jun 2026).
The paper grounds FJS in a monitor-control architecture and states five propositions: Framing gates the loop; Judging gates Steering; Dissociation; Asymmetric transfer; and Inverse of offloading. It operationalizes them in CoReasoningLab, which presents an ill-defined problem, then a plausible but flawed AI solution, and scores each skill independently. Framing is scored against a gold-standard framing and rubric; Judging is scored against a seeded issue set using recall/precision logic; Steering is scored against the trajectory of the output across revisions. In a crossed-factorial simulation with 8 competence profiles across 10 subjects, yielding 80 simulated learners, each grade showed strong own-skill effects and near-zero cross-skill effects. The pattern held across GPT-4o, GPT-4o-mini, and Llama-3.3-70B, and repeat grading showed 92% self-consistency. The paper also states the ceiling relation
and uses an ordinal grading scale of
Its broader claim is that AI-era competence should be diagnosed not by a single “prompting” score or by the final answer alone, but by separately measuring the hidden operations that govern good AI-assisted work (Apartsin et al., 4 Jun 2026).
A system-level analogue appears in work on competent AI systems, which distinguishes Capability—what a system is expected to deliver—from Competency—what it does successfully under real conditions. The paper explicitly frames this as a gap between “what it can do vs. what it does successfully.” It proposes a competency-oriented framework divided into micro-competency and macro-competency. Micro-competency covers representation competency, data competency, algorithmic competency, and functional competency at the level of modules or atomic tasks. Macro-competency covers flow competency, action competency, and solution competency at the level of the whole system, with overall performance limited by the “least competent module” or “weakest competent link.” The framework also introduces inside-out and outside-in paradigms for specifying competency (Karlapalem, 2023).
This work emphasizes that competency is often implicit, hidden in software or hardware components, and only inferable from confidence intervals, success rates, audits, scenario-based testing, or accumulated experience. It gives the example that “a stemmer with 92% accuracy has competency defined by the circumstances under which it performs optimally,” and uses the Glass Door Problem to show how a general capability can fail under specific environmental conditions. The practical consequence is that trust should rest on “explicit modeling and detailed specification of its competency,” not merely on claims about capability (Karlapalem, 2023).
These two lines of work converge on a common point: invisible competencies can be made more visible through decomposition, independent scoring, scenario-based evaluation, and explicit modeling. They also show that invisibility is partly an artifact of coarse-grained assessment.
4. AI literacy as hidden judgment, ethical appraisal, and social invisibility
AI literacy research has increasingly treated invisible competencies as the difference between declarative familiarity and responsible, context-sensitive use. A study of LIS professionals in the United Arab Emirates uses a three-part competency model: 2 3 4 The findings show strong cognitive competencies, but uneven behavioral competencies and the most significant gaps in normative competencies, especially identifying potential AI biases. Respondents were also concerned about privacy breaches, bias amplification, and transparency challenges, and the study found a disconnect between the perceived importance of AI skills and the effectiveness of existing training programs. Its Figure 1 proposes four connected actions for “Enhancing AI literacy in libraries”: Comprehensive AI education, Specialized training programs, Practical AI applications, and Ethical standards and guidelines (Khan, 3 Nov 2025).
This literature therefore locates invisible competencies not in the ability to name tools or explain AI basics, but in judgment-intensive capabilities such as bias detection, privacy awareness, transparency, accountability, critical evaluation of outputs, and responsible design of AI-enabled services. These capacities are less visible in routine operations because they often appear only when systems fail, when harms become salient, or when professionals are required to justify or constrain AI use (Khan, 3 Nov 2025).
A workplace study adds a distinct social dimension. Based on semi-structured interviews with 19 knowledge workers across sectors including IT, data science, law, management and HR, creative industries, research and academia, finance/accounting, architecture, logistics, and marketing, it identifies two core GenAI competencies: Awareness of potential GenAI use cases and Ability to remove cues indicating GenAI use. These are learned through two strategies: Learning from others and Learning through hands-on experience. The study organizes its analysis with the COM-B model—Capability, Opportunity, Motivation—and with three competency dimensions: Technological competencies, Learning and reflective competencies, and Socio-technical competencies (Qing et al., 1 Feb 2026).
The paper’s central finding is that GenAI competence can become socially invisible. Workers learn not only how to use systems, but how to erase cues that signal that AI was used. Concealment may be driven by stigma, organizational ambiguity, or fear of appearing lazy, but it may also be status-enhancing. The study distinguishes concealment from critique: hiding AI use to avoid judgment is different from identifying and removing AI “tells” as a form of quality control. Yet both processes make competence legible through the absence of visible AI traces. The title quotation—“If you’re very clever, no one actually knows you’ve used it”—captures a setting in which domain expertise, professional identity, and reputational capital are produced through the quiet control of AI rather than open disclosure (Qing et al., 1 Feb 2026).
A plausible implication is that AI literacy has two visibility problems. One concerns hidden judgment: ethical and evaluative capacities are not apparent from superficial competence. The other concerns hidden provenance: successful AI-assisted work may be socially reclassified as purely human expertise. Both limit shared learning if organizations lack norms that make AI-mediated practice discussable.
5. Invisible competencies in software engineering and requirements engineering
In software engineering, invisible competencies are often treated as the human side of sociotechnical work. A systematic mapping of 97 studies published between January 2000 and May 2025 identifies 33 distinct soft skills, consolidated into a taxonomy of 22 main interpersonal skills, across agile development contexts. The most frequently mentioned are Communication with 87 mentions (89.6%), Teamwork with 63 (64.9%), Leadership with 45 (46.3%), Adaptability with 36 (37.1%), Proactivity with 32 (32.9%), Motivation and Problem Solving with 29 (29.9%) each, and Creativity with 26 (26.8%). The paper distinguishes increasing trends—Communication, Adaptability, Problem Solving, Willingness to Learn, Knowledge Management, Critical Thinking—from stable skills such as Teamwork, Leadership, Proactivity, Motivation, Organization, Creativity, Responsibility, Socialization, Trust, Business Vision, Empathy, Negotiation, Active Listening, Ethics, Conflict Management, and Goal Setting. It also maps skills to roles and notes that Scrum dominates the literature, with 55 mentions in the RQ3 section and appearing in 21% of the publications (Lima et al., 30 Apr 2026).
The mapping’s significance lies in its explicit sociotechnical framing. Soft skills are not treated as peripheral “people skills,” but as competencies essential for quality, productivity, innovation, collaboration, and responsiveness in agile teams. The paper also identifies major literature gaps: lack of role-specific studies, limited attention to seniority, underrepresentation of newer or specialized roles, and the fact that 41 studies did not specify the agile methodology used. This suggests that invisible competencies are widely acknowledged yet still under-specified in the research base (Lima et al., 30 Apr 2026).
A Delphi study of exceptional software engineers provides a finer-grained taxonomy. Drawing on 36 experts from 11 countries and 21 institutions, it identifies and ranks 55 non-technical abilities in four categories: Communicative skills (11), Collaborative skills (16), Problem solving skills (13), and Personal skills (15). The study uses Kendall’s W with the stopping criterion
and reports second-round values of 0.74 for technical experts, 0.80 for business experts, and 0.58 for academics. Technical experts ranked systematically verify assumptions and validate results first; business experts ranked collaborate with others to achieve a shared goal first; academics ranked understand and engage with the people involved during development first. The paper also records a respondent’s objection to the label “non-cognitive”: “Non-cognitive seems not only incorrect but derogatory. All non-technical skills are cognitive, require practice, consideration and deliberate exercise to improve” (Groeneveld et al., 2019).
Requirements engineering extends the discussion from teamwork to contextual diagnosis. A vision paper argues that RE in dynamic environments is not primarily a matter of eliciting stable requirements, because modern problems are characterized by uncertainty, ambiguity, emergence, and co-evolution of system and environment. It proposes contextual intelligence as the central competency and defines it as “the ability to recognize and diagnose the plethora of contextual factors inherent in an event or circumstance and then intentionally and intuitively adjust behavior in order to exert influence in that context.” The paper uses the Cynefin distinction among Simple, Complicated, Complex, and Chaotic problems, and argues that for complex problems practitioners should probe, then sense, then respond. It further distinguishes horizontal development from vertical development, and identifies supporting competencies such as learning to learn, sensemaking, dialogue, mindfulness, and facilitative leadership (Jantunen et al., 2019).
These software and RE literatures broaden the meaning of invisible competencies. They show that hidden competence is not limited to AI-specific practice. It also includes the social, interpretive, and contextual capacities through which technical work becomes adaptive, collaborative, and resilient.
6. Educational consequences, dependency, and institutional governance
Educational research increasingly treats invisible competencies as capacities that may be obscured, weakened, or displaced by AI-supported performance. A study of Filipino college students uses a 24-item instrument across five domains—critical thinking, writing skills, learning independence, research skills, and academic engagement—and analyzes 651 valid responses through Latent Class Analysis. The selected four-class model had AIC = 109.4 and BIC = 4243.13, yielding four profiles: Highly Engaged Independent Learners, Selective AI Users, Moderate AI Users, and AI-Dependent Learners. The most concerning group, AI-Dependent Learners, had the weakest academic competencies across all dimensions, especially weak critical thinking, writing, and research skills, and very low learning independence and engagement. Descriptively, the highest dependency scores were Uncritical acceptance of AI outputs with , Dependency on AI for research suggestions with , and Substitution of AI for academic engagement with . The study found no significant gender differences in AI dependence patterns (0), no significant association with academic performance, and significant variation by academic year (1) (Fernando et al., 30 Apr 2026).
This work suggests that invisible competencies can be hidden behind polished AI-assisted outputs. A student may appear competent in writing or research while underlying critical thinking, learning independence, or academic engagement is weakened by substitution. The paper’s policy position is explicit: higher education institutions should integrate AI literacy while preserving essential skills, and AI should support rather than substitute learning (Fernando et al., 30 Apr 2026).
Institutional governance problems arise when invisible competencies are indispensable but unrecognized. In higher education GenAI implementation, workarounds may become “shadow IT” or “shadow system[s]”, solving immediate problems while operating outside official governance and obscuring the labor behind them. The same paper argues that institutions should stop treating workarounds as mere deviance and should instead see them as signals that “the system does not fit the work,” that implementation needs more sociotechnical attention, and that the labor of integration must be recognized, supported, and governed (Lee et al., 24 Dec 2025).
Library AI literacy research arrives at a parallel conclusion from a different angle. Because the weakest areas are ethical and judgment-based, the paper recommends specialized, tailor-made, practical AI literacy training, stronger integration of AI literacy into LIS curricula, deeper links to computer science, and clear ethical standards and guidelines for privacy, bias, accountability, and responsible service delivery (Khan, 3 Nov 2025). Workplace research likewise argues for open dialogue, visibility of user-generated knowledge, and reduced reputational penalties for disclosure, because invisibility may be individually advantageous yet institutionally harmful through reduced knowledge sharing and weaker transparency (Qing et al., 1 Feb 2026).
Across these settings, the governance challenge is consistent. Invisible competencies are not merely missing annotations on otherwise self-sufficient systems. They are often the actual mechanisms by which systems become usable, trustworthy, and educationally legitimate. When organizations ignore them, they risk mismeasuring competence, undervaluing labor, misgoverning informal infrastructures, and confusing visible output with underlying capability.