Rodriguez and Dimitrova’s (2011) four-tiered model is one of the most cited works in the literature on visual framing (Walter & Ophir, 2024), as it moves beyond atheoretical or exclusively semiotic approaches to engage with how image selection, cropping, and editing can capture attention, evoke emotions, carry meanings, and influence perceptions (Coleman, 2010; Geise & Baden, 2015). Their model disaggregates visual framing into four tiers. The denotative level constitutes meaning by identifying salient frames through who and what an image depicts based on the scene and surrounding texts (Schwalbe, 2006). Denotative frames can be context-specific (Goffman, 1974, 1979) or function as equivalency-based dyads, such as gain versus loss and war versus peace (Cacciatore et al., 2016). The semiotic level suggests social meanings by examining visual grammar and conventions, including camera angles, perceived viewer distance, body posture, eye contact, and facial expressions (Forgas & East, 2008; Hall, 1966; Kress & Leeuwen, 2006). The connotative tier considers abstract and figurative symbols (e.g., metaphors) in the shot that can “combine, compress and communicate social meaning” (Rodriguez & Dimitrova, 2011, p. 96). Finally, the ideological level expands on the visual elements, stylistic choices, and symbols to provide a holistic interpretation that underscores the political, religious, economic and/or demographic underpinnings of the visual constructions. Combined, the four-tiered model allows for a nuanced understanding of visual messaging.

Studies of visual media campaigns by Islamist extremist groups have examined all four levels of the 4-tiered model, albeit with less emphasis on semiotics. At the denotative level, for example, Al-Qaeda’s visual campaign encompassed predominantly militant visual frames, such as training, operations, and martyrdom (Farwell, 2010; Center, 2005), highlighting political Islamist connotations associated with objects like flags and swords (Coleman, 2006). At the semiotic level, studies of ISIS dissected several pictorial stylistic conventions, including viewer distance, camera angle, eye contact, facial expressions, point-of-view shots, and dynamic versus static imagery (e.g., Impara, 2018; Winkler et al., 2019). At the connotative level, ISIS and al-Qaeda utilized symbols like AK47s, the monotheism hand gesture, and depictions of death and dying to convey meaning (Wignell et al., 2017; Winkler et al., 2018). Ideologically, both al-Qaeda and ISIS espoused an extreme Islamist lens steeped in a clash of civilization narrative that promoted religious rule as an alternative to the nation state system (Ciovacco, 2009), but had differing views on the nature and timing of the Caliphate (Kuznar, 2015). Yet, existing literature on al-Qaeda and ISIS’s media campaigns, while providing nuanced understandings of the four visual framing tiers, has been mainly limited to manual coding.

Computational visual analysis on its own is not yet capable of fully engaging the four-tier model of visual framing. Computational attempts identify some denotative and semiotic elements (e.g., humans, race, age, color, public figures, rifles, umbrellas, facial expressions) (Chen et al., 2022; Joo & Steinert-Threlkeld, 2022; Muise et al., 2022; Zhang & Peng, 2022). Yet, such analyses stop short of gauging key symbolic and ideological visual components of scenes and their contexts due to what Peng, Lock, and Salah (2024) rightly argue is an automation-theoretical disconnect. This study addresses this gap by comparing manual and automated coding in the online extremism sphere in pursuit of a more effective, hybrid approach.

Traditional computational image analysis typically relies on task-specific computer-vision architectures such as YOLO-style object detectors, Mask R-CNN segmentation models, or transformer-based vision–language models like BLIP-2 or Flamingo. These systems are designed to extract concrete visual features (e.g., objects, faces, scene layouts) and perform discrete tasks with high accuracy when trained on large, labeled datasets. However, they are not built to apply multi-layered interpretive schemas such as Rodriguez and Dimitrova’s four-tier model without extensive task-specific fine-tuning. Because our study evaluates GPT-4o in a zero-shot setting—asking a general-purpose vision-language model to follow a human-developed codebook—we position this approach as complementary to, rather than a replacement for, traditional CV pipelines.

AI changes related to the scanning, sampling, and quantization of visual images are rapidly evolving, rendering accurate summations of developments difficult (Zhang & Dahu, 2019). Complicating the quickly changing terrain is that machines are now producing most images, often for other machines rather than the human eye (Paglen, 2019). Nonetheless, AI remains data-driven, rather than image-driven, meaning that understanding image-data relationships should remain a priority (Anderson, 2017). Whether computerized or not, visual cultures influence and are influenced by human biases in both production and consumption of online messaging (Bridle, 2023; Sezen, 2020). Combining quantitative and qualitative analyses of big data can augment visual framing. Dondero, for example, maintains that large-scale quantitatively produced diagrams produced through quantitative and semiotic analysis can assist in identifying “contrasting areas, opposite areas, or superposing of images on the plane of expression” useful for further quantitative and qualitative analysis (Dondero, 2019, p. 140). Such a combined approach preserves the importance of visual context within a specified corpus and across the image components. In short, she advocates for scholars to use her process as a metavisual device for four reasons:

Here, we agree with Dondero (2019) about the value of computerized quantitative analysis for assisting qualitative results. However, we add that quantitative human coding analysis, combined with statistical assessments, can further strengthen the tracked meaning of results. We begin by asking:

Large language models and NLP pipelines can efficiently process large text corpora, grouping semantically similar responses and surfacing recurring patterns. Gamieldien et al. (2023) find that transformer-based tools can generate highly granular codes across thousands of student reflections, substantially reducing human labor. This aligns with earlier infrastructure-oriented work showing NLP can automatically classify predictable text categories. Similarly, Morgan (2023) reports that ChatGPT performs well when themes are concrete and descriptive, requiring little interpretive inference. In hybrid interfaces, rule-based suggestions can be systematically extended to unseen data (Rietz & Maedche, 2021), increasing agreement rather than replacing human interpretation and underscoring that automation is most reliable for patterned, literal, and structurally evident meaning.

Despite scalability gains, current AI systems appear to underperform when meaning depends on tacit knowledge, ambiguity, or socio-cultural interpretations. Gamieldien et al. (2023) note the need for AI researcher oversight when semantic nuance matters. Studies of thematic automation note that disagreement among humans themselves reflects interpretive pluralism (Armstrong et al., 1997; Mackieson et al., 2019) — something models are poorly equipped to resolve. Transformer architectures excel at long-range dependencies (Lakretz et al., 2020), but still primarily attend to textual features rather than framing context, affective tone, or symbolic cues. These limitations indicate that interpretive coding requires judgment beyond probabilistic associations, particularly in indexical, connotative, or historically situated domains.

Across studies, researchers express caution about fully delegating thematic interpretation to automated systems. Marathe and Toyama (2018) report reluctance rooted in opacity, loss of theoretical accountability, and few questioning opportunities (Chen et al., 2018). Rietz and Maedche (2021) similarly find researchers use automated suggestions not to accelerate coding, but to reflect on needed codebook refinements. Such reflection aligns with iterative qualitative traditions emphasizing continuous interpretation (e.g., Braun & Clarke, 2006; Saldana, 2021 cited in Gamieldien et al., 2023). Thus, for tasks requiring contextual inference, socio-cultural reading, or interpretive framing, human coders continue to outperform computational models. Because visual framing often requires reading symbolism, composition, affect, and implied narratives, we ask:

To assess the effectiveness of GPT-4o for applying a human-developed content-coding schema to visual framing analysis, we began by conducting a human-coded content analysis of 7,292 images from al-Qaeda and ISIS’s English and Arabic magazines or newsletters distributed 2009-2020 (see Table 1). For al-Qaeda, the English issues included Inspire (1-17) and Jihad Recollections (1-4), and the Arabic issues were al-Masra (1-57). ISIS’s English issues included Dabiq (1-15) and Rumiyah (1-13), and al-Naba (1-229) in Arabic. All items were publicly available through Google, Jihadology (Zelin, 2021), or archive.org.

We utilized 13 expert coders from Egypt, Afghanistan, Turkey, Saudi Arabia, Syria, Poland, Vietnam, and the United States to create, refine, and utilize a visual analysis codebook. Our human coders had doctorates or graduate training in Communication Studies, Psychology, Political Science, and Education. The pilot phase involved three US and Egyptian coders who created coding categories inductively from images in Dabiq’s first issue until intercoder reliability was higher than 0.80 on Cohen’s kappa for all variables. Coders met weekly to identify and resolve discrepancies and cross-cultural differences that produced unacceptable reliability levels. In cases of disagreements, a bias toward the Middle Eastern perspective prevailed in line with the primary target audience. With a reliable codebook, 13 coders received oral and written training and analyzed each image in the dataset. The average intercoder reliability score using Cohen’s kappa across all categories was 0.91 (see Table 2). A third coder resolved discrepancies for statistical analysis.

We sorted our human coding categories into four levels of visual framing. While Rodriguez and Dimitrova (2011) note that the four tiers can overlap, we focused on the level that three coders considered most suited to the images’ characteristics. The meaning of denotative elements (or objects that bore an indexical relationship with an individual, thing, or place) involves two interrelated processes. First, the coding process accounts for elements the viewers can see in the shot, rendering them “closer to the truth than other forms of communication” (Messaris & Abraham, 2001, p. 217). Fully gauging the denotative meaning, however, requires generating frames that derive from a second process that looks not only at the textual context (Rodriguez & Dimitrova, 2011), but also into the syntactic relationships between images. Compared to words, images lack an explicit prepositional syntax, or the ability to propagate clear causal relationships, similarities, or other forms of connections (Messaris & Abraham, 2001). Our denotative categories included military fighters, death, humans, leaders, flags, and destruction.

The second level of semiotics (or stylistic, technical, and subject conventions) focuses on “signs and symbols, sign systems, and sign processes” (Moriarty, 2002, p. 20). Visual semiotics fulfills three meta functions: compositional, representative, and interactive (Kress & Leeuwen, 1996, 2006), as it emphasizes four types of signification: arbitrary (by convention), memetic (by iconic representations), evidential (by cues and codes), and signaling (by recognition) (Moriarty, 2002). Metaphors or metaphorical thinking (Feng & O’Halloran, 2013), and visual metonymics linked to abstract concepts and objects/events are also associated with semiotics (Feng, 2017). Accordingly, our semiotic categories included viewer position, image position, viewer distance, eye contact, and facial expressions.

The third framing level of connotation involves visual symbols linked to ideas or concepts associated with individuals, things, or places. Turner insists symbology draws its data from “cultural genres or sub systems of expressive culture…as well as narrative genres, such as myth, epic, ballad, the novel and ideological systems [and t]hey would also include non-verbal forms” (Turner, 1979, p. 12). Symbols can allude to power, authority, faith, rituals, and death, among others, to achieve personal or group goals (Turner, 1974). At the iconographical level, visual symbols go beyond the depicted object or person to connote ideas and concepts; they begin to reveal ideological meanings derived from backgrounds and the surrounding context (Panofsky, 1955; Leeuwen, 2001). Our connotative categories included about to die, religion, and state, with the latter disaggregated into state-building, law enforcement, allegiance pledges, and media propaganda for a more fine-grained analysis (see Appendix A for category meanings; Table A1 for examples).

We removed several manual coding categories from our computational model. We excluded image size and position because the workflow operated on individual images rather than publication layouts. We omitted gender because the overwhelming number of individuals displayed were male, with females only as occasional outliers. We removed age and social infrastructure as neither produced significant results across more than a dozen papers addressing how message strategies intersected with situational factors.

To compare manual content coding with AI visual coding, we built a lightweight, fully reproducible inference pipeline that (1) ingested image files, (2) encoded each image and sent it to GPT-4o together with a structured labeling prompt derived from our codebook, and (3) compiled the model’s outputs into a standardized dataset for evaluation against human annotations. The pipeline did not fine-tune model weights; it relied on constrained prompting and schema validation to ensure consistent, interpretable results (see Figure 2).

RQ1 asked how effectively a vision-language model (GPT-4o) could apply a human-developed content-coding schema to Rodriguez and Dimotrova’s four tiers of visual framing. The AI coding performance differed in substantial ways across variables and framing tiers (see Table 3 and Appendix B Table B1). Denotative variables showed the strongest alignment between AI and human coding. Humans, Destruction, Leaders, and Flags demonstrated the most consistent performance across metrics. Humans yielded an F1 of 0.80 and a Balanced Accuracy of 0.83. Destruction performed similarly (F1 = 0.79, Balanced Accuracy = 0.83). Leaders and Flags showed moderately strong performance, with F1 values generally ranging from 0.70 to 0.78 and Balanced Accuracy typically above 0.75. Krippendorff’s α values for these variables were moderately reliable, ranging from approximately 0.40 to 0.60. As shown in Figure B1, these categories were dominated by absent cases, inflating agreement due to prevalence rather than consistent recognition of positive cases. Weighted F1 values were very close to F1 for denotative variables, suggesting that class imbalance had limited influence on AI performance in this category.

Denotative variables requiring additional contextual interpretation showed more modest alignment. Military Role achieved an F1 of 0.38, a Balanced Accuracy of 0.73, and a Krippendorff’s α around 0.30, indicating reliability for distinguishing between combatants and non-combatants below acceptable thresholds. Death showed somewhat higher performance (F1 = 0.52, Balanced Accuracy = 0.65), although still below that of other denotative variables. For both Military Role and Death, weighted F1 scores were notably higher than F1, indicating that performance was disproportionately driven by the majority “not applicable” class and the model struggled to identify less frequent positive cases.

Semiotic variables requiring interpretations of bodily, expressive, or relational cues showed weaker correspondence. Viewer Position, Viewer Distance, Eye Contact, Stance, and Facial Expression produced F1 values generally 0.60-0.72, Balanced Accuracy values 0.72- 0.76, and Krippendorff’s α around 0.20 to 0.40, falling below the 0.67 acceptable threshold for exploratory research. Across all semiotic variables, Balanced Accuracy consistently exceeded F1, indicating that while the model could distinguish broad classes, it did not reliably identify positive instances. Weighted F1 values were consistently higher than F1, confirming that the predominance of absent codes influenced performance. These semiotic findings suggest that discerning gaze direction, bodily posture, or facial expression requires contextual and relational sensitivity that current image models do not robustly encode.

Connotative variables tied to symbolic or ideological meanings showed the weakest correspondence. State-building performed poorly (F1 ≈ 0.24, Balanced Accuracy ≈ 0.43, Krippendorff’s α≈ 0.05 to 0.10), indicating minimal reliability. About to Die and Religion displayed comparably low performance, with α values often approaching zero. Weighted F1 values for connotative variables were only slightly higher than F1, suggesting that performance limitations stemmed not only from class imbalance but also from the framing elements’ conceptual and inferential nature. Connotative categories contained very few positive cases, limiting model exposure and contributing to low performance (see Figure B1). While the model detected overt symbolic cues, it struggled with implicit or context-dependent signals of state-building, martyrdom, or religious practice that require interpretive inference beyond visual pattern matching.

RQ2 asked which dimensions of visual framing remain resistant to automation, and what these limitations reveal about the strengths of human coding in visual analysis.

Lesson #1: Denotative Interactions

Despite AI’s stronger performance at the denotative level, inaccuracies and omissions were present. Military roles, for example, showed only modest agreement levels, as manual coders examined taglines and accompanying text to distinguish al-Qaeda and ISIS militants from enemy fighters. Without these textual cues, AI required more training on choice of uniforms or human intervention. Additionally, AI could detect ISIS’s objects like coins, outdoor markets, and competing currencies, but required researchers to recognize their connotative meaning, such as ISIS’s desired frame of economic independence. Similarly, AI could detect bottles of alcohol, drugs, and cigarettes (see Figure 3), but required human coders to recognize them as critical components of ISIS’s moral policing apparatus. In short, the addition of human coding can render computational methods less likely to miss key objects or elements in big datasets and more likely to properly assess their functions.

Another key contribution of manual coding to AI processing involves aggregation of denotative elements. Our AI learning approaches generated disparate visual elements, such as militants fighting or training, raised index fingers, swords, maps, doctors treating patients, and sunsets, etc. Human coding, however, captured subtle visual relationships that revealed broader themes of military prowess and state-building, identified how relationships created cohesive narratives, showed how symmetry, repetition or alignment influenced viewer interpretations, and unveiled the visual syntax strategy. For example, human coders identified the interrelationship between images of beheadings, piles of cigarettes, and checkpoints as ISIS’s law enforcement apparatus, complete with punishments for alleged spies, the moral policing apparatus, and the access points for determining who could enter the caliphate. Entman’s framing associations were useful to unlock the messaging strategy. The visual law enforcement frame communicated that sins were rampant (problem definition) because of people smuggling contraband, committing treason, and ignoring Islamic rules (causal interpretation), which stained the society (moral evaluation), thus requiring punishments and crackdowns on contraband to ensure community safety (treatment recommendation). An unsupervised learning computational approach on its own would not fully reveal the visual narrative.

Lesson #2: Semiotic Interactions

For the most part, AI performed poorly on visual semiotic elements, rendering human coding highly valuable in this domain. With Krippendorff’s α consistently lower than .67 between AI and human coding for viewer distance, eye contact, body stance, and facial expressions, the use of supervised AI was neither efficient nor cost-effective in analyzing the semiotics of al-Qaeda and ISIS’s images. Human training and validation, however, could help refine steps for identifying more useful, valid, and efficient processing of visual semiotics patterns. One example is the use of direct eye contact because it conveys dominance (Appendix A). From 2016 to 2018, ISIS used direct eye contact as a frequent visual tactic, but the use of the strategy differed based on whether the depicted subject was a friend or foe and if the language of publication was English or Arabic. Yet, supervised training for AI only produced a .54 Krippendorf’s α with manual coding. Intercoder reliability difficulties related to eye contact suggest that a carefully constructed, rule-based definition of what constitutes eye contact and avoidance would be necessary to reduce noise in the AI extraction process.

Human coding could also help narrow categories down to coding options most useful for understanding visual strategies. For example, semiotic understandings of viewer distance focus on four categories: intimate, personal, public, and social (Jewitt & Oyama, 2008). Each conveys specific meanings associated with standard human interactions (e.g., intimate distances associated with distraught photo subjects; photo subjects shot at a public distance conveying group rather than individual identities). However, human coding and statistical analysis revealed that significant findings primarily appeared only after combining intimate and personal distance (with photo subjects photographed from less than four feet) and comparing them with the combined categories of public and social distance (i.e., greater than four feet). Such findings help avoid overlooking important patterns that could be missed with strict adherence to viewer distance’s four standards.

Lesson #3: Connotative Interactions

While computational coding could efficiently identify key symbols, human content expertise helped interpret needed cultural genres and subsystems. The black flag, for example, is a transhistorical object al-Qaeda and ISIS used as a symbol of adherence to Islam and the Prophet Muhammad’s path. It typically features the words “No God but Allah” with a white circle beneath carrying the words “Muhammad is Allah’s Messenger.” Yet, al-Qaeda often deviated from standard black flag depictions, also featuring white and other emblems used in Afghanistan and elsewhere (see Figure 4). Without such insights, computational extractions of flags as denotative elements would miss other symbolic variations essential for understanding the nuance of extremist groups’ visual messaging. Another frequent symbol in al-Qaeda and ISIS photographs was militants raising their index fingers. The diverse makeup of our manual coding team, including Muslim researchers, identified the gesture as part of Islamic culture, connoting monotheism and the dedication of deeds to the one God. Culturally informed, human content expertise was instrumental for properly training and validating computational visual analyses to generate insightful media campaign understandings.

When collaborating with AI, another beneficial area for human coders is adding useful insights about tropes and other visual commonplaces. A prime example in al-Qaeda and ISIS’s media was the excessive reliance on the about to die visual trope, which appeared in 75 percent of their images. Yet, AI missed many instances of the trope. About to die images assume three forms: presumed (i.e., showing implements of death like weapons and destruction), possible (i.e., showing photo subjects potentially dying without a confirmation their death occurred), or certain (i.e., showing photo subjects with accompanying text confirming their deaths) (Zelizer, 2010). An unsupervised learning method would likely group all three forms under a military visual frame, hence not distinguishing between the three constructs nor accounting for the groups’ emphasized use of the three disaggregated strategies. Instead, breaking down each form into its own core denotative components could facilitate training and validation. Labeled data for supervised learning could account for objects or elements, such as blood, swords, knives, guns, AK-47S, tanks, armored vehicles, rockets, ammunition, fire, smoke, explosions, and sniper crosshairs. Human coding could then complement the computational analysis by grouping the three about to die clusters, each with its own unique viewer interactions (Zelizer, 2010).

Lesson #4: Ideological Interactions

Increasingly, scholars have recognized the linkages between ideology and discourse. Fabiszak et al. define ideology as “systems of beliefs, shared by a social group, with the power to evaluate and explain the social world” (Fabiszak et al., 2021, p. 409). McGee adds that “ideology in practice is a political language, preserved in rhetorical documents, with the capacity to dictate decision and control public belief and behavior” (McGee, 1980, p. 5). Van Dijk (1998) explains that ideologies can influence the human mind’s cognitive structures. McGee (1980) posits that a full set of ideological propositions can be summed up in a single term, while Edwards and Winkler (1997) extend such reasoning to a small set of visual images.

Human coding can help AI users distinguish between visual markers of culture and other images not performing ideological functions. One example of how this process could work concerns ideographs. In his study of American culture, McGee (1980) defines a small subset of positive and negative words as ideographs (e.g., freedom, liberty, slavery, and terrorism). They serve as ordinary terms in political discourse, have abstract meanings that allow for collective commitment, warrant the use of power, guide behavior, and have culture-bound meanings. Consideration of the ideograph’s characteristics within social groups can assist AI users in narrowing large corpuses of visual images to specific objects that serve as cultural markers. With al-Qaeda and ISIS, for example, one visual ideological code is the group’s display of the monotheism gesture, whereby Muslims point their index fingers upwards toward heaven to connote their adherence to Allah. Al-Qaeda and ISIS frequently utilize images of the same gesture to signal potential Muslim recruits sympathetic to their groups’ causes.

However, a key function of human coders in the AI training process involves understanding the interactions of visual elements of ideographs within an image. Returning to the monotheism gesture example, AI would be able to scrape all images showing humans pointing their index finger upward. Such an approach, absent human coders’ insights, would initially yield many images of Muslims demonstrating their Islamic faith with no affiliation with extremism. AI might also retrieve images of athletes or other individuals raising their index finger to signify success and victory (see Figures 5 & 6). Rather than confound the results with too much noise to produce meaningful results, AI trainers could refine the scraping process by asking for the gesture along with the presence of militants, males, direct eye contact in photo subjects, personal distance, black flags, and number of humans, as each of these variables have a documented relationship with the monotheism gesture in extremism photographs (Winkler, 2022). By considering element constellations rather than single objects or elements, the dataset will become more accurate in discerning ideological frames.

Lesson #5: Image-Context Interactions

Context is a multifaceted resource involving a myriad of forms that can influence interpretations of texts, whether discursive or nondiscursive (Linell, 1998). It both shapes and is shaped by its textual interactions. As McGee notes, “Failing to account for ‘context,’ or reducing ‘context’ to one or two of its parts, means quite simply that one is no longer dealing with discourse as it appears in the world” (McGee, 1990, p. 283).

The complicated interactions between images and contexts in relation to al-Qaeda and ISIS emphasized the need for human coders to supplement AI for meaningful, efficient messaging processing. To begin, human coding aided in the identification of image codes that corresponded to changes in context factors over time. As Appendix C, Table C1 shows, 18 of 30 variables in our manual codebook bore a significant relationship to context variables (e.g., troop withdrawal announcements, online account suspensions, attack lethality, etc.). Those relationships suggest a productive, efficient training regimen for AI, as the context factors are revealing about the groups’ response interactions over time. The remaining coding categories, while potentially useful for assessing image meaning or relationships with other images, did not significantly change in frequency over time, making them less of an AI priority.

Human coding can also help verify the appropriate AI scope for assessing impact of context variables on strategic image use. The table reveals that the relationships with image strategies vary according to the context factor under consideration. As a result, the outputs of human coding point to context variables that need verification checks prior to premature conclusions that any single context variable is responsible for shifts in the image characteristics. For example, since censorship of militant group online accounts, announcements of anticipated troop withdrawals, and the relative rise in standing of competing ideological groups all relate to significant changes in the use of photo subject distance, users of AI should consider whether each of these context factors overlap during the timespan under evaluation before drawing any conclusions about the potential influence of any single context element.

Human coding can also yield insights regarding appropriate AI disaggregation of context elements in relation to image characteristics over time. For example, our assessment of leader loss based on human coding revealed that the deaths of different levels and types of ISIS leaders resulted in different corresponding changes in the group’s image characteristics. Political and military leader deaths corresponded to different image changes. Leaders at different statuses within the media hierarchy corresponded to different changes in visual strategies. Thus, fine-tuned human coding analysis can aid in the development of a more robust, efficient AI system for analyzing the image-context strategies of groups like al-Qaeda and ISIS.

This analysis demonstrates the advantages of having humans and AI functioning together to understand the visual framing of extremist group messaging. Human coding yielded benefits on the retrieval, analysis, and validation of results related to denotative, semiotic, connotative, and ideological framing. It also aided in understanding how AI-human interactions can maximize text-context relationships. Yet, validation checks between the human and AI coding revealed that the percent agreement levels on coding categories varied considerably, suggesting the need for a more robust AI training process, particularly on subjective variables, like facial expressions and perceived distance, and identifying visual constellations linked to connotative and ideological framing.

Examining the extremist visual context provided an ideal opportunity to test and compare manual and computational coding for MENA-based violent groups, but it does not necessarily apply to other types of violent protest visuals. The generalizability of the findings derived from al-Qaeda and ISIS photographic campaigns should be tested in relation to other forms of political violence (e.g., electoral protests, climate activism, and public vigils). Future studies should determine if the reliability of AI visual framing is transferable to these other settings.

Additionally, variables with inherently nuanced or subjective definitions, such as stance or eye contact, pose significant challenges for consistent annotation. These complexities are reflected in the low recall and F1-scores observed in these categories, as the AI model struggles to align with human coders’ interpretations. The limitations of computational methods in capturing subtle cultural or contextual cues further exacerbate these discrepancies, particularly for categories like state, religion, and impending death that rely heavily on contextual understandings. Future studies should examine alternative training protocols to maximize efficient and reliable extraction processes.

Data Availability: Replication materials for this study are hosted on the Open Science Framework (OSF). Due to the presence of violent and potentially harmful imagery, the image data are archived as a restricted-access component on OSF and may be accessed upon request, subject to review and approval: https://osf.io/r9tjm, project DOI: 10.17605/OSF.IO/R9TJM. All non-sensitive replication materials—including the coding instrument, model prompts, variable definitions, documentation, and analysis scripts—are publicly available on OSF and are also mirrored on GitHub for ease of access and version control: https://github.com/aysedeniz09/Visual-Framing-in-the-AI-Era.