Language, Credibility, and AI-Assisted Verification: Evidence from Jordanian Digital Journalism

Abstract

The rapid circulation of misleading information through digital media has made credibility assessment increasingly challenging for journalists and audiences. This study examines how linguistic and visual credibility cues contribute to the appearance of reliability in misleading digital news and explores how AI-assisted systems respond to these cues in Arabic and English. The study adopts a qualitative comparative case-study design based on five prominent and analytically suitable misinformation cases documented by AKEED, the Jordanian Media Credibility Monitor, in 2026. The cases represent different mechanisms of misleading communication, including AI-related visual content, numerical claims, recycled visual material, institutional attribution, and financial interpretation. Each case was evaluated by an AI system in Arabic and English using the same verification-oriented prompt structure. The findings from the five cases suggest that AI-assisted verification may help identify questionable claims and possible verification pathways, although the responses varied in their treatment of evidence, context, and interpretation. Greater challenges emerge when misleadingness depends on contextual displacement or unsupported interpretation. The study argues that AI can support digital journalism and verification, but effective assessment continues to require source checking, contextual knowledge, and human editorial judgment.

1. Introduction

Digital news increasingly circulates in environments where accurate reporting, misinformation, manipulated images, infographics, and AI-generated content coexist. In such environments, readers may find it difficult to distinguish reliable information from misleading content. Misinformation does not necessarily originate from sources that appear untrustworthy. Instead, it may include precise figures, refer to official institutions, use professional journalistic language, or feature images that appear to provide evidence of an event. These features can create an impression of credibility even when the information is inaccurate or insufficiently supported by evidence.

The growing use of artificial intelligence has added another layer of complexity to this issue. AI tools can assist journalists with information retrieval, summarization, translation, and the processing of large volumes of information, and they can also support verification and fact-checking. At the same time, generative AI facilitates the production of realistic images and persuasive textual content whose origins may not be immediately apparent to audiences. The relationship between AI and misinformation is therefore complex. Although AI can be used to create and disseminate misleading content, it can also support efforts to identify questionable claims and investigate the evidence underlying them.

Credibility assessment is central to understanding how audiences evaluate digital information. Online information is judged not only by the accuracy of its factual content but also by how it is presented. Linguistic choices, numerical details, source attribution, visual evidence, references to institutions, and expressions of certainty can all influence how a claim is perceived. Research on online credibility has shown that individuals may rely on credibility cues and cognitive heuristics when evaluating information (Metzger and Flanagin 2013). Research on misinformation has also demonstrated that misleading information can continue to influence people even after it has been corrected (Ecker et al. 2022). These issues are particularly relevant to journalism, where verification is a fundamental professional practice, especially in digital environments where information can be rapidly reproduced and detached from its original context.

These questions are particularly relevant in the Arabic digital environment. Although research on misinformation and automated fact-checking has expanded, the capabilities and limitations of multilingual AI systems in assessing Arabic-language claims warrant further investigation. For example, Gupta et al. (2025) examined the capabilities of multilingual large language models (LLMs) in evaluating Arabic claims and identifying supporting evidence. However, less attention has been paid to the specific linguistic and multimodal cues through which misleading information may appear credible, and to whether an AI system responds to these cues similarly when the same claim is presented in Arabic and English.

This study examines these issues through five misinformation cases documented by AKEED, the Jordanian Media Credibility Monitor, in 2026. The cases represent different forms of misleading digital information, including AI-related visual content, numerical and infographic-based claims, recycled or decontextualized images, misleading institutional attribution, and financial interpretation. The study first examines how credibility is constructed through these different mechanisms and then compares how an AI system evaluates the same claims when presented in Arabic and English. Accordingly, the study addresses three research questions: (1) Which credibility cues contribute to the appearance of credibility in the selected misinformation cases? (2) How does the AI system respond to these claims in Arabic and English? (3) What similarities and differences emerge between the Arabic and English responses?

The study does not treat AI as an independent or infallible fact-checker. Instead, it examines AI as a potential assistive tool in the broader process of verification, recognizing the continued importance of human judgment and contextual knowledge. By bringing together linguistic analysis, digital journalism, misinformation research, and AI-assisted verification, the study contributes to discussions of how credibility cues operate in contemporary digital news environments. More specifically, it examines how linguistic and visual cues may make misleading information appear credible and explores what an AI system can and cannot identify when evaluating such information in Arabic and English.

2. Literature Review

Audiences’ trust in online information is closely related to how they assess its credibility. In digital environments, readers often have to make judgments about the reliability of information without sufficient time or resources to verify every claim independently. Research on online credibility suggests that individuals may rely on cues such as source trustworthiness, the presentation of information, and the availability of supporting evidence when making these judgments (Metzger and Flanagin 2013). These cues are particularly relevant when misleading information is presented in a format resembling legitimate news reporting.

Research has also shown that evaluating news credibility involves more than simply distinguishing true information from false information. A systematic review and meta-analysis of news judgments found that audiences generally rated true news as more accurate than fact-checked false news, although judgments varied across information contexts and types of news (Pfänder and Altay 2025). These findings highlight the importance of examining the features that shape credibility judgments, rather than focusing exclusively on whether audiences accept or reject a claim.

In journalism, credibility depends not only on the accuracy of information but also on how it is sourced, contextualized, presented, and attributed. An inaccurate statistic, an old photograph presented as recent, or a reference to an institution without sufficient supporting information can create an unwarranted impression of reliability. This issue is particularly relevant to digital journalism, where information can be rapidly reproduced and circulated without its original context. Examining the relationship between presentation and evidence is therefore essential to understanding how misleading information may appear credible.

The emergence of generative artificial intelligence has added further complexity to this issue. Generative AI facilitates the production of persuasive text and realistic images, audio, and video, making the origins and authenticity of digital content more difficult to establish in some circumstances. A scoping review of research on generative AI and disinformation identified journalism, fact-checking, deepfakes, media literacy, and scientific disinformation as important areas of investigation (López-Borrull and Lopezosa 2025). The review also highlights the dual role of AI: it can facilitate the production and dissemination of misinformation while offering potential tools for its detection and verification.

Research on AI-generated images in news contexts has examined how audiences interpret synthetic visual content and how the use of such images may affect expectations of transparency and trust in news (Strikovic and Cools 2025). Research on the attribution of news content to human or AI authors has likewise suggested that the relationship between authorship and perceived credibility is not straightforward. The findings indicate that credibility judgments do not always differ significantly according to whether a story is attributed to a human or an AI system, although some measures have shown a preference for human-authored content. Taken together, these findings suggest that the relationship between AI and credibility cannot be reduced to a simple distinction between human-generated and AI-generated information. How information is presented, attributed, and interpreted also matters. This reinforces the need to examine linguistic and visual features when investigating misleading digital content.

At the same time, AI is increasingly being considered as a potential tool for countering misinformation. A scoping review of 199 studies on automated fact-checking found that this research area has been dominated by computational approaches, with comparatively less attention to journalistic perspectives and professional verification practices (Kavtaradze 2025). This distinction is important because journalistic fact-checking involves more than computational assessments of factuality; it also requires contextual interpretation, professional judgment, and human oversight. The distinction highlights the need to consider AI-assisted verification not only in terms of computational performance but also in relation to the practices and contextual judgments involved in journalism.

Empirical research has demonstrated both the potential benefits and the risks of using AI to support verification. In an experiment involving 80 participants, Si et al. (2024) found that explanations generated by large language models (LLMs) could help users verify claims more efficiently than traditional search engines. However, participants also tended to rely excessively on the models when their explanations were incorrect. This finding illustrates a central challenge of AI-assisted verification: a system may provide useful explanations while still producing inaccurate or misleading information that users may accept too readily.

The problem extends beyond the accuracy of individual facts. Research on LLM factuality has shown that these systems can produce persuasive and seemingly accurate responses that nevertheless contain false or misleading information. At the same time, LLMs can perform subtasks that may support different stages of the fact-checking workflow (Augenstein et al. 2024). AI-assisted verification should therefore be understood as a means of supporting journalistic work rather than replacing independent evaluation, source checking, and professional judgment.

These challenges become more complex when verification involves multiple languages. Research on misinformation detection and automated fact-checking has developed substantially, but the capabilities and limitations of multilingual AI systems in evaluating Arabic-language claims warrant further investigation. Arabic-language content may contain culturally specific references, local institutional names, regionally familiar expressions, and context-dependent vocabulary that require more than literal interpretation. Without sufficient linguistic and contextual understanding, an evaluation may address the apparent factual content of a claim while overlooking the circumstances that shape its meaning. Similarly, Khoury (2026) highlights the challenges of preserving contextual meaning in machine-assisted Arabic–English translation, particularly when dealing with figurative expressions.

Recent research provides empirical evidence that LLM performance can vary across languages and prompting strategies. Gupta et al. (2025) used 771 Arabic claims from the X-Fact dataset to evaluate four LLMs under different prompting strategies. Their findings indicated that performance varied according to the prompting method and that cross-lingual prompting improved fact-checking performance. These findings suggest that the evaluation of Arabic claims may depend partly on how prompts are formulated and how linguistic resources are used. They also provide a rationale for examining Arabic and English responses comparatively rather than assuming that the same claim will be evaluated identically in both languages.

The Jordanian context provides a relevant setting for examining these issues. References to government institutions, official statistics, local organizations, and familiar journalistic expressions may influence how claims are interpreted. The significance of such cues depends on their relationship to the underlying evidence and the context in which a claim circulates. An assessment that focuses only on isolated factual details may therefore overlook contextual information needed to evaluate the claim adequately. Examining misinformation cases documented in Jordan allows these issues to be investigated through examples grounded in a specific digital information environment.

Although existing research has examined misinformation, AI-generated content, automated fact-checking, and the use of LLMs in verification, these topics have often been investigated separately. Research on AI-generated misinformation frequently focuses on synthetic content and its detection, whereas research on automated fact-checking often emphasizes computational performance. Less attention has been given to how linguistic and visual features contribute to the appearance of credibility in misleading digital information and how AI systems interpret these features when evaluating claims across languages.

A related gap concerns the relationship between linguistic analysis and journalistic verification. A misleading claim may appear credible not because its factual content is complex, but because of the way it is worded, the precision of its numerical details, the authority attributed to its source, or the visual material accompanying it. Examining these features alongside AI-generated evaluations can provide a more detailed account of the information an AI system identifies, the evidence it considers relevant, and the contextual issues it may overlook. Addressing this gap, the present study analyzes five misinformation cases documented by AKEED, the Jordanian Media Credibility Monitor, in 2026. It examines the credibility cues represented in these cases and compares how an AI system evaluates the same claims in Arabic and English. In doing so, the study brings linguistic analysis and AI-assisted verification into a common qualitative framework while keeping the findings specific to the selected cases and evaluation conditions.

The eight analytical dimensions were developed for this study based on the literature reviewed on online credibility assessment, misinformation verification, and the use of large language models in fact-checking. They were designed to examine complementary aspects of AI-assisted verification, including claim identification, evidence assessment, source selection, uncertainty, contextual interpretation, and the recognition of credibility cues. The dimensions provide a structured qualitative framework for examining the responses without reducing their evaluation to a numerical performance score.

Credibility assessment involves more than evaluating the factual content of a claim; it also requires attention to the features through which information is presented and interpreted. Institutional attribution may lend a claim an appearance of authority, while numerical precision may create an impression of objectivity even when the underlying figures lack adequate sourcing or context. Similarly, visual realism can make an image appear authentic without establishing its provenance, and temporal specificity may make a claim seem relevant to a current event even when the supporting material is outdated. Professional language may also contribute to perceived legitimacy, although stylistic formality does not establish factual accuracy. These features are therefore treated as potential credibility cues whose significance depends on their relationship to verifiable evidence, source reliability, and contextual consistency. Building on research on online credibility assessment and misinformation verification, the present framework brings linguistic, numerical, visual, institutional, and temporal features into a common qualitative analysis of AI-assisted verification. It extends this perspective by examining not only which cues appear in a claim, but also how an AI system identifies, evaluates, and contextualizes them when responding in Arabic and English.

In this study, credibility cues are treated as observable features of misleading information rather than as direct measures of its truth or reliability. These features include word choice and expressions of certainty, numerical precision, institutional attribution, temporal references, contextual framing, and visual presentation. Their significance depends on how they function within each case. For example, numerical precision may create an impression of authority without establishing the accuracy of a statistic, while a genuine photograph may become misleading when presented as evidence of a different event. The analysis therefore examines these features in relation to the specific claim and its documented verification. It then considers how the AI system responds to them through eight qualitative analytical dimensions: claim identification, scope and reference, evidence evaluation, source selection, uncertainty management, contextual reasoning, recognition of credibility cues, and actionability. These dimensions provide a structured basis for examining the AI responses without treating credibility as a numerical measure or assuming that the presence of a particular cue necessarily makes information misleading.

The present study addresses this gap by examining five misinformation cases documented by AKEED in Jordan in 2026 and comparing how an AI system evaluates the same claims in Arabic and English. Rather than assessing the system solely in terms of binary correct-or-incorrect outcomes, the study examines how it identifies claims, evaluates evidence, selects sources, manages uncertainty, incorporates contextual information, and recognizes linguistic and visual cues that may contribute to the appearance of credibility.

3. Methodology

This study adopts a qualitative comparative case-study design to examine how misleading digital information creates an appearance of credibility and how an AI system responds when evaluating the same claims in Arabic and English. The study is exploratory rather than quantitative and does not seek to establish an accuracy rate or rank performance across the two languages.

The study focuses on the Jordanian digital information environment and draws on cases documented by AKEED, the Jordanian Media Credibility Monitor, in 2026. AKEED publishes verification reports documenting misinformation claims and the evidence used to assess them.

During the first half of 2026, AKEED reported 697 rumors, of which 525 circulated through social media and 172 through media outlets. These figures provide the broader monitoring context for the present study.

Purposive criterion sampling was used to select five cases from the 2026 AKEED monitoring corpus based on their analytical relevance and suitability for comparison. The study does not claim that these were the five most widely circulated rumors, as comparable audience-reach data are not publicly available for every individual case.

Cases were included if they met the following criteria:

• They were documented by AKEED during 2026.
• They circulated publicly through digital media, social media, or online news outlets.
• Sufficient documentation was available to establish the original claim and AKEED’s verification.
• They contained identifiable credibility cues relevant to the study.
• They represented an analytically distinct mechanism of misleadingness.

Each case was selected for its relevance to a distinct credibility mechanism. The first case illustrates how AI-generated visual material may affect perceptions of journalistic credibility. The second examines numerical precision and source attribution in a misleading infographic. The third addresses the misleading reuse of an older image in a new temporal context. The fourth focuses on institutional attribution and the credibility associated with an international organization. The fifth examines financial framing and the interpretation of registered capital in relation to project investment. Together, these cases provide analytical variation across visual, numerical, temporal, institutional, and financial credibility mechanisms. The selection was based on analytical relevance rather than statistical representativeness or audience reach.

The five claims were evaluated using the same AI system and model version. Each claim was presented separately in Arabic and English, yielding ten AI responses. The corresponding AKEED verification reports were withheld from the AI system during the evaluations.

The English-language prompt was:

“Information is circulating on social media in Jordan claiming that [claim]. Assess the reliability of this information. Do not assume that the information is true or false. Analyze the claim, explain what needs to be verified, and identify the sources that should be consulted to verify it.”

The Arabic equivalent was:

“تنتشر على منصات التواصل الاجتماعي في الأردن معلومات تفيد بأن [الادعاء]. قيّم مدى موثوقية هذه المعلومة. لا تفترض أن المعلومة صحيحة أو خاطئة. حلّل الادعاء، ووضّح ما الذي ينبغي التحقق منه، وما المصادر التي يمكن الاعتماد عليها للتحقق منه.”

The AI-assisted evaluations were conducted using GPT-6 Sol with High reasoning in ChatGPT Plus between 5 and 8 September 2026. The same model and reasoning setting were used to evaluate all five claims in Arabic and English, producing ten responses. Each claim was submitted separately using the corresponding verification-oriented prompt presented above. The AKEED verification was not provided to the model before evaluation. The Arabic and English evaluations were conducted independently using equivalent prompts and identical evaluation settings. Web browsing and external search tools were disabled during the evaluations to ensure that the model’s responses were based on the information provided in the prompts and its existing knowledge rather than live retrieval of external sources. All responses were retained in their original form and analyzed qualitatively using the eight analytical dimensions described below.

To support methodological equivalence, the Arabic and English prompts were designed to convey the same verification task and preserve the same core instructions. Both versions introduced the claim as information circulating on social media in Jordan and used the corresponding claim without providing the AKEED verification. Equivalence was assessed in terms of the prompts’ intended meaning, task requirements, neutrality, and the information supplied to the model. Particular attention was paid to ensuring that neither version introduced additional contextual information, stronger evaluative language, or explicit cues about the expected conclusion. The prompts were used separately for the Arabic and English evaluations, and the same evaluation procedure was applied in both languages.

The analysis distinguishes between the credibility features observed in the original misinformation cases and the analytical dimensions used to examine the AI responses. The former include linguistic, numerical, institutional, temporal, contextual, and visual features; the latter organize the examination of how the AI system interprets claims and approaches verification. The relationship between these two elements is interpretive rather than one-to-one, as a single credibility feature may be relevant to more than one analytical dimension.

AKEED served as a benchmark verification reference rather than an unquestionable ground truth. The comparison focused on the AI responses’ assessments of the claims, use of evidence, reasoning, and proposed verification steps. Agreement with AKEED was not treated as proof of the AI responses’ objectivity or reliability.

The responses were analyzed thematically using eight qualitative dimensions, as presented in Table 1.

Table 1. Qualitative Coding Framework for AI Response Analysis.

	Analytical Dimension	Focus of Analysis
1	Claim identification	Whether the AI response identifies the central claim requiring verification.
2	Scope and reference	Whether the response clarifies relevant dates, locations, quantities, and references associated with the claim.
3	Evidence evaluation	How the response identifies and evaluates evidence relevant to verifying the claim.
4	Source selection	Whether the response identifies appropriate and credible sources for verification.
5	Uncertainty management	How the response acknowledges missing information, uncertainty, and limitations in reaching conclusions.
6	Contextual reasoning	Whether the response considers relevant temporal, institutional, and situational contexts.
7	Recognition of credibility cues	Whether the response recognizes linguistic, numerical, institutional, or visual features that may make a claim appear credible.
8	Actionability	Whether the response proposes specific and relevant steps for verifying the claim.

The eight dimensions were used as predefined qualitative coding categories. Each AI response was examined for how it identified the claim, evaluated evidence and sources, managed uncertainty, interpreted context and credibility cues, and suggested verification steps. The analysis was conducted by a single researcher, without independent second-coder assessment or intercoder-agreement testing, which is acknowledged as a limitation. The same coding framework was applied to all ten responses to support systematic qualitative comparison between the Arabic and English evaluations, without numerical scoring.

The analysis proceeded from individual case analysis to Arabic–English comparison and then cross-case interpretation. The AI responses were also compared with AKEED’s documented verification at the levels of claims, evidence, and reasoning.

The study uses publicly available misinformation cases and does not involve human participants, interviews, surveys, or private personal information. AI responses are treated as generated textual material rather than participant data. The researcher selected and analyzed the cases independently using publicly available AKEED verification reports. AKEED served as the source of the documented cases and verification evidence, rather than as a research collaborator. The analysis was conducted for academic purposes, with the aim of examining credibility cues and AI-assisted verification from a linguistic perspective.

The study is limited by its small, purposively selected sample and by the dependence of AI responses on the model version, prompt, available information, and date of interaction. The findings are therefore interpreted as observations specific to the conditions of the present study rather than as permanent characteristics of AI systems.

4. Data Analysis and Discussion

The final dataset consists of five AKEED cases selected for their prominence and analytical relevance. They represent different mechanisms through which information can acquire an appearance of credibility as indicated in Table 2.

Table 2. Selected 2026 AKEED Cases.

Case	AKEED Case	Date	Source	Primary Type	Main Credibility Mechanism
1	AI-related illustrative image used in local news coverage	3 April 2026	AKEED (2026e)	AI-assisted/AI-involved visual content	Visual realism, AI disclosure, journalistic transparency
2	Inaccurate infographic presenting unemployment rates in Arab countries	21 January 2026	AKEED (2026a)	Statistical/infographic misinformation	Numerical precision, source attribution, temporal accuracy
3	Old image presented as showing the Roman Theatre flooded during a recent weather event	16 January 2026	AKEED (2026b)	Recycled/decontextualized image	Image provenance, temporal context, visual evidence
4	Claim concerning a supposed UN “Quality of Life Initiative” ranking cities	1 March 2026	AKEED (2026c)	Misleading institutional attribution	Institutional authority, source credibility, terminology
5	Claim that the registered capital of a green-ammonia company indicates that a large-scale project is not credible	11 May 2026	AKEED (2026d)	Misleading financial/economic interpretation	Numerical framing, financial context, institutional reasoning

The five cases were selected to provide mechanism diversity rather than topic diversity. They illustrate different ways in which misleading information can acquire an appearance of credibility. These include the use of AI-assisted illustrative imagery, precise but insufficiently sourced statistics, photographs detached from their original temporal context, institutional attribution, and financial figures used to support unsupported conclusions. Examining these cases provides a basis for analyzing how linguistic and visual cues shape the interpretation of information and how the AI system responds to the resulting verification challenges.

Across the five cases, the AI responses generally recognized the need for verification, although the depth of their explanations varied. The responses appeared to address claims involving identifiable sources, dates, images, or numerical information more directly than cases requiring contextual interpretation or evaluation of inferences drawn from otherwise accurate information. These observations are specific to the cases examined and do not establish general differences in verification performance.

The study also identified recurring processes through which misleading information may acquire an appearance of credibility, including visual realism, numerical precision, institutional attribution, professional journalistic style, and the presentation of isolated details without sufficient contextual information to substantiate a claim.

The Arabic and English responses showed similarities in their general approaches to verification, alongside differences in contextual explanation and source specificity. In the cases examined, the English responses provided somewhat more explicit descriptions of verification procedures, whereas the Arabic responses placed greater emphasis on the local context of the claims. These observations are specific to the cases analyzed and should not be interpreted as evidence of systematic differences in AI performance between Arabic and English.

The first case concerns an AI-assisted illustrative image used in local news coverage. AKEED’s Arabic headline describes the case as follows

“صورة تعبيرية وإدخال الذكاء الاصطناعي .. وسيلة إعلام تضلل الجمهور بصريًا”

The phrase “صورة تعبيرية” (“illustrative image”) is particularly important from a linguistic perspective. Although it indicates that the image is not necessarily a documentary photograph, AKEED’s verification questioned whether this label provided sufficient transparency, given that the image depicted identifiable public figures and had been generated using AI.

The Arabic response recognized that the image should not automatically be treated as documentary evidence and emphasized the need to establish its origin and purpose. The English response followed a similar line of reasoning but discussed image provenance and comparison with the original news report in somewhat greater detail.

The case illustrates the distinction between visual authenticity and contextual appropriateness. An image may appear realistic without providing evidence that the depicted event actually occurred. In this instance, the label “صورة تعبيرية” becomes part of the credibility issue because it may not adequately communicate that the image was artificially produced.

The case also demonstrates that transparency is not solely a technical issue; it has a linguistic dimension. Labeling an image as “illustrative” is not equivalent to explicitly stating that it was generated using artificial intelligence. The distinction matters because the two formulations communicate different information about the image’s origin and evidential status.

This interpretation is consistent with research emphasizing the importance of transparency and audience interpretation in the use of AI-generated images in news contexts (Strikovic and Cools 2025). It also supports the argument that AI-assisted verification should examine not only the image itself but also the language used to frame it.

The second case concerns an infographic presenting unemployment rates in Arab countries. The original Arabic title was:

“البطالة في الدول العربية 2026”

The wording gives the infographic a factual and statistical appearance. The use of “البطالة” (“unemployment”), the geographical specification “في الدول العربية” (“in the Arab countries”), and the year “2026” presents the information as a defined statistical account of a particular region and period.

AKEED identified temporal and source-related problems with the infographic. It attributed the figures generally to the International Labour Organization without identifying a specific report, publication date, or statistical indicator. AKEED also questioned whether finalized figures for 2026 could reasonably have been available at the beginning of that year.

Both AI responses identified the need to verify the source and date of the unemployment figures. They also recognized that statistical information should be examined in relation to the population and period to which the figures refer.

The case illustrates how numerical precision can function as a credibility cue. Percentages such as 17.7%, 17.3%, and 16.4% create an impression of measurement and objectivity, but precision alone does not establish accuracy. The figures must be linked to a verifiable source, a clearly defined reference period, an appropriate statistical measure, and the population being described.

The linguistic framing is also relevant. The expression “لعام 2026” (“for the year 2026”) provides temporal specificity that may lend the information an appearance of authority. However, a precise date or period can itself be misleading when the underlying data do not substantiate the stated timeframe. This case therefore highlights the importance of examining the relationship between numerical claims and their supporting metadata rather than treating specificity as evidence of reliability.

The third case concerns an older image presented as evidence of recent flooding at the Roman Theatre. AKEED’s title explicitly states:

“الصورة التي تُظهر غرق المدرج الروماني قديمة ولا تعود للمنخفض الجوي الأخير”

The central distinction is between what the image depicts and the event to which it is attributed. The image may genuinely show water around the Roman Theatre, but the claim that it relates to a particular recent weather event constitutes a separate factual proposition. AKEED’s verification identifies the image as old and unrelated to the recent weather event.

The Arabic AI response recognized that the image needed to be checked against its original publication and contextual history. It emphasized determining when the image first appeared and whether it had previously been associated with another event. The English response similarly emphasized image provenance and comparison with contemporaneous reports.

This case illustrates the distinction between visual authenticity and contextual authenticity. An image may be genuine while the claim accompanying it is misleading. The misinformation arises from the relationship between the image and the linguistic framing that assigns it a particular temporal and situational context.

From a linguistic perspective, the expression “لا تعود للمنخفض الجوي الأخير” (“does not belong to the most recent weather system”) highlights the importance of temporal reference. The issue is not necessarily what the image depicts, but when the depicted event occurred. This distinction matters for AI-assisted verification because establishing that a photograph is genuine or has not been digitally manipulated does not establish that it provides evidence for the event being claimed. Verification must also consider when and where the image originated and whether its original context supports the accompanying claim.

The fourth case concerns the claim that a purported United Nations quality-of-life initiative ranks cities competitively. AKEED’s Arabic headline states:

“مبادرة جودة الحياة الأممية .. لا تُصدر ترتيبًا تنافسيًا للمدن من حيث جودة الحياة”

The expression “مبادرة جودة الحياة الأممية” (“the UN Quality of Life Initiative”) functions as an institutional credibility cue. The adjective “الأممية” associates the initiative with international institutional authority and may make the claim appear credible before its source has been established.

The case illustrates how institutional attribution can operate as a linguistic shortcut to credibility. References to recognized organizations may lend information an appearance of authority, but the presence of an institutional name does not establish that the organization issued the information or supports the specific claim.

Both language versions treated the institutional reference as requiring verification rather than as evidence in itself. The Arabic response focused on whether the purported initiative could be traced to an official United Nations source. The English response more explicitly distinguished between the existence of a United Nations body or programme and the specific ranking attributed to it.

This distinction is important because institutional names can be used rhetorically to strengthen the apparent legitimacy of a claim. Expressions such as “بحسب الأمم المتحدة” (“according to the United Nations”) may imply institutional endorsement even when that endorsement has not been established. Verification therefore requires identifying the original source and determining whether it actually issued or supports the information attributed to it.

The case reinforces the broader argument that credibility cues must be evaluated in relation to verifiable evidence. Institutional attribution may shape the presentation of a claim, but its evidential value depends on whether the cited institution can be connected to the specific information under investigation.

The fifth case concerns the interpretation of a company’s registered capital in relation to a large-scale green-ammonia project. AKEED’s Arabic title states:

“رأس مال شركة إنتاج الأمونيا الخضراء لا يُعد وحده مؤشرًا للحكم على جدية المشروع”

The wording contains specialized financial terminology, including “رأس مال” (“registered capital”), “شركة إنتاج الأمونيا الخضراء” (“green-ammonia production company”), and “جدية المشروع” (“seriousness of the project”). Such terminology can give a claim an appearance of technical expertise, although the use of specialized language does not necessarily establish the validity of the reasoning presented.

The case involved public discussion of a project estimated at approximately USD 1 billion and a company whose registered capital was JOD 5000. AKEED explained that the company had been established as a special-purpose vehicle and that registered capital alone was not an adequate measure of the project’s financing or seriousness.

The Arabic AI response recognized that registered capital alone could not establish whether a large-scale project was genuine, viable, or adequately financed. The English response reached a similar conclusion and more explicitly distinguished between registered capital and the total financing requirements of a project.

This case illustrates the difference between factual accuracy and interpretive accuracy. A financial figure may be correctly reported, yet the conclusion drawn from it may not follow. The registered capital of a company does not, by itself, establish the financial structure, funding arrangements, or viability of a large-scale project.

The case therefore demonstrates that misleading information can arise through an unsupported inference rather than through the fabrication of a fact. AI-assisted verification needs to examine not only whether a number is accurate but also whether the reasoning connecting that number to a broader conclusion is justified. In this instance, the central verification task is to assess the relationship between the reported figure and the conclusion drawn from it.

Three observations emerged from the five cases. First, the AI responses generally identified credibility cues, including realistic images, numerical details, institutional references, and professional presentation, as features requiring further verification rather than as sufficient evidence of reliability. Second, the responses addressed some verification tasks, such as checking dates, image provenance, institutional information, and statistical sources, more directly than tasks requiring contextual interpretation or evaluation of inferences. Third, the cases illustrated how misleading information may arise from relationships between facts, sources, images, dates, and conclusions rather than from entirely fabricated content. These observations are specific to the cases analyzed and do not constitute a general measure of AI verification performance.

The linguistic evidence helps explain how these credibility cues operate. In the first case, “صورة تعبيرية” frames the evidential status of an image. In the unemployment case, “البطالة في الدول العربية 2026” combines a statistical concept, a geographical category, and a precise temporal reference. In the institutional case, “مبادرة جودة الحياة الأممية” invokes authority through institutional association. In the financial case, “رأس مال شركة إنتاج الأمونيا الخضراء” employs specialized terminology that may make an inference appear financially grounded.

Across these cases, lexical choices and grammatical framing contributed to the way information was presented. However, the linguistic form of a claim does not establish its truth. Rather, it provides a basis for examining how the claim is framed and how its apparent authority relates to the evidence available to support it. The analysis therefore treats credibility cues as features of information presentation, not as direct measures of factual accuracy or audience perception.

The Arabic responses also reflected aspects of the local context of the claims, including relevant institutions, terminology, events, and information sources. However, these observations are limited to the selected cases and do not establish whether similar patterns would occur across other Arabic and English verification tasks. The differences observed between the two language versions should not be interpreted as evidence that verification in one language is inherently more effective than in the other. Rather, the findings suggest that responses may differ in explanatory detail, source specificity, and contextual emphasis. Determining whether these differences are systematic would require a larger dataset and further comparative research.

The findings suggest that AI-assisted verification can help identify claims requiring further investigation, suggest relevant sources, and highlight missing contextual information. However, identifying a verification need does not necessarily mean that a claim has been adequately verified. Verification also requires assessing whether the available evidence is relevant to the claim and whether the conclusions drawn from that evidence are justified. As research on large language models and fact-checking has highlighted, fluent explanations should not be treated as proof of factual accuracy (Augenstein et al. 2024). The present study contributes to this discussion by examining how an AI system evaluates misleading information in Arabic and English, with particular attention to linguistic framing, source attribution, and contextual reasoning.

The findings have several implications for digital journalism. First, visual transparency is important when AI-generated or AI-assisted images are used in news reporting. Clearly distinguishing illustrative images from documentary photographs and disclosing AI involvement where relevant can help clarify the evidential status of visual material.

Second, source attribution requires independent verification. Numerical information, institutional references, and apparently authoritative statements may contribute to the credibility of a claim, but journalists need to establish whether the cited source actually supports the information presented.

Third, contextual information should be treated as an integral part of verification. The Roman Theatre case illustrates how an authentic image can be misleadingly associated with a different event, while the green-ammonia case shows how a financial figure can be used to support an unjustified conclusion. Both cases demonstrate that misleading information may arise from the way facts are contextualized and interpreted, rather than from the fabrication of individual facts.

Fourth, linguistic framing deserves attention in verification workflows. Expressions such as “صورة تعبيرية,” “البطالة في الدول العربية 2026,” and “مبادرة جودة الحياة الأممية” contribute to how information is presented, but they do not independently establish its truth. Verification processes that focus exclusively on factual propositions may therefore overlook how lexical choices, temporal references, and institutional terminology contribute to the appearance of credibility.

Finally, AI-assisted verification should support, rather than replace, journalistic judgment. AI systems can help formulate verification questions, organize information, and suggest sources for consultation. Journalists remain responsible for assessing the relevance and reliability of the evidence and determining whether it supports the claim under investigation.

Theoretically, the findings support a qualitative approach to credibility analysis that considers linguistic and visual presentation alongside evidence, attribution, and context. The five cases illustrate how credibility cues may contribute to the appearance of authority without establishing factual accuracy. By examining these cues alongside AI responses in Arabic and English, the study connects linguistic analysis with research on AI-assisted verification. Its contribution lies in bringing these elements together within a common analytical framework and demonstrating the importance of contextual reasoning when evaluating potentially misleading digital information. Given the small purposive sample and the use of a single AI model, however, the framework should be understood as exploratory rather than as a validated or universally applicable model.

5. Conclusions

This study examined how credibility cues operate in five AKEED-documented misinformation cases and how an AI system evaluated the claims in Arabic and English. The findings indicate that the responses identified relevant verification needs, including checking sources, dates, visual provenance, and numerical information. However, the responses also showed limitations when evaluating claims that depended on contextual interpretation or inferences drawn from otherwise accurate details. These findings address the research questions by illustrating how credibility cues appear in the selected cases and how the AI system responded to them under the evaluation conditions of this study.

The study’s main contribution is a qualitative framework for examining AI-assisted verification through credibility cues and related analytical dimensions, including evidence evaluation, source selection, uncertainty management, and contextual reasoning. By comparing responses in Arabic and English across different types of misleading information, the study connects linguistic analysis with the practical evaluation of AI-assisted verification. The findings suggest that the usefulness of AI in this context depends not only on identifying potentially questionable claims but also on evaluating the evidence and context needed to assess them.

The findings must be interpreted in light of the study’s limitations. The analysis was based on five purposively selected cases and one AI model under specific prompting and evaluation conditions. The cases are not representative of all misinformation in Jordan, and the study does not establish general differences in AI performance between Arabic and English. The qualitative analysis was conducted by a single researcher without independent second-coder assessment. Future research could examine a larger and more diverse set of cases, compare multiple AI models, and incorporate independent coding to assess the consistency and transferability of the framework.