Artificial intelligence (AI) has entered higher education with remarkable velocity. Generative AI systems, large language models (LLMs) capable of producing fluent, contextually coherent text, are already being adopted or actively considered by students, academics and administrators across both the Global North and South (Baidoo-Anu & Ansah 2023; Kasneci et al. 2023). In South Africa and across the continent, this uptake unfolds within educational landscapes marked by uneven resource distribution, contested epistemologies and enduring debates about the university’s role (Badat 2016; Fataar 2018). Recent commentary in the higher education press underscores the urgency of these questions: Fataar (2026) reflects on lessons from AI benchmarking in Asia, while McKenna (2026) publicly advocates for embedding critical AI literacy in curricula. Such journalistic commentaries, cited here as indicators of sectoral concern, highlight that preparing learners to engage with AI is not a merely technical challenge but a profoundly educational, political and institutional one.

Responses to this challenge have coalesced around the concept of ‘AI literacy’. Yet, a critical examination of how this construct is being operationalised reveals a troubling pattern: most frameworks equate AI literacy with what may be termed ‘prompt literacy’, the ability to use AI tools effectively; to craft better prompts, to select the right tool for a task, to understand broadly how these systems function (Long & Magerko 2020; Ng et al. 2021). Whilst such competencies are not without value, they represent a fundamentally instrumental orientation that fails to address the most consequential challenge posed by generative AI in educational and intellectual contexts: that what these systems produce often looks credible without being accurate. In the South African higher education context, where epistemic transformation is an institutional imperative, this failure is not incidental. Artificial intelligence tools that are adopted without adequate epistemic scrutiny risk becoming new instruments for the reproduction of the very knowledge hierarchies that transformation seeks to dismantle.

Large language models are designed structurally, not incidentally, to maximise plausibility rather than truth (Almufarreh & Arshad 2023). This distinction between the plausible and the accurate is not merely semantic; it is the epistemological fault line that prevailing AI literacy frameworks have failed to adequately theorise. The result is a discourse that produces graduates adept at prompting AI tools but poorly equipped to interrogate what those tools actually produce. In a scholarly environment where the markers of academic authority - citation, qualification, disciplinary register - are precisely what LLMs have learned to reproduce, this gap has consequences for higher education.

The purpose of this article is therefore to interrogate the epistemological assumptions embedded in prevailing AI literacy frameworks and to develop, in their place, a theoretically grounded reconceptualisation of AI literacy adequate to the evaluative demands that generative AI places on scholars, students and institutions, particularly those committed to epistemic transformation in South African and African higher education. Two research questions guide the analysis:

Pursuant to these questions, this article argues that prevailing frameworks of AI literacy are epistemologically inadequate. It advances the claim that AI literacy should be reconceptualised as a practice of epistemic discernment: an evaluative orientation towards AI outputs that distinguishes between surface-level plausibility and substantive epistemic warrant. The argument unfolds through three interrelated moves. Firstly, it surveys the dominant AI literacy landscape, identifying its structural limitation as a failure of epistemic rather than technical imagination. Secondly, it draws on Connell and Keane’s (2004, 2006) Knowledge-Fitting Theory of Plausibility, virtue epistemology, critical information literacy, and the lateral reading research of Wineburg and McGrew (2019) to construct a theoretically grounded alternative. Thirdly, it situates this alternative within the broader context of epistemic transformation in South African and African higher education, contending that the biases embedded in globally deployed AI systems render epistemic discernment not merely a general educational good but a prerequisite for genuine intellectual autonomy among scholars working within non-dominant knowledge traditions. The article then engages the most plausible objections to its central claims, before translating the theoretical framework into concrete institutional and pedagogical implications.

This article proceeds as a theoretical position paper, grounded in a critical and explicitly normative reading of the existing literature. Rather than a systematic review aimed at synthesising empirical findings, its purpose is analytical: to interrogate the theoretical assumptions embedded in existing AI literacy frameworks and to develop an epistemologically adequate alternative. Selection in this kind of analysis is purposive rather than exhaustive: sources illuminate the argument being developed rather than providing comprehensive coverage of the field. This is characteristic of theoretical scholarship in the critical tradition, as distinguished from systematic review, and is a legitimate mode of scholarly inquiry (Barnett 1997; Freire 1970).

The theoretical architecture draws on six disciplinary traditions, and the logic of their selection and synthesis warrants explicit statement. Each tradition is enlisted because it answers a distinct sub-question within the overall argument, as no single tradition can answer them all. Artificial intelligence literacy scholarship supplies the object of critique: the frameworks whose epistemological assumptions the article interrogates. Cognitive science, through the Knowledge-Fitting Theory of Plausibility (Connell & Keane 2004, 2006), supplies the mechanism: an account of how plausibility judgements are formed and why they can be systematically misled. Philosophy of knowledge and virtue epistemology supply the normative standards: accounts of epistemic warrant, calibrated trust, and the dispositions of responsible knowers against which evaluative practice can be judged (Fricker 2007; Zagzebski 2012). Critical information literacy supplies the political dimension: an insistence that knowledge evaluation is always situated within institutional and social conditions (Elmborg 2006). Verification pedagogy supplies the empirical warrant for teachability: evidence that evaluative practices can be explicitly taught (Wineburg & McGrew 2019). Higher education theory supplies the institutional location: an account of the university as an epistemic community with distinctive purposes and responsibilities (Barnett 1997; Morrow 2009).

The synthesis across these traditions is problem-driven rather than eclectic, organised around a single question: what does it take to evaluate LLM output responsibly? Each tradition contributes what the others lack, with tensions acknowledged and managed rather than dissolved. The most significant tension lies between the individual-cognitive register of plausibility research and virtue epistemology, on the one hand, and the structural-political register of critical information literacy and decolonial scholarship, on the other. This article manages the tension by assigning distinct explanatory tasks: cognitive accounts clarify how the plausibility trap operates within individual judgement, while critical and decolonial accounts situate that mechanism within unequal conditions of knowledge production, conditions that determine who is most vulnerable to it and whose knowledge is distorted. The two registers are thus treated as complementary levels of analysis, not competing explanations.

This article adopts an explicitly normative approach: it does not only describe what current AI literacy frameworks contain, but argues for what they should contain. The philosophical justification rests on three grounds. Firstly, it aligns with critical educational theory, which treats the purposes of higher education as contested and regards normative argument about what ought to be as a legitimate scholarly mode (Barnett 1997; Freire 1970). Secondly, its conclusions are offered for evaluation against criteria appropriate to theoretical work: internal coherence, adequacy to the phenomenon, explanatory power relative to rival accounts, and the capacity to generate researchable propositions and actionable guidance. Readers are invited to assess the argument by these standards rather than by empirical generalisation, which the article does not claim. Thirdly, the approach resonates with the journal’s concern for transformation in higher education: transformation is treated not only as an institutional and social project but also as an epistemological one, where preparing learners to know and evaluate knowledge is inseparable from the question of what kind of university the institution ought to be.

A note on positionality is necessary, and it must extend beyond ritual acknowledgement. The author writes from his position as a higher education scholar within the South African public university system, with sustained involvement in national programmes on the recognition and development of university teaching. This location shapes the analysis in two key ways. Firstly, it frames the South African transformation agenda as the evaluative context for testing AI literacy frameworks, rather than treating it as one application among many. Secondly, it foregrounds institutional dimensions of AI literacy, reflecting direct experience with how policy commitments succeed or fail at the organisational level. Positionality, in this account, is not a disclaimer but an explanation of the analytical choices it enables.

The most widely cited frameworks for AI literacy share a common structural logic. Long and Magerko (2020), in a foundational synthesis, identified competencies that include understanding AI’s capabilities and limitations, critically evaluating outputs, integrating AI into everyday life, and collaborating with AI systems. Ng et al. (2021) proposed four dimensions: awareness, usage, evaluation and ethical reflection. Cetindamar et al. (2024) examined AI literacy in digital workplaces, highlighting the need for contextual embedding. In the United States, the AI4K12 initiative articulated five ‘big ideas in AI’ to scaffold conceptual development across educational levels (Touretzky et al. 2019). More recent work has emphasised the community-embedded nature of AI literacy, particularly the role of informal online communities in shaping how learners evaluate AI-generated content (Stewart 2026). Luckin et al. (2016), in an early and influential account of AI’s educational potential, argued for cultivating ‘meta-learners’ who not only use AI tools but also learn alongside them, a formulation that anticipates epistemic engagement, though without fully theorising it.

Across these frameworks, several patterns emerge. Firstly, it is important to acknowledge that the evaluation of AI outputs is present in the most influential accounts: Long and Magerko (2020) list ‘critically evaluating AI’ among their 17 competencies, and Ng et al. (2021) position evaluation as a core dimension. The substantive concern is not the absence of evaluation but its epistemological thinness. What it actually means to evaluate an AI output, the epistemic standards relevant to such evaluation, and the specific challenges that LLMs pose to conventional evaluative practice are insufficiently examined. Evaluation functions, across most frameworks, as a checklist item rather than a theorised practice. Secondly, the framing of AI literacy is predominantly individual: it addresses what a person should know and be able to do, with limited attention to the institutional and structural conditions that shape whether such capacities can be cultivated and exercised. Thirdly, ethical dimensions, where addressed, tend to focus on bias, privacy, and fairness, all important concerns, but distinct from the epistemic question of whether outputs can be trusted as accurate.

An important qualification accompanies this critique. The frameworks examined are among the most widely cited and structurally influential, and the patterns they display are taken here as representative of the dominant discourse. However, a purposive reading of leading frameworks cannot amount to a comprehensive census of the field. The claim advanced is therefore about prevailing tendencies in AI literacy as reflected in its most influential accounts, not about the impossibility of exceptions. The net effect is that contemporary conceptions of AI literacy align more closely with what Cope and Kalantzis (2009) term operational literacy: the capacity to make the system function, rather than critical literacy, which entails understanding the conditions of its production and the interests and limitations embedded within it. This distinction is far from trivial. In educational contexts where LLMs are increasingly deployed to support research, writing and knowledge synthesis, the inability to interrogate their outputs constitutes an epistemic vulnerability of the highest order.

To grasp why the epistemological gap in prevailing frameworks is so consequential, it is essential to be precise about how LLMs generate text. Trained on vast corpora, LLMs operate by predicting the next most probable word or token in a given sequence (Bender et al. 2021). Crucially, they do not retrieve information from a database of verified facts; rather, they generate text that reflects statistical regularities in their training data. The outputs are calibrated to be coherent and contextually appropriate; that is, to appear plausible, but plausibility here is a function of textual patterning, not of epistemic warrant.

The Knowledge-Fitting Theory of Plausibility, developed by Connell (2004) and elaborated by Connell and Keane (2004, 2006), provides a theoretical account of plausibility as a cognitive judgement about how well new information aligns with an individual’s existing knowledge of the world. It distinguishes between two forms of fit: word-coherence – the extent to which an expression conforms linguistically to the conventions of credible discourse, and concept-coherence, the extent to which its claims correspond with verified knowledge (Connell & Keane 2004). Large language models excel at generating word-coherence. Trained on vast corpora of academic and professional writing, they reproduce vocabulary, rhetorical conventions, hedging strategies, citation patterns, and registers of expert discourse with striking fidelity. What they cannot guarantee, however, is concept-coherence: their outputs may diverge from established evidence while remaining linguistically indistinguishable from authoritative scholarship. This dynamic produces what this article terms the plausibility trap: the systematic misrecognition of linguistic credibility as epistemic validity. In scholarly contexts, this conflation constitutes the defining epistemic hazard of LLM use.

A brief example illustrates the plausibility trap in practice. A postgraduate student preparing a literature review asks a generative AI tool for the five most important sources on epistemological access in South African higher education, with summaries. The output appears exemplary: it names recognised scholars, provides plausible article titles, journals, dates, and page ranges, and summarises each source in confident disciplinary prose. Read vertically, assessed on fluency, citations, and scholarly tone, the text passes every surface test. Read laterally, however, it unravels: a database search shows that two references do not exist as cited, a third misattributes an argument, and one summary reverses the position of the original text. None of these failures is signalled in the output, which performs the markers of scholarship flawlessly. A student relying only on its internal coherence has no way of detecting that word-coherence has been substituted for concept-coherence.

This composite illustration, reflecting failure patterns documented in the literature on AI-fabricated references (Alkaissi & McFarlane 2023), exemplifies the very scenario that an epistemologically adequate AI literacy must prepare learners to navigate. Language models (LMs) can generate outputs that are fluent, authoritative in tone, internally consistent and entirely inaccurate. They may cite sources that do not exist, misattribute quotations, or produce confident summaries that diverge materially from the scholarly record (Alkaissi & McFarlane 2023). These are not incidental errors but structural features of how such systems generate text. The technical literature often labels this tendency ‘hallucination’, yet the term understates its epistemological weight by framing it as a malfunction rather than a constitutive risk. Many deployed AI systems now attempt to mitigate this risk through retrieval-augmented generation, browsing functions, or citation tools. While these architectures reduce some errors, they do not resolve the epistemic problem: retrieval quality, source selection, and summarisation fidelity remain uneven, and outputs continue to perform authority in ways users cannot easily interrogate without the very verification practices this article argues must be cultivated.

To understand this phenomenon, we need a more theoretically precise term. This article proposes the concept of the ‘stochastic alibi’ to describe how LLMs generate outputs that are probabilistically assembled yet rhetorically presented as credible knowledge. The model provides students and researchers with polished, authoritative-sounding explanations that mask factual uncertainties behind a facade of credibility. While ‘hallucination’ is an empirically measured construct in applied studies (Alkaissi & McFarlane 2023), the stochastic alibi is not a competing empirical measure. Instead, it is a conceptual redescription meant to illuminate what empirical framing obscures. Whereas ‘hallucination’ views errors as temporary failures within a system meant to tell the truth, an ‘alibi’ highlights a structural, pre-emptive strategy: it is designed to deflect suspicion by pretending to be something it is not.

Under this view, an LLM is not an occasionally unreliable informant; rather, it is a system that simulates credibility itself. This builds on Floridi’s (2023) analysis of LLMs as producing syntactically plausible but semantically ungrounded outputs, signifying agency without intelligence. This simulation risks exploiting the misplaced trust of readers who mistake stylistic markers like citations, disciplinary jargon and confident assertions for actual reliability. It also echoes older philosophical critiques of representation that mask an underlying absence, such as Baudrillard’s (1994) simulation and Derrida’s (1976) supplement.

The educational stakes of this reframing are significant. If the phenomenon is understood as episodic error, the appropriate response is improved prompting and more vigilant fact-checking. However, if it is understood as a constitutive feature of the system’s generative logic, one that is largely unchanged under prompt refinement, then the response must be more fundamental, requiring a reconfiguration of how epistemic authority is recognised and validated. The stochastic alibi, on this account, is the mechanism through which the plausibility trap is sprung: it names the rhetorical performance that induces users to accept word-coherence as evidence of concept-coherence. The remainder of this article deploys the concept in exactly this role, tracing how the alibi operates differentially across knowledge traditions, how institutions can be complicit in accepting it, and how each dimension of the proposed framework functions to dismantle it.

The field of critical information literacy offers important theoretical resources for addressing the plausibility problem. Elmborg (2006) argued that information literacy should be understood not as a set of skills for locating and using information, but as a practice of critical consciousness; the ability to read the world through texts and to understand the social and institutional conditions that shape what counts as knowledge. This framing, drawing on Freirean pedagogy, insists that literacy is always already political and epistemological: it concerns not just access to information but the capacity to interrogate its conditions of production.

Within the field of AI literacy, this tradition reframes the central inquiry: the crucial issue is not simply whether learners can operate the tool, but whether they can critically interrogate its outputs and discern the epistemic status of the knowledge it appears to generate. Building on virtue epistemology’s concern with the dispositions that constitute responsible knowers (Zagzebski 2012), including intellectual humility before the limits of one’s own knowledge, epistemic vigilance towards sources that present themselves as authoritative, and calibrated trust that proportionally aligns confidence with evidential reliability, it becomes possible to theorise a third, insufficiently developed pillar of critical AI literacy: epistemic discernment.

While critical prompting optimises inputs, epistemic discernment interrogates outputs, shifting the evaluative gaze from tactical interaction to systemic critique across three interrelated dimensions. Firstly, source interrogation recognises that LLMs are not neutral retrieval mechanisms but value-laden artefacts whose outputs reflect the ideological provenance and material composition of their training corpus (Stewart & Rodgers 2025). Secondly, semantic depth analysis enables scholars to look past a model’s surface-level fluency (word-coherence) to rigorously evaluate its substantive engagement with evidence and argument (concept-coherence) (Bender & Koller 2020). Finally, algorithmic auditing exposes how training data can privilege dominant Western epistemologies while enforcing a structural erasure of non-dominant knowledge systems, a dynamic with acute consequences for scholars navigating African and other marginalised intellectual contexts (Bender et al. 2021; Essenfelder 2026).

Philosophy of knowledge offers a complementary framework through the lens of epistemic trust, the conditions under which we are warranted in accepting the testimony of others (Fricker 2007; Zagzebski 2012). The foundational insight of this literature is that trust must be dynamically calibrated against a source’s reliability, the stakes of the claim and the availability of independent verification. Large language models fundamentally disrupt this framework by mimicking the rhetorical markers of reliability while entirely lacking the constituent properties such as genuine knowledge, moral accountability and verifiable truth-seeking processes that ordinarily justify belief. Restated in the vocabulary of testimony, the stochastic alibi represents a structural deception: the system flawlessly performs the role of a trustworthy testifier while existing completely outside the relations of accountability that make testimony epistemically valid.

Fricker’s (2007) concept of epistemic injustice is also relevant, though in an unexpected way. Her account of testimonial injustice describes how credibility is unjustly deflated for knowers from marginalised groups. Large language models, by contrast, risk inflating credibility: their stylistic markers of authority, such as citations, disciplinary vocabulary, and confident tone, may persuade users who lack domain knowledge to accept outputs uncritically. In contexts where access to expert verification is uneven, as in many resource-constrained higher education environments, this inflation of credibility becomes especially hazardous. Artificial intelligence literacy must therefore address the differential epistemic vulnerabilities of learner populations, recognising that some are more exposed to the risks of misplaced trust than others.

Critically, epistemic discernment is not merely an individual cognitive skill. It is also, as Almufarreh and Arshad (2023) suggest, a disposition cultivated through engagement with AI as a mediational tool rather than an answer engine. When AI outputs are treated as first drafts for critical interrogation rather than final products for consumption, the epistemic relationship between learner and tool is fundamentally reoriented. This reorientation is as much pedagogical as it is cognitive, and it requires deliberate instructional design rather than the incidental development of good habits.

To translate the theoretical construct of epistemic discernment into pedagogical practice, verification literature provides concrete empirical models, most notably through Wineburg and McGrew’s (2019) framework of lateral reading. In analysing source credibility, their research contrasted the ‘vertical reading’ characteristic of historians and undergraduates, who remain confined within a document to evaluate its internal consistency and stylistic markers, with the lateral strategy deployed by professional fact-checkers. This lateral practice demands that evaluators immediately exit the target text to open external tabs, rapidly verifying the source’s provenance and claims across the broader information ecosystem before deeply engaging with the content itself. Crucially, the systemic reliability of this method stems not from specialised domain knowledge or general intelligence, but from its status as a highly transferable epistemic discipline that can be explicitly operationalised and explicitly taught.

The plausibility trap described in the previous section is structurally a variant of the vertical reading problem: users who engage with AI-generated text on its own terms, assessing it by its internal fluency, its apparent citations, its confident register, are performing the equivalent of reading vertically. They are evaluating word-coherence as a proxy for concept-coherence. What is required is a lateral turn: the practice of exiting the AI output and interrogating its claims against independent, verifiable sources before accepting them.

This task extends beyond the mechanical verification of citations, though that remains an indispensable first step given LLMs’ propensity to fabricate references that appear credible but do not exist (Alkaissi & McFarlane 2023). What is required is the cultivation of a broader intellectual disposition, one exemplified by the fact-checkers studied by Wineburg and McGrew. This disposition entails a default scepticism towards authoritative presentation, a disciplined habit of seeking independent corroboration, and a metacognitive sensitivity to the distinction between textual fluency and factual accuracy. Together, these qualities constitute what this article has termed epistemic discernment.

The pedagogical implication for AI literacy is clear: verification must be taught as a concrete behavioural practice rather than left as an abstract conceptual injunction. It is insufficient to exhort students to adopt a sceptical stance towards AI-generated outputs; what is required is systematic instruction in the operational mechanics of verification, tracing citations back to their source documents, employing database searches to cross-check factual claims, and discerning the absence of conceptual coherence beneath superficially fluent prose. These are not innate dispositions but learnable skills, and evidence from research on lateral reading indicates that they can be effectively acquired through relatively brief, targeted instructional interventions (Wineburg & McGrew 2019).

The concerns raised in the preceding sections acquire a particular urgency in the South African and broader African higher education context, where they intersect with a long-standing and contested debate about the epistemological character of the university. The decolonisation movement in South African higher education, galvanised by the #RhodesMustFall and #FeesMustFall protests of 2015 and 2016, raised fundamental questions about whose knowledge is legitimated, whose epistemological traditions are centred in curricula, and what access to knowledge means in a context shaped by colonial and apartheid legacies (Badat 2016; Heleta 2016). These questions are not resolved by the arrival of AI; they are deepened and complicated by it. The decoloniality literature provides the conceptual frame within which this deepening can be understood. Ndlovu-Gatsheni (2018) argues that the struggle for epistemic freedom in Africa is a struggle against the coloniality of knowledge itself: the enduring arrangement by which the Global North functions as the site of theory and the Global South as a site of data. Mignolo and Walsh (2018) similarly insist that coloniality persists in the very structures through which knowledge is produced and legitimated, and that delinking from those structures is a precondition for epistemic plurality. If these analyses are correct, then technologies trained overwhelmingly on Global North textual corpora are not epistemically neutral instruments arriving into a settled knowledge order; they are new carriers of an old asymmetry.

Heleta’s (2016) analysis, published in Transformation in Higher Education, argued that South African universities continue to reinforce the epistemic violence of Eurocentrism by treating Western knowledge traditions as the unmarked default of scholarly production. This argument has direct relevance to the question of AI literacy. Large language models are trained predominantly on English-language text produced within Western academic, journalistic and cultural contexts. The statistical regularities they learn, the patterns that constitute their capacity for word-coherence, are shaped by this provenance. When an LLM produces text that ‘sounds scholarly’, the rhetorical markers of scholarship it reproduces are those of a particular, geographically and historically situated intellectual tradition. Claims, framings and knowledge structures that align with this tradition will be generated with greater fluency and apparent confidence than those drawn from African, indigenous or other non-dominant epistemological traditions. In effect, the plausibility trap has a cultural topology: some knowledge traditions are more likely to be represented as plausible, others to be marginalised, distorted or absent. The stochastic alibi, correspondingly, is not culturally neutral: the credibility it performs is the credibility of a particular tradition, so that the alibi is most persuasive precisely where the knowledge it displaces is least represented in the training data.

This claim intersects with an emerging body of African and South African scholarship on AI and coloniality. Writing from South Africa, Adams (2021) asks whether AI can be decolonised, arguing that both AI systems and the ethics discourses surrounding them are embedded in colonial rationalities that reproduce racialised hierarchies of knowledge and personhood. Birhane (2020) characterises the uncritical importation of Western-developed algorithmic systems into African contexts as algorithmic colonisation, in which technologies encode the values and interests of their sites of production while presenting themselves as universal. This scholarship strengthens the article’s decolonial commitments by demonstrating that the concerns raised here are not speculative extrapolations from Northern critical theory but are being actively theorised from within African critical technology studies.

The illustrative examples that follow are hypotheses: theoretically motivated predictions, not verified empirical findings. Their hypothesised implications for African higher education extend beyond the factual unreliability of AI outputs to the deeper issue of epistemological misrepresentation. Large language models, trained predominantly on Global North corpora, may be expected to reproduce Western epistemic dominance while marginalising African scholarly traditions. This structural bias would risk reducing Ubuntu’s relational ontology to managerial slogans, flattening Ngũgĩ wa Thiong’o’s (1986) call to ‘decolonise the mind’ into generic diversity discourse, and erasing indigenous knowledge frameworks in agriculture and ecology by supplanting them with Western disciplinary equivalents. Such hypotheses are not idly formed: scholars have long warned that digital technologies reinforce epistemic inequalities in African higher education (Mhlongo et al. 2023; Nyamnjoh 2019), that Ubuntu pedagogy is vulnerable to trivialisation when detached from its philosophical grounding (Letseka 2013; Waghid & Smeyers 2012), and that indigenous knowledge systems face systematic erasure under Western paradigms (ed. Odora Hoppers 2002). Rigorously testing it through systematic prompting across current LLMs and analysis of outputs against the African scholarly record is among the most pressing items on the research agenda this article opens.

Maton’s (2014) Legitimation Code Theory (LCT) provides a productive conceptual vocabulary for analysing this problem. Legitimation Code Theory distinguishes between epistemic relations, which orient knowledge practices towards the properties and organisation of knowledge itself, and social relations, which establish the legitimacy of knowers within particular knowledge communities. Large language models instantiate a distinctive configuration of these relations: they reproduce the surface epistemic features of scholarly discourse, argumentative structure, citation conventions and rhetorical qualification, while remaining indifferent to the social relations through which knowledge is produced, scrutinised and authorised. As a consequence, they are structurally incapable of recognising or sustaining the accountability practices that confer epistemic warrant within specific disciplinary and cultural traditions. Their outputs may therefore achieve rhetorical credibility while lacking epistemic provenance, detached from the legitimating conditions that ground scholarly authority.

Two further concepts from LCT deepen this analysis: semantic gravity, the extent to which meaning is dependent on its context, and semantic density, the extent to which meaning is condensed within concepts, symbols and specialised terminology (Maton 2014). Powerful disciplinary knowledge characteristically unfolds through semantic waves, oscillating between dense theoretical abstraction and its unpacking into contextualised, empirically grounded particulars. Large language models can convincingly reproduce the surface markers of high semantic density by deploying sophisticated theoretical vocabulary with apparent fluency. However, these abstractions are frequently detached from the contextual anchoring, cumulative conceptual development and inferential relationships that constitute their disciplinary meaning. The result is text that performs abstraction without epistemic grounding: semantically dense in appearance but semantically diminished in substance. This sharpens the diagnosis of the plausibility trap while extending semantic depth analysis as a dimension of epistemic discernment. The critical question is not merely whether claims can be verified, but whether the abstractions invoked can be meaningfully unpacked into the contextual and empirical particulars from which their explanatory power derives.

Morrow’s (2009) distinction between physical access and epistemological access provides a valuable conceptual lens for situating the AI literacy debate. Physical access, the opportunity to enter an educational institution or engage with its resources, is a necessary but insufficient condition for meaningful learning if the epistemic practices of the institution remain inaccessible to those admitted. Epistemological access, by contrast, entails induction into the disciplinary norms, standards of justification, and modes of reasoning through which knowledge is generated, evaluated and legitimised. Within the context of AI, this distinction assumes renewed significance. Providing students with access to powerful generative AI tools without developing their capacity to interrogate, evaluate, and contextualise AI-generated outputs effectively grants access to the appearance of authoritative knowledge while withholding access to the epistemic practices required to assess its validity.

Algorithmic auditing therefore should be understood as more than a technical exercise aimed at detecting bias or correcting deficiencies in training data. It is equally a political and epistemological practice, concerned with interrogating whose knowledge is recognised, legitimised and amplified by AI systems, and whose knowledge is rendered invisible, marginal or unintelligible. Within African higher education, this reframing positions algorithmic auditing as a central dimension of epistemic discernment. The task is not simply to evaluate the accuracy or fairness of AI-generated outputs, but to examine the knowledge structures, epistemic assumptions and relations of authority that those outputs reproduce.

Most existing AI literacy frameworks are oriented towards individual learners, specifying what a person should know and be able to do. While this framing has value, it carries structural limitations. By locating the problem of AI credibility primarily with the individual user, it risks reproducing a logic in which the harms of AI misuse are attributed to inadequate personal skills rather than to the systemic features of the technologies themselves and the institutional conditions under which they are deployed.

Higher education institutions, however, are not neutral conduits for transmitting AI literacy competencies; they function as epistemic communities governed by distinctive standards, practices and forms of accountability (Barnett 1997). Assessment frameworks, research conventions and disciplinary norms together constitute systems of epistemic governance that, when recalibrated, provide institutional scaffolding beyond what individual competencies alone can sustain. When institutions require students to verify sources, model critical engagement with evidence, and design assessments that demand genuine epistemic reasoning rather than AI-facilitated reproduction, they enact a collective role in epistemic governance that transcends the remit of any singular literacy framework.

Institutional responses to generative AI have often been framed primarily through the lens of academic integrity: concerned with whether students use AI to circumvent assessment rather than with whether they develop the epistemological tools to engage with it critically (Essenfelder 2026). While integrity concerns are legitimate, they are epistemologically secondary. A student who incorporates AI outputs uncritically into their thinking, yet refrains from presenting them as original work, remains epistemically vulnerable even if not in breach of formal policies. Indeed, an institution that polices attribution while leaving the stochastic alibi unexamined effectively accepts the alibi at scale: it treats AI-generated text as credible knowledge whose only problem is ownership. Institutional responsibility must therefore extend beyond policing, towards cultivating capacities for critical evaluation, shifting emphasis from managing AI as a threat to honesty to interrogating AI-mediated knowledge production as an epistemic practice.

An argument of this kind gains analytical force by engaging directly with its strongest objections. Four merit explicit consideration.

The first is that the critique is uncharitable: prevailing AI literacy frameworks do, in fact, include evaluation, and the article therefore attacks a straw target. This objection has already been anticipated but warrants restatement. The article neither denies nor minimises the centrality of evaluation; indeed, it explicitly acknowledges that evaluation is a recognised competency across the most influential frameworks. The critique concerns not its omission but its epistemological underspecification. While evaluation is consistently named as a desirable capability, the frameworks examined rarely theorise what evaluation entails when its object is text generated by a system optimised for probabilistic coherence rather than epistemic warrant. The distinction between identifying a competency and specifying its epistemic content is precisely the conceptual gap this article seeks to address.

The second objection is technological: advances in retrieval-augmented generation, grounded citation and model alignment will progressively reduce fabrication, rendering the epistemological critique obsolete. This objection merits serious consideration, and meaningful progress in these areas is already evident. Yet the argument advanced here does not depend on any particular rate of error or fabrication. Its concern is structural rather than empirical. So long as generative AI systems produce text through statistical patterning rather than participation in the social and epistemic accountability practices through which knowledge is created, scrutinised and authorised, their outputs will continue to perform epistemic authority that cannot be established from the text itself. Even retrieval-augmented systems remain contingent on the quality of retrieval, the selection of sources, and the fidelity with which retrieved material is synthesised and represented. Ironically, improvements in reliability may intensify rather than diminish the epistemological challenge. As fabrication becomes less frequent, users are more likely to generalise trust, making the stochastic alibi increasingly persuasive precisely because it fails less often. Epistemic discernment is therefore required not because AI systems are invariably unreliable, but because the grounds of their reliability remain opaque and cannot be inferred from the rhetorical surface of their outputs.

The third objection posits that epistemic discernment is merely critical thinking under a different label, and that universities already cultivate this capacity. The response is that while generic critical thinking is indispensable, it is no longer sufficient. The capacities advanced here of recognising the architecture that produces lexical coherence without conceptual coherence, tracing citations to authoritative and indexed sources, interrogating semantic depth, and auditing outputs for the cultural topology of their training data all require both an understanding of how LLMs generate text and a repertoire of verification practices that conventional critical thinking curricula seldom address. The distinction is therefore not one of intellectual sophistication but of epistemic specification. As Wineburg and McGrew (2019) demonstrate, expert historians, highly accomplished critical thinkers, were consistently outperformed by professional fact-checkers, not because they lacked analytical ability, but because they employed a distinct, teachable verification practice. Epistemic discernment, as theorised here, should therefore be understood not as a replacement for critical thinking, but as its recalibration for a new epistemic object: text that is rhetorically coherent yet produced through probabilistic prediction rather than participation in practices of knowledge production and epistemic accountability.

The fourth objection comes from within the transformation agenda itself: that emphasising scepticism towards AI risks denying historically disadvantaged students the access benefits these tools provide, including access to academic registers from which they have been excluded. This objection identifies a genuine tension, and the article’s position is not AI-rejectionist. The framework insists on coupling access with interrogation rather than choosing between them. What would genuinely disadvantage marginalised students is not the demand for discernment but its absence: students with the least access to expert verification networks are precisely those most exposed to the stochastic alibi, and most harmed when it fails them. Equipping them to see through the alibi is an equity intervention, not an obstacle to one.

The preceding analysis underscores the need for a reconceptualised framework of AI literacy grounded in epistemology rather than mere operational proficiency. At minimum, such a framework should encompass four interrelated dimensions, each aligned with the theoretical strands outlined above. The four dimensions are neither hierarchical nor strictly sequential; they are mutually constitutive. Conceptual understanding without verification practice is inert; verification practice without conceptual understanding is mechanical; both remain fragile unless consolidated into the disposition of epistemic discernment; and none of the three individual capacities can be reliably cultivated or exercised outside institutional conditions that require, model and reward them. Equally, institutional epistemics without individually capable members is an empty policy shell. Table 1 summarises the four dimensions, their theoretical sources, and their principal pedagogical and institutional implications.

The first dimension is conceptual understanding of generative AI architecture. This is not a call for universal computer science education, but for a conceptually adequate grasp of how LLMs generate text. Learners must understand that outputs are calibrated for plausibility rather than truth, and that word-coherence is not evidence of concept-coherence. Recognising the plausibility trap is the precondition for epistemic discernment: one cannot adopt appropriate scepticism towards a mechanism one does not understand. Nor can one resist the stochastic alibi without recognising that it is structural rather than episodic. Luckin et al. (2016) argue that effective AI-supported education requires learners to know what AI can and cannot do. This article extends that claim: such knowledge must be epistemological, oriented towards what AI outputs can and cannot know, not merely functional.

The second dimension concerns verification practices, operationalised through the method of lateral reading. AI literacy should encompass explicit, discipline-specific instruction in strategies for evaluating AI-generated content. These include consulting external sources to corroborate claims, tracing citations back to their original entries in indexed databases, and conducting semantic depth analysis to interrogate whether fluent-sounding assertions are genuinely supported by evidence. Vygotsky’s (1978) concept of the zone of proximal development, the interval between what a learner can accomplish independently and what they can achieve with guidance, provides a useful theoretical lens. The cultivation of verification practices is quintessentially a matter of scaffolded instruction within this zone. In this process, instructors and curricula serve a pivotal scaffolding function, guiding students from prompted verification towards the habitual scepticism that underpins mature academic inquiry.

The third dimension is epistemic discernment, cultivated as a disposition through what may be termed prompt criticism: the systematic and analytical examination of AI outputs as artefacts whose conditions of production are integral to their evaluation (Essenfelder 2026). This is not reducible to generic critical thinking, as argued in the engagement with counter-positions above. It demands specific knowledge of LLMs, targeted verification skills, and the metacognitive awareness to recognise the limits of one’s own expertise as a verification resource. It is, in effect, the internalised refusal of the stochastic alibi: the settled disposition to treat performed credibility as a claim requiring warrant rather than as warrant itself.

The fourth dimension is institutional epistemics, with particular salience for South Africa’s transformation agenda. Institutions committed to epistemic transformation must ask a pointed question: does their adoption of AI advance or undermine decolonisation? Uncritical deployment, without attention to the provenance and cultural topology of training data, risks reinstating the epistemic hierarchies that transformation seeks to dismantle; it amounts to accepting the stochastic alibi as institutional policy. Algorithmic auditing must therefore be an institutional responsibility, not an individual one. Concretely, this means developing procurement criteria that require transparency about training data composition. It means involving African-language scholars and indigenous knowledge specialists in the evaluation of AI tools before institutional endorsement. It means aligning AI use policies with existing decolonial curriculum commitments rather than treating them as separate administrative domains. In the South African context, AI literacy is not a neutral technical exercise. It coincides with the larger project of building an epistemically just university, in the sense theorised by the decoloniality scholarship discussed above (Adams 2021; Ndlovu-Gatsheni 2018), and must be governed as such.

These recommendations are offered at the level of principle, and intellectual honesty requires acknowledging the gap between principle and practice. A sceptical reader will rightly ask: how does lateral reading instruction function in a first-year cohort of 500 students when library and academic development resources are stretched? What happens when instructors themselves lack the disciplinary expertise to model citation verification in unfamiliar fields? And what does institutional procurement reform actually look like when most South African universities are negotiating with the same limited pool of vendors with limited institutional bargaining power? These are genuine constraints, and they do not dissolve at the level of framework. This article does not pretend to resolve them. It argues, rather, that the epistemological case for addressing them is established, and that the next phase of work, empirical, curricular and political, must grapple with the structural conditions under which these principles can be operationalised at scale.

This article has argued that the core of this problem is the plausibility trap: the structural conflation of word-coherence with concept-coherence that is endemic to LLM outputs. It has proposed the stochastic alibi as a conceptual redescription of what the technical literature terms ‘hallucination’. Addressing this problem requires a reconceptualised AI literacy centred on epistemic discernment: a theoretically grounded, practically enacted capacity for source interrogation, lateral verification, semantic depth analysis, and algorithmic auditing, cultivated both in individual learners and through the institutional practices of the university.

It is, however, the transformation argument that gives this reconceptualisation its sharpest edge, and it belongs at the centre of the conclusion rather than its margins. The plausibility trap, this article has argued, has a cultural topology: the credibility that LLM outputs perform is the credibility of the Global North traditions on which they were trained, so that the stochastic alibi operates with greatest force precisely where African, indigenous and other non-dominant knowledge traditions are least represented. The decolonisation scholarship of Heleta (2016), Ndlovu-Gatsheni (2018), and Adams (2021) converges on the implication: uncritically adopted AI systems are new carriers of an old coloniality of knowledge, and epistemic discernment is therefore not a generic study skill but a condition of epistemic freedom. For universities committed to transformation, the cultivation of discernment and the institutional auditing of AI systems are continuous with, not adjacent to, the decolonial project. Equipping learners to navigate the AI-saturated knowledge environment of the contemporary university requires more than teaching them to prompt better. It requires teaching them to know better: to distinguish the plausible from the accurate, to insist on that distinction under pressure, and to do so within institutional communities that model and reward such insistence.

Such a reconceptualisation shifts the horizon of AI literacy from managing AI to critiquing AI-mediated knowledge production. It repositions the university not as a site of anxious adaptation to technological change, but as an epistemic community with particular responsibilities to cultivate the capacities for rigorous, sceptical and independent thought that the current moment demands. In higher education contexts committed to transformation, to building universities that are genuinely epistemically just, this is not a peripheral concern. It is among the most fundamental ones.

The framework advanced here is deliberately theoretical: it specifies and justifies the epistemological requirements of an adequate AI literacy, rather than prescribing its large-scale implementation. The research agenda it inaugurates is correspondingly substantial. It includes empirical investigation of the cultural-topology hypothesis; specifically, whether and how LLMs systematically misrepresent non-dominant knowledge traditions in ways that can be documented and compared across systems. It calls for the design and evaluation of pedagogical interventions rooted in lateral reading and epistemic discernment, tested within African higher education contexts and attuned to their distinctive resource and institutional conditions. It also requires institutional research into the curricular, structural, and political circumstances under which epistemologically adequate AI literacy can be genuinely institutionalised, rather than merely articulated as a policy aspiration. The argument advanced here is that such work is not ancillary to the AI literacy project in higher education; it constitutes its necessary next chapter.