Aware but not verifying: media literacy, verification behaviour, and misinformation resilience at three universities in Somalia

Introduction

Digital media have transformed how information is produced, circulated, and trusted, and the transformation has not been uniformly benign. Alongside unprecedented connectivity has come what Wardle and Derakhshan (2017) term information disorder: the circulation of misinformation, disinformation, and malinformation at a speed and scale that legacy gatekeeping institutions were never designed to manage. The World Economic Forum (2024) now ranks misinformation and disinformation as the most severe short-term global risk. Young people, who adopt new platforms fastest, are among the most exposed, and international policy has responded by positioning media and information literacy (MIL) as a core twenty-first-century competency (Grizzle et al., 2013; UNESCO, 2013).

The case for treating media literacy as a defence rather than a palliative rests on two bodies of evidence. First, a meta-analysis of 51 intervention studies found that media literacy education produces positive effects on knowledge, critical evaluation of messages, and related attitudes and behaviours (Jeong et al., 2012). Second, cognitive research indicates that susceptibility to false content is driven less by partisan motivation than by a failure to engage analytic reasoning at all, together with the familiarity that repeated exposure confers (Pennycook et al., 2018; Pennycook and Rand, 2019). The decade since has produced a much larger body of intervention evidence, recently consolidated by Kozyreva et al. (2024), and its lesson is narrower than the original case: what works tends to change what a user does at the moment of encounter rather than what the user knows in general. If the problem is largely one of unreflective processing, then skills and habits that prompt reflection are a plausible remedy, and universities, where a country's future professionals and public officials are concentrated, are a natural site for building them.

Somalia presents an unusually demanding test of that logic. Three risk factors converge. State institutions remain fragile after three decades of conflict, and since the collapse of central government in 1991 clan and kinship networks have functioned as the principal channels of both authority and information (Chonka, 2019). Deep social divisions mean that a single rumour framed through clan identity can reactivate old grievances, and practitioners working in the country have warned that fabricated content routinely outpaces the capacity to correct it (Wahlberg, 2017). Meanwhile digital access has expanded very quickly: by January 2025 Somalia had roughly 10.7 million internet users, about 55% of the population (Kemp, 2025). That expansion has outrun the development of verification infrastructure. Fact-checking capacity is limited to a small number of organisations, and Somali is a low-resource language for which automated detection tools barely exist (Mohamed et al., 2025). The result is a setting in which smartphones and connectivity are widespread, critical engagement with content is thin, and the social cost of error is high.

Despite this, empirical research on media literacy in Somalia is almost non-existent. The small literature on the Somali information environment has concentrated on news media and political contestation (Chonka, 2019), on the use of social media by conflict actors (Cooley and Jones, 2013), and more recently on computational detection of Somali-language fake news (Mohamed et al., 2025). No published study has measured the media literacy competencies of Somali university students, examined how those competencies relate to the capacity to resist misinformation, or tested whether students’ self-assessments bear any relationship to what they can actually do when confronted with a misleading post. This article addresses that gap.

Drawing on a survey of 393 students at three universities in the country's three main regions and on interviews and a performance task with 15 of them, the study asks three questions. RQ1: What is the present level and profile of media literacy among Somali university students? RQ2: How is media literacy related to students’ competency in identifying and resisting misinformation? RQ3: What contextual factors shape how misinformation spreads among these students, and what do students themselves recommend? The quantitative strand is descriptive and exploratory rather than inferential, and the three universities were selected purposively, so the patterns reported here describe the participating students rather than Somali university students as a population. The central pattern observed in this sample is an awareness–action gap: students know that misinformation is harmful and believe themselves reasonably able to detect it, but their verification behaviour is weak and their demonstrated discernment weaker still. We argue that this pattern, and the clan-mediated credibility judgements that appear to accompany it, provide grounds for reconsidering how media literacy programming in Somalia is oriented, away from awareness-raising and towards the formation of verification habits.

Literature review and theoretical framework

Information disorder and its cognitive foundations

Wardle and Derakhshan's (2017) typology distinguishes misinformation (fake content shared without intent to harm), disinformation (fake content created and shared to cause harm), and malinformation (genuine content weaponised to damage a person or group). Tandoc et al. (2018) show that the popular label “fake news” collapses these categories along with satire, propaganda, and manipulated content, and they recommend that researchers specify both the facticity of a message and the intent behind it. We follow that advice: the survey items used here address misinformation broadly, while the performance task deliberately spans fabricated claims, decontextualised imagery, health rumours, and identity-targeted content.

The consequences of information disorder extend beyond individual error. Exposure to untrustworthy content erodes trust in institutions and in the news media (Guess et al., 2020b; Stubenvoll et al., 2021) and, in already divided societies, sharpens polarisation and makes collective decision-making harder (Lewandowsky et al., 2017). Lazer et al. (2018) argue that responses must operate at two levels: empowering individuals to evaluate content and changing the structures that reward its spread. Platform architectures that prioritise engagement over accuracy, and the echo chambers they help to produce, are part of the structural problem (Tandoc et al., 2018).

At the individual level, the cognitive literature offers a consistent picture. People preferentially accept information congruent with prior beliefs (Nickerson, 1998), and mere repetition increases the perceived accuracy of false headlines (Pennycook et al., 2018). Yet Pennycook and Rand (2019) find that analytic thinking predicts better discernment regardless of whether a headline flatters one's own side, which implies that the primary deficit is not motivated reasoning but the absence of reasoning. Identity remains important, however: Efstratiou and De Cristofaro (2022) synthesise evidence that group-based processes, including the desire to signal belonging, shape both belief in and sharing of misinformation, and Van Bavel and Pereira (2018) describe how group identity can override accuracy goals when the two conflict. These two strands, low analytic engagement and identity-driven acceptance, provide the interpretive frame for our findings.

Work published since 2022 has sharpened both strands and added a third. Ecker et al. (2022) separate the cognitive drivers of misinformation belief—fluency, repetition, source cues, emotional arousal—from the social and affective ones, and show that both survive correction, which is why resistance to correction rather than exposure itself has become the design problem for interventions. The third addition is modality. Weikmann and Lecheler (2023), reviewing a literature that had until recently treated falsehood as a textual phenomenon, argue that visual disinformation differs from its textual equivalent in production, processing and effect: images carry a presumption of evidence, are judged faster, and are remembered longer, and the cheapest form of visual falsehood is not a deepfake but an authentic photograph stripped of its date and place. That point matters here for an empirical reason. The decontextualised image was one of the two items our participants judged least accurately, and it was the item on which almost none of them attempted any check at all.

Media literacy as competency and habit

Media literacy is most commonly defined as the ability to access, analyse, evaluate, and create messages across media forms (Hobbs, 2010; Livingstone, 2004). Potter (2004) frames it cognitively, as a set of knowledge structures and skills that allow individuals to exert control over how media messages affect them, and later notes how fragmented the field's definitions have become (Potter, 2010). Kellner and Share (2007) push the concept further, insisting that a critical media literacy must interrogate power and representation rather than merely teach technical competence. Buckingham (2015) makes the point that matters most for the present study: literacy is not a stock of knowledge but a practice, and young people need the four key concepts of media education (representation, language, production, audience) applied habitually to the digital environments they actually inhabit.

The distinction between knowing and doing recurs throughout the intervention literature. Jeong et al. (2012) report that media literacy programmes reliably change knowledge and attitudes, but behavioural effects are smaller and depend on programme design. McDougall et al. (2023) argue that evaluations must therefore track behavioural change, not only knowledge gains, if they are to demonstrate resilience. This is the premise on which we treat verification behaviour, rather than awareness, as the outcome that matters.

That premise now has a good deal more evidence behind it than it did when the meta-analytic case was first made. Kozyreva et al. (2024) sort individual-level countermeasures into nine families and find that the better-evidenced ones share a form rather than a content: they act on the moment of encounter. Accuracy prompts are the clearest instance. Pennycook et al. (2021) showed across survey experiments and a field experiment on Twitter that briefly drawing a user's attention to accuracy improves the quality of what that user shares next, and Pennycook and Rand (2022) read the effect as evidence that people share carelessly rather than indifferently, reporting an improvement in sharing discernment of roughly 72% across twenty experiments. Psychological inoculation applies the same logic before the encounter rather than during it; Roozenbeek et al. (2022) found that short videos teaching common manipulation techniques improved recognition of those techniques among YouTube users under field conditions.

Two qualifications keep this literature from settling the question for a setting such as Somalia. The first is transfer. Guess et al. (2020a) tested a set of digital media literacy tips in the United States and in India and found improved discernment in both, but the Indian effect did not persist and was absent among the rural respondents most exposed to misinformation in the first place. An intervention that works where fact-checking infrastructure, search habits and platform design are already favourable cannot be assumed to work where they are not. The second is measurement. Maertens et al. (2024), introducing a validated performance measure of veracity discernment, observe that this field has leaned heavily on improvised self-report scales of uncertain psychometric quality, and insist that what a person believes they can do and what they demonstrably do are two quantities requiring two instruments. That insistence is the methodological premise of the performance task reported below.

Evidence from Sub-Saharan Africa and Somalia

Much of the media literacy canon was built in North America and Western Europe and travels imperfectly to contexts with different media systems, infrastructures, and educational histories (Grizzle et al., 2013). African research has begun to fill the gap and points to a pattern directly relevant here. In a two-wave survey across Kenya, Nigeria, and South Africa, Wasserman and Madrid-Morales (2019) found high perceived exposure to disinformation alongside low trust in media, with many respondents admitting to having shared content they later learned was false. Focus groups with university students in six sub-Saharan countries showed that students frequently share misinformation knowingly, for reasons of civic duty, humour, or social standing (Madrid-Morales et al., 2021). Among Ghanaian and Nigerian students, most “verification” consisted of asking peers or checking other social media rather than using fact-checking resources, of which few were aware (Adjin-Tettey and Amenaghawon, 2024). A quasi-experiment in Nigeria found that awareness campaigns raised intentions to verify, which implies that baseline awareness alone was not sufficient to produce the behaviour (Apuke et al., 2023). Mustapha et al. (2024) put a name to the underlying misconception. Surveying students across Anglophone West Africa, they found that media literacy and critical thinking were understood as tools to be picked up when a suspicious post appeared rather than as dispositions exercised continuously, and they treat that understanding as itself an obstacle to countering misinformation. It is close to the declarative–procedural distinction we draw from our own data. Studies of information literacy in Ghanaian universities similarly document low use of library and database resources despite their availability (Yebowaah and Sanche, 2021). Across these studies the same awareness–behaviour disjunction appears that we document in Somalia.

Post-conflict settings add a further layer. Reporting on South Sudan, the Sentinel Project's rumour-verification work found that misinformation was routinely weaponised along ethnic lines and that literacy efforts had to be tied explicitly to intergroup trust if they were to reduce harm (Otunga, 2020). In Somalia, Chonka (2019) describes a transnational Somali digital public in which online news outlets are deeply entangled with political and clan contestation, and Cooley and Jones (2013) show how early social media use in the country was shaped by conflict actors. Recent computational work confirms that Somali-language fake news and toxic content circulate widely on social platforms and that detection is hampered by the language's low-resource status (Mohamed et al., 2025). What is missing from this literature is an audience-side account: how Somali users, and particularly the educated young, actually evaluate what they see.

Two features of the Somali environment make that account especially necessary. The first is the speed of the digital transition relative to institutional development. Internet penetration has passed the halfway mark (Kemp, 2025) in a country whose formal media regulation, professional journalism training, and fact-checking capacity remain thin, so that the gatekeeping functions journalism performs elsewhere are weak precisely where the volume of unverified content is growing fastest (Wahlberg, 2017). The second is the way information authority is distributed. Because the state has for long periods been absent or contested, credibility has been vested in persons and lineages rather than in institutions, and news travels through networks in which the identity of the messenger has traditionally certified the message (Chonka, 2019). Social platforms have not displaced these networks; they have digitised them. A WhatsApp group of kin, or a Facebook page associated with a particular political and clan constituency, reproduces online the trust structures that governed oral communication, with the difference that content can now be fabricated, decontextualised, and forwarded to thousands in minutes. Understanding how young, educated Somalis evaluate information therefore requires attention not only to their cognitive skills but to the social criteria of credibility they have inherited.

Theoretical framework

Defining and distinguishing the constructs

This article uses four closely related terms, and they are not interchangeable. Because one of them appears in the title, we set out what each means, which measure carries it, and where the boundaries between them are soft.

Media literacy is the general, domain-independent capacity to access, analyse, evaluate and create messages across media forms (Hobbs, 2010; Livingstone, 2004; Potter, 2004). It is indexed here by the Media Literacy Index (MLI), a self-report measure.

Misinformation-fighting competency is its domain-specific counterpart: self-assessed awareness of misinformation as a phenomenon, understanding of how it circulates and what it does socially, and a disposition to participate responsibly. It is indexed by the Misinformation-Fighting Competency Index (MFCI), also self-report. Both indices record perceived capability.

Misinformation resistance is what a person actually does when a specific item is in front of them: the accuracy of the veracity judgement and the checking, if any, that precedes it. In this study it is measured behaviourally, by the five-item performance task, and deliberately not by self-report.

Misinformation resilience is the higher-order outcome: a durable capacity to meet repeated exposure to false content over time without believing it or passing it on. Resilience is not a single measurement. It is a property that emerges from competency, from resistance exercised often enough to become habitual, and from environmental supports—verification tools in the user's own language, credible correction, institutions worth trusting—that make the habit sustainable. Establishing it therefore calls for designs that observe durability across time or across repeated challenge, as the inoculation and discernment literatures do (Kozyreva et al., 2024; Maertens et al., 2024; Roozenbeek et al., 2022). This study uses no such design. We retain the term in its framing sense, as the outcome the measured components feed into and the standard against which a media literacy curriculum should eventually be judged, and we state plainly what is measured instead: reported competency and demonstrated resistance, both at a single point in time. Readers should treat the study's contribution to the resilience literature as diagnostic—a description of the starting condition—rather than as an assessment of resilience itself.

Two consequences follow for measurement. The first concerns overlap. The MLI and MFCI are separated by scope rather than by subject matter: the MLI asks about evaluation and sharing across media generally, the MFCI about misinformation in particular. Several MLI items nonetheless refer to verifying before sharing and to telling reliable from unreliable information, which is MFCI territory, and no design-stage procedure was used to remove the overlap. It is a plausible contributor, alongside common-method variance, to the size of the correlation reported under RQ2. Since the factor structure of neither index was examined, discriminant validity cannot be demonstrated here and both are treated as exploratory composites. The second consequence is that the MFCI cannot stand in for resistance or resilience. It records what students believe they can do, and the performance task shows that this is a different quantity from what they did.

How the two frameworks are used

Two theories organise the analysis, and they answer different questions. Media literacy theory supplies the competency model. Following Livingstone (2004), Hobbs (2010), and Potter (2004), we treat media literacy as a set of access, analysis, evaluation, and production competencies, and we distinguish general media literacy from the more specific competency of identifying and resisting misinformation. The former is a necessary but, we hypothesise, insufficient condition for the latter. Its contribution to this study is concrete rather than decorative: the three domains of the MLI—access, analysis and evaluation, production and responsible sharing—are taken directly from this tradition, and Buckingham's (2015) argument that literacy is a practice rather than a possession supplies the analytic move on which the findings turn, the separation of declarative from procedural competence. RQ1 and the gradient reported in Table 1 derive from this framework, as does the expectation, examined descriptively under RQ2, that general literacy and domain-specific competency should be related.

Table 1. Media literacy and misinformation awareness indicators, ranked (N = 393; selected indicators from the MLI and MFCI).

Scale: 1 = strongly disagree to 5 = strongly agree. Labels in the Domain column describe the competency each item taps; they are descriptive and do not correspond to the three domains used to construct the MLI.

Uses and gratifications theory (Katz et al., 1973) supplies the behavioural lens. Audiences select media actively to satisfy informational, social, and identity needs. In the Somali context this explains why students gravitate to particular platforms, why speed and social affirmation can outweigh accuracy as gratifications, and why platform choice itself structures exposure to misinformation. It supplies what a competency model cannot, namely a reason why a capability might go unused, and it generates a second set of study elements: the attention to platform differentiation, the questions about purposes of use and language of consumption, and the treatment of sharing as a socially rewarded act. RQ3, the platform patterns in Table 2, and the forwarding-culture and clan themes reported under it all derive from this framework rather than from the first.

Table 2. Media access and usage patterns (N = 393; multiple responses permitted for platforms and outlets).

The two are complementary, and their complementarity is the study's argument in miniature. Media literacy theory specifies what a student is able to do; uses and gratifications theory specifies whether, in the conditions of a particular platform at a particular moment, that ability is exercised. A student may hold every evaluative competency the first framework describes and still not use it, because the gratifications the second describes—being first with the news, being seen to be informed, affirming one's own people—arrive faster when nobody pauses to check. Taken together the frameworks anticipate a gap rather than a deficit: not that students lack knowledge, but that knowledge and action come apart wherever the reward structure favours speed over scrutiny. That is the expectation this study examines. Media literacy should be associated with misinformation-fighting competency; both should exceed demonstrated resistance; and the excess should be widest where identity-based trust and speed-based gratification are strongest. The expectation is examined descriptively. Neither framework, and no analysis reported here, licenses a causal claim.

Methods

Design

The study used a convergent mixed-methods design with an embedded behavioural component (Creswell and Plano Clark, 2018), in which a cross-sectional survey provided breadth and semi-structured interviews, supplemented by a performance-based verification task, provided depth and a behavioural check on self-report. Quantitative and qualitative strands were collected in the same period and integrated at the interpretation stage, with interview material used to explain survey patterns and the performance task used as a behavioural check on self-assessed competency. The survey used non-probability, purposive sampling and was administered at a single time point; the quantitative strand is therefore treated throughout as descriptive and exploratory, and no causal or population-level inference is attempted.

That classification needs defending, because the qualitative strand is used in the Findings and the Discussion to explain and elaborate what the survey showed, which is the signature of an explanatory sequential design. Three features place the study on the convergent side of the line. Timing is the first: survey, interviews and performance task were fielded in the same period, and interview recruitment was not conditioned on survey results. Independence of instrument design is the second: the interview guide and the five posts used in the performance task were specified before any survey analysis began, so the qualitative instruments were not derived from quantitative findings in the way an explanatory sequential design requires. The point of integration is the third and the most important: the strands were analysed separately and brought together only at interpretation, where agreement and disagreement between them were examined and meta-inferences drawn.

Two qualifications keep the description honest. During the review of candidate themes, the survey's weakest indicators were used as sensitising concepts, which introduces a sequential element at the analysis stage even though data collection was concurrent. And the performance task is properly an embedded component nested inside the qualitative strand rather than a third strand of its own: it was administered only to the 15 interviewees and serves a supplementary, validity-checking function. We therefore describe the design as convergent with an embedded behavioural component and one partially sequential analytic step, rather than claiming pure concurrence. Integration itself proceeded in two ways: each qualitative theme was mapped onto the survey indicator it bore upon, a mapping set out in the Findings under RQ3 and, for the students’ own proposals, summarised in Table 3; and, for the 15 participants who appear in both strands, self-rated confidence was compared directly against performance-task accuracy.

Table 3. Interviewees’ recommendations for universities (N = 15).

Setting and participants

Three universities were purposively selected to span Somalia's three principal regions and political-administrative contexts: Somali National University in Mogadishu (south-central Somalia), the University of Hargeisa (Somaliland, north-west), and East African University in Garowe (Puntland, north-east). The Federation of Somalia Universities (2025) lists around 100 institutions nationally; the three selected are the largest public or established universities in their respective regions and concentrate digitally engaged student populations. Their combined enrolment, which constituted the accessible population (Asiamah et al., 2017), was approximately 23,750 students (Somali National University about 16,000; University of Hargeisa about 7,000; East African University Garowe about 750).

The target size of the survey sample was set by benchmarking against the conventional requirement for a population of this magnitude. For an accessible population of roughly 23,750, a 95% confidence level and a ±5% margin imply a minimum of 379 responses under Cochran’s (1977) formula with a finite-population correction, and 378 under the equivalent table published by Krejcie and Morgan (1970); the 393 responses obtained exceed both. We want to be exact about what this calculation does and does not establish. It shows that the achieved sample is large enough to produce stable descriptive estimates for the students who took part and to support the proportional allocation described next. It does not confer the properties that a confidence level and a margin of error ordinarily denote, since those presuppose probability selection, which this study did not use. No margin of error is reported for any estimate in this article. Allocation across the three universities followed enrolment share, which is also why the Garowe sub-sample is so small in absolute terms: proportional allocation reproduces the disparity in institutional size rather than correcting it, and East African University Garowe accounts for about 3% of the accessible population.

Purposive sampling (Etikan et al., 2016) was used at the institutional level, with the sample of 393 allocated across the three universities in proportion to enrolment (271, 110, and 12 respectively) and, within each university, across faculties in proportion to their size. Within faculties, students on official enrolment lists were invited to participate until each quota was met. The design therefore aims at cross-institutional and cross-disciplinary coverage rather than at probabilistic generalisation to all Somali students. Although the three sites occupy important regional and political-administrative positions, they were not drawn at random from the approximately 100 institutions nationally and cannot be considered statistically representative of Somali universities or of Somali university students as a population; nor were individual respondents selected by any probability mechanism. The very small Garowe sub-sample (n = 12) in particular precludes institution-level statistical inference for that site, and results for that subgroup are reported as descriptive and exploratory only. No sampling weights were applied, and weighting would not in any case compensate for the size of the Garowe subgroup or for the purposive selection of institutions.

For the qualitative strand, 15 survey respondents (8 men and 7 women, aged 20–31) were purposively selected to span the three universities, all four year-groups, and eleven fields of study (Table 4). Interviewees are identified by letter codes throughout. Selection followed a sampling matrix that crossed institution, year-group and field, so that the qualitative sample reproduced the structural variation present in the survey rather than its numerical proportions.

Table 4. Profile of interview participants (N = 15).

Instruments

Survey

The questionnaire, administered in English via Google Forms in January 2026, had four parts. Part I collected demographic information (university, gender, age group, year of study, field). Part II recorded media access and use: platforms and outlets used, frequency of news checking, daily time on social media, main purposes of use, and language of consumption. Part III measured media literacy with Likert-type items (1 = strongly disagree to 5 = strongly agree) organised around three competency domains drawn from the media literacy framework: access to information (for example, “Instead of depending on just one source, I actively look for information from various media outlets”); analysis and evaluation (“I can distinguish between reliable information and misinformation”); and production and responsible sharing (“I verify information before sharing it on social media”). Twelve items were retained to form the Media Literacy Index (MLI), which showed good internal consistency (Cronbach's α = .84; mean correlation correlation = .32). Part IV measured misinformation-fighting competency with nine items covering awareness of misinformation in Somalia (“I am aware of the websites that are most likely to spread fake news”), understanding of its mechanics and social effects, and responsible participation; these were averaged to form the Misinformation-Fighting Competency Index (MFCI).

The two indices are separated by scope rather than by subject matter, as set out above, and some MLI items therefore overlap with MFCI content. Internal consistency was computed for the MLI (α = .84) but not for the MFCI, and the dimensionality of neither index was examined. Both are accordingly reported as exploratory composites, and the MFCI in particular, which combines awareness, knowledge and behavioural items, should not be read as a validated unidimensional scale.

Language of administration

The questionnaire was administered only in English, and the reasons are worth stating because the choice has consequences the results themselves expose. English is the medium of instruction and assessment at all three participating universities, so every respondent was studying in English at the time of the survey and the register of the items matched the register of their coursework. A single English version also held item wording constant across three sites that differ in administrative context and in conventions of written Somali, which a hastily produced translation would not have done. And there is at present no validated Somali-language instrument for media literacy or misinformation competency from which items could have been adapted.

None of that amounts to evidence that the items were understood as intended. Respondents’ English proficiency was assumed from their enrolment rather than assessed; the items were not cognitively pretested or formally piloted; and no Somali version was produced, so no translation and back-translation procedure (Brislin, 1970) was carried out. The risk this creates is specific, and one of the study's own findings sharpens it. Somali was the dominant language of media consumption for half the sample (Table 2), so for many respondents the questionnaire asked in one language about practices conducted in another, and credibility, source evaluation and verification are precisely the vocabulary in which a proficiency gap would surface. Differential comprehension cannot be ruled out, and it may have contributed to the dispersion on the behavioural items, which carry the widest standard deviations in the instrument. We report this as a limitation of the design rather than a feature of it. A Somali-language instrument developed through translation, back-translation and cognitive pretesting is the first methodological priority we would set for further work in this setting.

Interviews

A semi-structured guide covered everyday media use, criteria for trusting sources, experience of comparing sources, understanding of misinformation, explanations for its spread in Somalia, its perceived social effects, and recommendations for universities. Interviews lasted 40–60 min, were conducted online, and were audio-recorded and transcribed with consent. Recruitment stopped when the sampling matrix described above was filled. We do not claim theoretical saturation, which fifteen interviews across three institutions cannot establish. The more defensible description is that the sample carried sufficient information power for the explanatory purpose it serves here: the aim is narrow, the population specific, the guide theory-informed, and the analysis is supported by quantitative material from the same respondents.

Performance-based verification task

To check self-reported competency against demonstrated behaviour, each interviewee was shown five real posts drawn from Somali Facebook, WhatsApp, and Telegram circulation and asked to (a) judge whether each was reliable or misleading, (b) explain the reasoning, and (c) describe any verification step they would take. Posts were selected to represent five categories prevalent in the Somali environment: a fabricated political claim attributed to a government minister; an out-of-context image re-shared as current conflict footage; unverified health advice; a clan-targeted rumour designed to inflame division; and, as a control, a verified factual item from an established outlet. Participants also rated their confidence in their own ability to detect misinformation on a five-point scale. Because the judgement required of participants was binary, chance performance on this task is 50%, which is the benchmark against which the accuracy figures below should be read.

Analysis

Survey data were analysed in SPSS. Descriptive statistics summarise demographics, media use, and item-level and composite scores; a Pearson correlation describes the MLI–MFCI association within the sample (RQ2), with r² reported as an indicator of shared variance (Cohen, 1988). The quantitative analysis is descriptive and exploratory: no multivariable modelling, sample weighting, or formal hypothesis-testing framework was applied, and the correlation is interpreted as a statistical association among these respondents rather than as evidence of a causal relationship or of a relationship in the wider student population. Internal consistency was examined for the MLI (Cronbach's α = .84); the underlying factor structure of neither index was formally tested, so both composites, and the MFCI in particular, are treated as exploratory indices.

Both composites were computed as the unweighted mean of their constituent item scores (MLI, 12 items; MFCI, 9 items), so that each index keeps the 1–5 metric of the underlying scale and can be compared directly with the item means in Table 1; no item required reverse coding. All survey items were set as required in the web form, so the 393 cases analysed contain no item-level missing data, and neither imputation nor case deletion was necessary. For the principal study variables we report the mean, standard deviation and observed range.

Interview transcripts were analysed using reflexive thematic analysis (Braun and Clarke, 2006, 2021), worked through the six phases that approach specifies. Each transcript was read in full at least twice before any coding, with notes kept on first impressions. The author then coded all fifteen transcripts line by line, inductively and at both semantic and latent levels, with codes derived from the data rather than from a prior framework; the whole data set was coded before theme development began. Codes were collated into candidate themes on the basis of shared meaning rather than of how often they occurred. Candidate themes were then checked against their collated extracts and against the full data set, and were revised, merged or abandoned where they did not hold; it was at this stage that the survey's weakest indicators were used as sensitising concepts, to test whether patterns visible in the quantitative strand had qualitative counterparts, and it was here that the three drivers reported under RQ3 took their final form. Each retained theme was then given a definition fixing its boundary and a name reflecting its content, and extracts were selected for the reporting stage because they were analytically illustrative rather than merely quotable, with each theme reported alongside the survey indicator it bears on. Performance-task responses were scored as correct or incorrect against the known status of each post, and the verification strategy each participant described was categorised.

Coding was carried out by the author; codes, candidate themes and the final theme definitions were reviewed with the author's supervisor, and disagreements were settled by returning to the transcripts. We did not compute an inter-coder reliability coefficient, and this is a position rather than an oversight. In reflexive thematic analysis themes are treated as analytic constructions rather than as objects lying in the data waiting to be found, so agreement between coders is not the relevant test of quality (Braun and Clarke, 2021). We pursued quality instead through the procedures set out above—coding the entire data set, reviewing themes against it, keeping disconfirming material such as the minority of participants who described systematic verification, and reporting analytic decisions in enough detail for a reader to judge them. One reflexive note belongs here as well. The author is Somali and was researching Somali institutions from a university abroad, a position that plausibly made it easier for participants to discuss clan-related material than it might have been for an outside interviewer, and that equally supplied a shared frame of reference in which some assumptions will have gone unexamined.

Ethics

Participants received an information sheet describing the study's purpose and their right to decline or withdraw without penalty. Survey responses were anonymous; interview recordings and transcripts were de-identified before analysis, and no participant's clan or other potentially identifying affiliation was recorded in the dataset, although participants discussed clan dynamics in general terms. Because misinformation and social division are sensitive subjects in a post-conflict setting, interview questions were framed neutrally, participants were reminded that there were no right or wrong answers, and the clan-related post in the performance task was selected to avoid naming any real lineage. The study was conducted under the supervision of the Institute of Communication Studies, Communication University of China, and participating universities will receive the findings for curriculum use.

Findings

Sample characteristics

Table 5 summarises the 393 valid responses. All quantitative results in this section are descriptive and exploratory and characterise the 393 participating students rather than any wider population. Consistent with proportional allocation, Somali National University supplied 69.0% of respondents, the University of Hargeisa 28.0%, and East African University Garowe 3.1%. Because the Garowe figure represents only 12 students, site-level percentages for that university are reported for completeness and should not be read as estimates for its student body; the combined results are driven overwhelmingly by the Mogadishu and Hargeisa samples. Men made up 65.6% of the sample, which reflects enrolment patterns in Somali higher education. A majority (56.7%) were aged 21–25, and final-year students formed the largest year-group (38.7%). Economics (20.4%), medicine (14.2%), and education and science and engineering (12.2% each) were the most common fields, with law, art, management, agriculture, history, and philosophy also represented.

Media access and use

Social media dominate students’ information environment. Eighty-five per cent named social media as their main information platform; Facebook was used by 87.3% of respondents, WhatsApp by 67.2%, YouTube by 61.1%, and TikTok by 45.3% (Table 2). Traditional outlets played a secondary role: SNTV (44.5%), Universal TV (34.1%), and Garowe Online (24.7%) were the most used, and interviewees described television mainly as a place to confirm what they had already seen online. News engagement was intense, with two-thirds of respondents checking news at least once a day and 35.4% several times a day. Time on social media was substantial: 38.9% reported one to three hours daily, 29.8% three to five hours, and 14.2% more than five. The main purposes were everyday information (65.6%), social connection (61.5%), and study-related communication (61.2%). Somali was the dominant language of consumption (49.4%), followed by English (25.7%) and mixed Somali–English (21.9%), a distribution with direct implications for the usability of fact-checking resources, which are overwhelmingly English-language.

Table 5. Demographic characteristics of survey respondents (N = 393).

Percentages are of N = 393 and may not total 100.0 owing to rounding. Education and science and engineering are separate fields of 48 respondents each.

Interviews clarified why Facebook, rather than the faster-growing TikTok, remains the default news source. Students described a clear functional division of labour between platforms. Facebook is where “the news pages and journalists post” (Participant D) and is treated as “a central hub” for Somali affairs (Participant B); WhatsApp is for private sharing; TikTok is for entertainment and short commentary; YouTube is for learning. As Participant J put it, “the popularity of TikTok at the moment is tremendous, particularly among the youth. Still, I prefer Facebook for more in-depth news.” Participant A described Facebook as “the first thing that I look into in the morning. It is the quickest method of knowing what is happening even before the official TV channels broadcast it.” The platforms are not interchangeable in students’ minds, and this differentiation shapes the pathways through which misinformation reaches them: political and civic misinformation via Facebook pages, forwarded rumour via WhatsApp groups.

Self-reported media literacy and misinformation-fighting competency (RQ1)

On the composite measures, students rated their general media literacy as moderately high (MLI: M = 3.82, SD = 0.74, range 1.83–5.00) and their misinformation-fighting competency as moderate (MFCI: M = 3.45, SD = 0.81, range 2.11–5.00). The 0.37-point difference between the two indices is the first indication of a pattern that recurs throughout the data: students feel more competent with media in general than with the specific task of identifying and resisting false content.

Item-level results sharpen the picture (Table 1). The highest-scoring item was awareness that “the spread of unverified information has the power to hurt society” (M = 4.15, SD = 0.82). Understanding of how fake news affects social media (M = 3.84) and confidence in distinguishing reliable from unreliable information (M = 3.72) were also comparatively high. Scores fell steadily as items moved from awareness towards practice. Actively seeking multiple outlets scored 3.58; the ability to check a story across traditional sources, 3.12; verifying before sharing on social media, 3.08 (SD = 1.21, the most dispersed item); and awareness of which websites are most likely to spread fake news, 2.91 (SD = 1.24), the lowest score in the instrument.

The gradient in Table 1 is the awareness–action gap in numerical form. The distance between the awareness item (4.15) and the verification-behaviour item (3.08) is 1.07 points on a five-point scale; the distance to the source-identification item (2.91) is 1.24 points. Students know misinformation is harmful, believe they can spot it, and yet report only middling verification practice and poor knowledge of where false content originates.

Interviewees described the same gap in their own conduct, frequently in the register of admission. Participant A: “When I encounter news on Facebook, I tend to accept it as it is. I do not visit another website or ask the television news to verify it. I guess I should, but I don’t.” Participant G: “When I read news on a page that I trust, I tend to believe it. Perhaps I ought to make more comparisons, yet I lack time and habit.” Participant M used almost the same words: “Maybe I should compare more, but I don’t have the habit.” A minority described more systematic practice. Participant E checks “whether they reference an official source, or whether they have a logo and an identifiable face. I would never believe a random individual who makes a big statement but has no proof.” Participant H relies on what he called “a kind of group verification”: when historians, reporters, and other informed people debating an item in Facebook groups or on X converge, he treats it as more reliable. Notably, even these more careful strategies depend on source recognition and social consensus rather than on independent checking of the claim itself.

Demonstrated discernment: the performance task

The performance task provided an exploratory behavioural check on self-assessment. Among these 15 participants the two did not align (Table 6). Because the task was administered to a small, purposively selected subgroup, the results below are descriptive and exploratory: they cannot be generalised to the full survey sample or to Somali university students, and they do not statistically validate the self-report instrument. Across the 75 judgements made by the 15 participants, 32 were correct (42.7%). On the four misinformation items alone, 21 of 60 judgements were correct (35.0%). Mean self-rated confidence in detecting misinformation was 3.61 on the five-point scale, equivalent to about 72%, a gap of roughly 30 percentage points relative to overall accuracy and 37 points relative to accuracy on misinformation items. Accuracy on both counts—42.7% overall and 35.0% on misinformation items—falls below the 50% that guessing alone would have produced on a binary judgement, as does accuracy on each of the four misinformation items taken separately. With 15 participants and five items, these percentages carry wide uncertainty and are reported as an exploratory indication rather than as precise estimates.

Table 6. Performance-based verification task results (N = 15).

The judgement required was binary, so chance performance is 50%. Verification-strategy counts total 15 in rows 1–3; in row 5 they refer to the 11 participants who judged the control item correctly.

The pattern of errors is more informative than the aggregate. Participants performed best on the control post, but their success rested on recognising a familiar outlet, not on any checking procedure; recognition is a useful heuristic that offers no protection against false content carrying a familiar look. The health item was judged correctly at a somewhat higher rate than the other misinformation posts, but almost entirely by medical students drawing on subject knowledge that does not transfer to other domains. The two weakest results, the decontextualised image and the clan-targeted rumour (both 27%), point to two weaknesses of particular consequence in the Somali environment: an inability to interrogate visual material, and the capacity of identity-framed content to switch off evaluation altogether. Only 4 of the 15 participants used any systematic verification tool (cross-referencing, reverse image search, or a fact-checking site) on any of the five posts. The image result is the one that most closely matches what the recent literature would predict: Weikmann and Lecheler (2023) argue that images are granted a presumption of evidence that text has to earn, and a photograph separated from its date and place exploits that presumption at almost no cost to whoever reposts it.

The association between media literacy and misinformation-fighting competency (RQ2)

The MLI and MFCI were strongly and positively correlated, r(391) = .62, p < .001. By conventional benchmarks the coefficient is large (Cohen, 1988); the two indices share 38.4% of their variance (r² = .384). The coefficient describes the association among these 393 respondents; because the sample was not drawn probabilistically, it should not be read as an estimate of the corresponding association in the wider population of Somali university students. Students who report stronger access, evaluation, and responsible-sharing competencies also report substantially stronger competency in recognising and resisting misinformation. Given the cross-sectional and self-reported nature of both measures, the coefficient indicates association rather than causation, and some shared-method variance is likely, as is some inflation from the conceptual overlap between the two instruments noted in the Theoretical Framework. Nonetheless the size of the relationship, combined with the 0.37-point gap between the index means, is consistent with the study's framing: in this sample, general media literacy is closely bound up with self-reported misinformation resistance yet does not on its own appear to guarantee it.

Interview data suggest one mechanism linking the two. The students who described the most careful sharing behaviour framed it as a matter of accountability rather than technique. Participant L, a medical student: “As a future doctor, I have a duty to do no harm. That applies to what I post online too.” Participant F: “People will no longer trust me if I provide fake information. So I am very careful.” Participant I: “I do not write anything without considering whether it is true, whether it is fair, and whether it can hurt someone.” For these students media literacy operates as an ethic of publication, in which sharing is understood as an act with consequences for one's reputation and for others, and that ethic motivates the verification that less engaged peers described lacking the “habit” for.

Contextual factors students associate with misinformation spread (RQ3)

Asked why misinformation travels so easily in Somalia, interviewees converged on three explanations. Each maps onto a weakness visible in the survey.

Critical-thinking capacity in the wider population, as students describe it

Participants located the problem in society at large rather than in themselves, while acknowledging that students are products of the same environment. “Anything posted on Facebook is believed by many people, especially the older generation. They don’t question it” (Participant A). “A person who cannot read or write well, who has not received much education, easily becomes a victim of misinformation” (Participant D). “The major problem is that many users cannot recognise and understand fake news. They do not have the critical thinking ability to handle it” (Participant H). This is the societal face of the survey's weakest indicators: source identification (2.91) and verification (3.08).

A forwarding culture that rewards speed

The second theme is the pace of sharing, especially in WhatsApp groups. “WhatsApp group forwarding culture is very common, and people forward without any confirmation. They don’t even consider whether it's fake news” (Participant C). “They find something interesting or outrageous, and they share it immediately without taking a moment to think; they just want to be the first to share” (Participant E). Participant K attributed the same behaviour to attention-seeking: “They just want attention.” In uses-and-gratifications terms, the gratifications of immediacy and social visibility are being obtained at the expense of accuracy, and they are obtained fastest by not pausing to check.

Clan affiliation as a credibility filter

The third and most context-specific theme concerns clan. “They see something that reflects their emotions or their loyalty to their clan, and they post it immediately to support their side, regardless of whether it is true” (Participant F). “In Somalia, misinformation is commonly used to control people's emotions, particularly regarding clan and politics” (Participant I). Across the interviews, information originating from one's own clan or close social network was described as being accepted faster and questioned less than information from unfamiliar sources; emotionally charged political content tied to clan identity was believed and forwarded because it arrived through trusted channels; and students acknowledged a readiness to defend or repost material favourable to their own group even when unsure of its accuracy. In these accounts clan affiliation operates as an informal credibility filter, and participants described it as accelerating the circulation of congenial falsehoods. The performance task corroborates this directly: the clan-targeted rumour produced the joint-lowest accuracy, and participants’ explanations of their judgements on that post referred to identity rather than to evidence.

Students’ recommendations

Interviewees were unanimous that universities should move from bystanders to active providers of media literacy education, and their proposals target the gaps the survey identified (Table 3). Fourteen of 15 called for workshops and seminars; 12 wanted media literacy embedded as a compulsory part of the curriculum (“It shouldn’t be optional”, Participant I); 11 asked specifically to be taught fact-checking; and 8 wanted to understand how platform algorithms shape what they see. Participant H framed the demand in terms of citizenship: “Every student should be taught to be a responsible online citizen. It is a necessary skill for the 21st century.” Several also described a multiplier effect that extends the case for intervention beyond campus: “When we learn here, we can educate our families and communities” (Participant A).

Discussion

An awareness–action gap visible in two kinds of measurement

The study's central result is that, among the students who took part, awareness of misinformation substantially exceeds reported verification behaviour, and self-assessed discernment substantially exceeds demonstrated discernment. Two different kinds of measurement point the same way. They are not independent samples—the 15 who took the performance task are a purposive subset of the 393 surveyed, assessing themselves on the same construct—but they differ in kind, one asking what students believe they can do and the other observing what they did. In the survey, awareness (4.15) outstrips verification (3.08) and source identification (2.91) by more than a full scale point. In the performance task, students who rated their detection ability at roughly 72% correctly classified 35% of misinformation items, a rate below the 50% that guessing alone would have produced on a binary judgement, and below it on every one of the four misinformation posts.

This pattern is consistent with a large body of evidence on the limits of self-report. Overconfidence is greatest precisely where competence is lowest, because the skills required to perform a task well are the same skills required to recognise poor performance (Kruger and Dunning, 1999). It is also the reason Maertens et al. (2024) argue for measuring discernment by performance rather than by self-assessment, and our two instruments produce exactly the divergence their argument anticipates. It also fits the intervention literature, in which knowledge and attitudes move more readily than behaviour (Jeong et al., 2012), and the regional literature documenting the same disjunction among students in Ghana, Nigeria, Kenya, and South Africa (Adjin-Tettey and Amenaghawon, 2024; Madrid-Morales et al., 2021; Mustapha et al., 2024; Wasserman and Madrid-Morales, 2019). What the present study adds is a behavioural measure in a context where none existed, and preliminary, exploratory evidence that the gap is present among a group of educated young Somalis; whether it holds in the wider student population remains an open question.

The error profile in the performance task points to where competence is thinnest. Success on the control item and on the health item came from recognition and prior knowledge respectively, neither of which is a verification skill. The two lowest scores involved a manipulated image and an identity-targeted rumour. Visual verification (checking the provenance and date of images and video) and resistance to identity-framed appeals are therefore plausible priorities for instruction, although, with 15 participants and five items, this should be treated as a hypothesis for further testing rather than as an established finding; neither capacity is well served by general awareness-raising.

Media literacy matters, but as practice rather than knowledge

The strong correlation between the media literacy and misinformation-fighting indices (r = .62) is consistent with the premise that media literacy is a meaningful resource in Somalia, although a cross-sectional design cannot establish that it is protective, in line with Livingstone (2004), Potter (2004), and the meta-analytic evidence (Jeong et al., 2012). But the descriptive results qualify how that resource should be understood. The competencies that scored highly were declarative: knowing that misinformation harms, understanding that it circulates on social media. The competencies that scored poorly were procedural: checking, comparing, identifying sources. Buckingham’s (2015) argument that literacy is a practice to be habituated rather than a body of knowledge to be transmitted is borne out almost item by item. Participants who did verify described it not as an application of knowledge but as a habit sustained by a sense of accountability, and those who did not verify described their failure as an absence of habit, not of knowledge. The implication for curriculum design, which follows from descriptive patterns rather than from tested effects, is that instruction should be organised around repeated practice of verification routines in realistic conditions, with feedback, rather than around lectures about the dangers of fake news.

This is also where the intervention literature of the last five years is most useful, and where it should be read with most caution. The approaches with the best evidence behind them share the property our data point towards: accuracy prompts work by interrupting the moment of sharing rather than by adding to what the user knows (Pennycook et al., 2021; Pennycook and Rand, 2022), and inoculation works by rehearsing the manipulation techniques in advance so that recognition is available when it is needed (Roozenbeek et al., 2022). Both are procedural. Kozyreva et al. (2024) make the same observation across nine intervention families. The caution is that almost none of this evidence comes from a setting resembling Somalia’s. Guess et al. (2020a) are the instructive case: their digital media literacy tips improved discernment in the United States and in India, but the Indian gains did not last and were absent among rural respondents—that is, among the people the intervention most needed to reach. A curriculum built in Mogadishu on findings generated in Cambridge and New Haven would be borrowing a mechanism, not importing a result, and it would need evaluating locally.

Platform gratifications and the architecture of exposure

Uses and gratifications theory explains a pattern that a purely cognitive account would miss. Students select platforms for distinct gratifications, and those selections structure their exposure. Facebook satisfies the need for fast, comprehensive Somali news and is treated as a quasi-official infrastructure; WhatsApp satisfies the need for social connection and is the vector of the forwarding culture; TikTok satisfies entertainment needs and is, revealingly, not trusted for news. Facebook's centrality is consequential because it is the platform on which Somali political and clan contestation is most visibly conducted (Chonka, 2019) and on which page identity, rather than claim verification, was students’ dominant credibility cue. When the gratification being sought is speed or social visibility, as interviewees described for WhatsApp forwarding, the act of verifying is not merely neglected but actively costly, because it delays the reward. Interventions that ignore these motivations, and treat misinformation as a knowledge deficit alone, are unlikely to alter what a student does in the second before forwarding. Programmes and platform measures that act at that second—an accuracy prompt of the kind Pennycook et al. (2021) tested, or friction deliberately introduced into forwarding—engage the gratification structure directly, which a lecture delivered several weeks earlier cannot. That is the sense in which the problem is not only a cognitive one: verification has to compete with a reward that is social, immediate, and rewarded in real time.

Clan as an informal credibility filter

The most context-specific contribution of this study is qualitative and exploratory evidence that clan affiliation appears to operate as a credibility heuristic in these students’ information processing. Interviewees described in-group provenance as accelerating acceptance and suppressing scrutiny and reported forwarding identity-congruent political content without checking it, and the clan-targeted post produced the joint-lowest accuracy in the performance task. This finding aligns with work on group-based processes in misinformation (Efstratiou and De Cristofaro, 2022) and on the priority of identity goals over accuracy goals (Van Bavel and Pereira, 2018), and it extends to Somalia the lesson drawn from South Sudan that in post-conflict settings literacy efforts must engage intergroup trust directly (Otunga, 2020). Two qualifications matter. First, Pennycook and Rand (2019) show that analytic engagement improves discernment even for identity-congruent content, so identity is a barrier to be designed around rather than an insurmountable one. Ecker et al. (2022) put the same point in terms of correction: identity-driven belief is unusually resistant to correction after the fact, which is an argument for intervening before or during the encounter rather than afterwards. Second, in-group trust is not irrational in an environment in which formal institutions have historically been unreliable (Chonka, 2019); the same networks that transmit rumour also transmit the information people need. The task for media literacy education is not to dismantle that trust but to insert a verification step into it, so that in-group provenance ceases to be treated as sufficient evidence of truth.

The language constraint

Half of respondents consume media primarily in Somali, and a further fifth in mixed Somali and English. Fact-checking resources, browser tools, and reverse image search interfaces are predominantly English-language, and Somali remains a low-resource language for automated detection (Joshi et al., 2020; Mohamed et al., 2025). This constraint helps to explain why so few participants used any tool in the performance task and why “teach us how to fact-check” featured so prominently among their recommendations. It also defines a policy priority that universities cannot meet alone: Somali-language verification tools and locally grounded fact-checking capacity. The constraint runs through the study itself as well as through its subject matter: the questionnaire was administered in English to students who consume media mainly in Somali, and that mismatch is discussed among the limitations below.

Contributions to theory and measurement

Three contributions follow from these findings. First, the study gives empirical content to a distinction that media literacy theory often leaves implicit: between declarative literacy (knowing what misinformation is and that it is harmful) and procedural literacy (being able and disposed to verify). The competency models of Livingstone (2004) and Hobbs (2010) list evaluation alongside access and analysis as if they were acquired together; our data suggest that in this sample they were not, and that the procedural component was the one that lagged.

Second, the study extends uses and gratifications theory into the domain of misinformation by identifying identity affirmation, the gratification of supporting one's own group, as a motive for sharing that competes directly with accuracy, and by showing that the platform through which content arrives changes the credibility test applied to it. Third, and methodologically, the study illustrates how a small performance task embedded in a qualitative phase can flag a possible self-report validity problem that a survey alone would conceal. The 30-to-37-point gap between confidence and accuracy found in this 15-participant sub-sample is large enough to suggest that self-report instruments, which dominate media literacy research in Africa, should not be used as the sole indicator of competency, a conclusion Maertens et al. (2024) reach from a very different evidential base. Where full-scale behavioural testing is infeasible, even a five-item task with a purposive sub-sample provides a valuable calibration check.

The concept of the clan credibility filter that emerges from the interviews and the performance task may also have purchase beyond Somalia. In any setting where trust in formal institutions is low and social identity is a primary organising principle, in-group provenance is likely to function as a substitute for verification. The South Sudanese case documented by Otunga (2020) and the identity-based dynamics reviewed by Efstratiou and De Cristofaro (2022) suggest as much. Comparative research across post-conflict and segmentary societies would establish whether the mechanism generalises and whether the interventions proposed below travel with it.

Limitations

Several limitations qualify these conclusions, and together they define the quantitative strand as descriptive and exploratory rather than inferential. The sampling was non-probabilistic and purposive at both the institutional and the individual level, and the institutional sub-samples were highly disproportionate (271, 110 and 12); the findings describe students at three urban universities and are driven overwhelmingly by the Mogadishu and Hargeisa samples. The three universities were chosen for their regional and political-administrative significance rather than sampled at random from the roughly 100 institutions nationally, so no estimate reported here can be treated as statistically representative of Somali universities or of Somali university students as a population. The Garowe sub-sample of 12 students is far too small to support institution-level statistical inference and is reported descriptively only; no sampling weights were applied, and weighting would not have compensated either for the size of that subgroup or for the purposive selection of institutions. Men are over-represented, in line with enrolment, but gender differences in media use could not be examined reliably.

The survey instrument measured self-reported competencies and is therefore subject to the very overestimation the performance task suggested; the correlation between the two indices may also be inflated by common-method variance and by the conceptual overlap between them. Internal consistency was reported for the MLI (α = .84) but not for the MFCI, and the factor structure of neither index was formally examined; the MFCI in particular, which combines awareness, knowledge and behavioural items, should therefore be read as an exploratory composite rather than as a validated unidimensional scale. The performance task was administered to only 15 purposively selected participants and used five posts, so its accuracy estimates carry wide uncertainty and are exploratory; they neither validate the self-report instrument statistically nor establish a population-level awareness–action gap. The design is cross-sectional and observational and cannot establish causal direction between media literacy and misinformation resistance; every relationship reported here is a statistical association within the study sample, and expressions such as “associated with” or “statistically related to” should be read literally throughout. Finally, the questionnaire was administered in English to students consuming media largely in Somali, which may have affected comprehension of some items.

Three further limitations concern the instruments and the qualitative strand. The first extends the language point just made: the questionnaire was neither cognitively pretested nor formally piloted, and no Somali version was produced, so nothing in the design establishes that the more nuanced items about credibility and verification were read as intended. The second concerns the qualitative analysis, which was coded by a single analyst and reviewed by a second. It follows reflexive thematic analysis and therefore makes no claim to coding consensus, and it makes none to theoretical saturation either; the themes reported are analytic constructions that a different analyst might have drawn along different lines. The third concerns the title. Misinformation resilience, which frames the argument, was not measured directly. What the study measures are two of its components at one point in time, reported competency and demonstrated resistance on five items, and resilience proper—a durable capacity across repeated exposure—would require the longitudinal or experimental designs used in the inoculation literature (Roozenbeek et al., 2022). The study's standing in that literature is diagnostic rather than evaluative: it describes a starting condition against which interventions might later be measured.

It is also important to distinguish three forms of inference that are sometimes treated as equivalent. The first is statistical inference within the observed sample, which concerns the precision of estimates for the respondents who took part. The second is generalisation to the population represented by a sampling frame, which requires a probability-based design that this study did not employ. The third is causal inference, which requires a design capable of establishing temporal order and of ruling out confounding. This study attempts only the first, and in a descriptive form; it makes no claim to the second or the third, and the additional procedures that might have strengthened the statistical analysis—weighting, bootstrapping, correlation or regression—could not by themselves have established national representativeness or causal relationships. Future research should employ probability-based sampling, sufficiently large samples from each region and institution, formal assessment of scale reliability and dimensionality, and appropriate multivariable and inferential methods in order to test whether the patterns described here hold in the wider population of Somali university students.

Conclusion and recommendations

The Somali university students who took part in this study are heavy, daily consumers of social media who are well aware that misinformation harms their society, believe themselves reasonably capable of recognising it, and in practice verify little and detect less. Within this sample, media literacy was positively associated with self-reported misinformation-fighting competency, but the competencies students reported were largely declarative, and the behavioural link that would translate them into resilience is weak. Interview accounts suggest that the gap is sustained by a forwarding culture that rewards speed, by platform gratifications that make verification costly, by clan affiliation acting as a substitute for evidence, and by the absence of Somali-language verification tools. Interviewees at all three universities described media literacy provision as absent or optional, and twelve of the fifteen wanted it made compulsory. These conclusions are descriptive and exploratory, and they characterise the three participating universities rather than Somali higher education as a whole.

Because the recommendations that follow rest on descriptive findings from three purposively selected universities, they are offered as provisional priorities rather than as conclusions supported by population-level or causal evidence. For universities, media literacy should be integrated into the core undergraduate curriculum as a required, practice-based component rather than an optional or awareness-oriented one. Instruction should be organised around repeated verification exercises using real Somali-language content, with explicit attention to image and video provenance, source identification, cross-platform checking, and the recognition of identity-framed appeals; it should be reinforced across disciplines by faculty trained in media literacy pedagogy and supported by librarians equipped to assist with verification. Discipline-specific modules can build on the accountability ethic that the most careful students already articulate, for example through professional codes in medicine, law, and education.

For policy, the Federal Government and Federal Member States should adopt media literacy as a national educational competency with a multi-stakeholder coordinating body spanning the Ministries of Education and Information, universities, media organisations, and development partners, and should fund the development of Somali-language fact-checking resources and support existing local verification initiatives so that students can check claims in the language in which they encounter them.

For research, priority should go to probability-based sampling with adequate numbers of students from each region and institution, so that the patterns described here can be tested inferentially; to larger-scale performance-based measurement that replaces self-report as the primary indicator of competency; to the development and validation of a Somali-language instrument, since the language constraint documented here applies as much to research as it does to fact-checking; to studies of faculty media literacy, since instructors cannot teach what they do not practise; to populations that interviewees identified as more vulnerable than themselves, notably older adults and rural communities; to longitudinal designs that track whether skills acquired at university persist; and to rigorous evaluation of the interventions that this study recommends, so that what is implemented in Mogadishu, Hargeisa, and Garowe is shown to work rather than assumed to.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Anastassia Zabrodskaja, Tallinn University, Estonia

Reviewed by: Maryam Tavosi, National Library and Archives of Iran, Iran

Abd Allah Sultan Aljalabneh, Zarqa University, Jordan