Cultural and multimodal realization in various AI models: text to image, image to text

1 Introduction

Visual and written communication has been greatly affected by the advent of AI. With the recent advances of multimodal artificial intelligence, a new generation of models have emerged that are capable of both describing images in textual prompts, as well as generating images from text (Cao, 2025). Image to text AI models are trained to capture images and transform them into language description like humans' text descriptions (Envisioning, 2014). The main feature of these models is that they can comprehend the visual image through search engines and, at the same time, generate texts through the integration of computer vision and natural language processing (Envisioning, 2014). There are text to image AI models that are trained based on a large number of images and text datasets to break into parts, accurate patterns, and linguistic features. Text to image AI models need to translate the textual prompts that carry semantic meaning into a visual output that has a high degree of realism (Ghazvineh, 2024; Sobey, 2025). Suchman (2006, p. 6) proposes that AI “works as a powerful disclosing agent for assumptions about the human”. This is true as “genAI images can be understood to be a social artifact, propagating a specific form of social imaginary derived from their data and the rules they adhere to” (Sobey, 2025, p. 1). The recent availability of AI imagery to the public calls for more empirical research on AI generated images in order to refute or support claims of social bias within AI reproduction (Westberg and Kvåle, 2025).

The ability of AI to generate texts from images and images from text prompts has facilitated many practices that previously required time, effort, money and mediation. Many sectors, including marketing, have benefited from this technological revolution. It is estimated that the market in generated AI would grow to 1.3 trillion US dollars within the next 10 years (Bloomberg, 2023). Many agencies are already using AI generated images to advertise their products (Bain, 2023; Hero, 2023; Maison Meta, 2023). These advertisements not only communicate marketing messages but also create a cultural experience for consumers that carries a number of social and ideological perspectives. A major concern with AI generated texts and images though is their degree of sensitivity, as well as the representation accuracy of cultural facets (Hurley, 2024; Sobey, 2025; Westberg and Kvåle, 2025). Having specific cultural norms, Saudi Arabia has excelled in preserving traditions and respecting Saudi heritage. All Saudi sectors carry this vision into their own practices. Product companies ensure that the way they advertise within Saudi Arabia adheres to Saudi traditions, reflects strong ideologies and represents culture accurately (Nassif and Gunter, 2008). An exploration of how AI generated text and images combine marketing concepts with Saudi culture is therefore important for marketing agencies, linguists and computer program engineers.

This study explores AI generated texts and images within the framework of social semiotic technology (Westberg and Kvåle, 2025). Such a revolutionary expansion of AI blends with human creativity and requires a deep understanding of how the processes that control the production of concepts and ideologies are translated through different representational modes (Ghazvineh, 2024). Hence, this study views AI models as semiotic affordances that create meaning and reflect social norms. The comparison of generated images and texts from different AI models is followed by an assessment of AI models in relation to their representation of a number of social and linguistic variants. To the best of our knowledge, this is the first study that holistically analyzes the sequence of AI textual prompts and generated images through the lens of multimodality.

2 Literature review

2.1 Studies in language and AI

Recent developments of AI imagery means research in this area is scarce, yet what is available offers findings about social representation (Westberg and Kvåle, 2025). Few studies have looked at different representations generated by AI images. Some studies conducted empirical studies in an attempt to understand how AI takes in text and transforms it into visual images. Thomson (2025) examined how language models describe two data sets from journals. Descriptions were then fed into the same language model to understand how AI tools recreate these images. Westberg and Kvåle (2025) looked at the generated AI images for the prompt “teenagers”. They conclude that AI tends to create images that are authentic in context and reveal teenagers of different ethnicity in a positive and relaxed mood. However, Westberg and Kvåle's (2025) “romanticizing” of AI images is not in line with other studies such as those of Chun (2025), Hurley (2024) and Putland et al. (2025), who all highlight some negative representations.

AI image generation has also been studied in relation to social concepts such as stereotyping and female representation. Chun (2025) shows how AI generated images are stereotypical and gender biased with the help of content analysis tools. Such findings recall Goffman's (1979) typical gender representation. Another social dimension is representation of family in AI generated photos. Hurley (2024) investigated text to image generation of family portraits from a collected corpus and identified techno semiosis “hallucination” as a recurrent aspect that can be categorized into three types: “perceived errors”, “distorted view of reality”, “regimes of Whiteness”. The study identified various racial issues and calls future studies to address them. Stereotyping and racial biases are among ideological representations that are also detected in Putland et al. (2025). This study focuses on how the idea of dementia is visually represented through AI generated images from the prompt “dementia” and found a lack of diversity and negative perspectives on the disease.

2.2 Social semiotic technology and multimodal discourse analysis

Social semiotic technology is an extension of Multimodal Discourse Analysis (MDA), adding a technological dimension to the analytical framework. Hence, a combination of tools are utilized to analyze data that includes linguistic, social, as well as technological artifacts. MDA originates from Halliday's (1978) socio-semiotic approach that examines an amalgam of textual, social and cultural aspects. Two distinctive approaches are identified within the literature; the contextual approach of Kress and Van Leeuwen (2006), and the grammatical approach of O’Toole (2010).

The MDA approach that is adopted in this study is the contextual one (Kress and Van Leeuwen, 2006) due to its inclusion of the different ways of meaning making via textual, verbal and visual medium (e.g., color, modality, logos, font, etc.). According to Zhang, “The relationship between a multimodal text and context is dynamic” (2024, p. 2). A multisemiotic text is in a reciprocal relationship with contextual factors; however, the text establishes the context (Halliday and Matthiessen, 2014). Kress and Van Leeuwen (2006, p. 20) view multimodality as the “use of several semiotic modes in the design of semiotic product or event”. O’Halloran's (2011, p. 135) states that multimodal discourse analysis “reveals how instances of multimodal semiotic choices function inter semiotically in ways which ultimately create and answer to larger patterns of social context and culture”. Social semiotics revolves around representation of participants, action and location (Westberg and Kvåle, 2025). Social semiotic technology goes beyond MDA that analyzes texts and multimodal data to include the technologies that produce, shape and reshape data (Eisenlauer, 2013; Ledin and Machin, 2021; Westberg and Kvåle, 2025).

This article analyses the semiotic choices made in different AI models in a way that responds to the cultural contexts that the original images reflect. Adopting the social semiotic technology, MDA approach, this study addresses the following research questions:

• What cultural elements in advertisements are represented when generated by AI models in multimodal contexts?
• What are the strengths and weaknesses of different AI tools with regard to multimodal representation when generating advertisements?

3 Methodology

The aim of this study is to reach a deeper understanding of the social and cultural practices of semiotics technology (textual prompts and images) generated by different AI models through an exploratory case study limited to a specific time frame and context. Analysis evaluates whether modern AI models (both text-to-image and image-to-text) can understand, generate, and culturally adapt visual content that functions as effective advertisements, not just as generic images. To reach an equivalent comparison between the different AI models we chose models that have the same features, which allow for transforming the image into textual prompts and, at the same time, have the features to enable the transformation of the textual prompts into generated images.

Seven models were included: ChatGPT, Gemini, Meta AI, Grok, Qwen, Leonardo AI, and Recraft V3. These models were chosen not only because they support both image-to-prompt and text-to-image functions but because they are also easily accessible through online websites, making them practical for testing without the necessity of an advanced setup or coding. In addition, selection was guided by rankings from Artificial Analysis' Text-to-Image Leaderboard, which evaluates the performance and accessibility of current leading image-generation models (Artificial Analysis, n.d.). It is important to note that some of the 7 AI models that were used to produce the textual data have sub models to create the image data resulting in 9 AI models: Google includes two image generating models i.e., Google Imagen 4 and Google Gemini 2.5 Flash (Nano Banana) whereas Leonardo AI includes both Phoenix 1.0 and Lucid Origin.

The data collection stage took place between October 1, 2025, and November 1, 2025. Six advertisements from the Saudi dairy company Almarai were selected as the essential data for the current study. Almarai is one of the largest dairy companies in Saudi Arabia ranking 46 among top 100 most valuable companies in the Middle East (Forbes Middle East, 2006). The criteria for selection includes time frame, ads that are recently posted on the company's official Instagram page in 2025, with high views (10K+), and inclusive of a variety of Saudi cultural and social aspects such as dress, family, gender, age and setting. With regard to ethical considerations, the current data is brand-produced public communications and not considered private. Therefore, the data is “fully replicable and raises no concerns regarding informed consent or data protection” (Park, 2026, p. 4).

For each model, the same prompts were given in both stages. First, we uploaded the six advertisements and gave the AI models the same prompt, “turn it into a text prompt” (see Figure 1 for the process of the study). This provided an outcome of 42 textual prompts. Then, using a different account of the same model so it would not use memory or context, we uploaded the generated textual prompts and gave the AI models the same prompt, “create an image”. When the AI company had multiple image models, we used the same generated prompt for each one. For example, with Gemini, the same generated prompt across Google's different image generation models [Google Imagen 4 and Google Gemini 2.5 Flash (Nano Banana)] was used. Similarly, with Leonardo AI, the same prompt for both Phoenix 1.0 and Lucid Origin models was used. Thus, the image-generation component of this study included nine contemporary AI models. The outcome for this process was 54 generated images (see Supplementary Appendix Table S1 for an overview of all the AI models included in this study). The discrepancies in the numbers of textual and image data is due to the differences in AI textual and image model numbers explained above.

Generated textual prompts and images were assessed using open reading inductively to interpret the different represented semiotic resources (Kress and Van Leeuwen, 2006). We concluded that the semiotic resources found within the generated textual prompts and images are never fully constrained, although a specific context can be represented (Westberg and Kvåle, 2025). In other words, the meaning making process involves “ways that have become historically established and that are routinely employed for specific purposes” (Ledin and Machin, 2018, p. 9). When discussing visual and linguistic strategies, Machin and Mayr (2012) state that even if they seem normal or neutral on the surface, it is important to be aware that they can carry inherent ideology that can be used to shape the representation of events and persons for particular ends.

4 Reliability analysis

Intercoder reliability was assessed to establish whether the qualitative judgements used to construct the quantitative scores are applied consistently. The two authors independently coded each output using the same coding framework and without access to the other author's decisions. Each observation represented one model, advertisement and semiotic feature, which was coded as 1 where criterion was fulfilled and 0 when it was not fulfilled. It produced 906 paired decisions: 420 for the image to text stage and 486 for the text to image stage. Reliability was calculated in Stata before the authors discussed any disagreements, preserving the independence required for an assessment of agreement. The analysis combined observed percentage agreement with Cohen's kappa, which adjusts observed agreement for the agreement expected by chance (Cohen, 1960; Klein, 2018). Separate estimates were calculated for both stages.

Cross classification of coder decisions is provided in Table 1. The authors assigned 0 on 324 occasions and jointly assigned 1 on 509 occasions, producing 833 agreements. They disagreed on 73 decisions: Coder 1 assigned 0 and Coder 2 assigned 1 in 29 cases, whereas the opposite occurred in 44 cases. Coder 1 recorded 553 positive decisions compared with 538 for Coder 2. This difference suggests a more conservative coding threshold for Coder 2, although the difference of 15 judgements is small across 906 observations.

Table 1. Cross classification of coder decisions.

Overall observed agreement was 91.94% (Table 2), compared with expected agreement of 52.07%. Cohen's kappa was 0.8319, with a 95% confidence interval from 0.7949 to 0.8689. Under the applied Landis and Koch (1977) categories, the point estimate indicates almost perfect agreement, the lower confidence limit remains within the substantial range. Reliability was consistent across stages. Image to text coding produced 92.14% agreement and a kappa of 0.8224, while text to image coding produced 91.77% agreement and a kappa of 0.8342. The close correspondence between these estimates indicates that coding reliability remained stable across both directions of transformation.

Table 2. Observed agreement and Cohen's kappa.

All kappa coefficients were statistically significant at p < 0.001.

Table 3 tests whether this conclusion depends on the choice of reliability coefficient. Overall estimates ranged narrowly from 0.8318 for Scott and Fleiss' pi to 0.8453 for Gwet's AC1. Krippendorff's alpha was 0.8319 and the Brennan Prediger coefficient was 0.8389. The same convergence appears in both stages, where all coefficients exceeded 0.82. Gwet's AC1 is especially useful as a check where category prevalence may influence kappa, whereas Krippendorff's alpha provides an assessment of coding consistency (Klein, 2018).

Table 3. Alternative reliability coefficients.

Values in parentheses represent 95% confidence intervals.

Overall, the high agreement, strong chance corrected coefficients and closely aligned robustness estimates demonstrate that coding framework was consistently applied. Therefore, the quantitative comparisons reported from these qualitative judgements have a reliable empirical basis. The further discussion of disagreements did not change these reliability estimates, which were based solely on the original independent ratings.

5 Analysis

Analysis is divided into two sections. The first deals with the cultural element in advertisements generated by AI models. Seven AI text models were utilized for retrieving texts from 6 advertisements resulting in a total 42 texts. These texts vary in length ranging between 63 and 303 words. Analysis shows that a number of themes and cultural aspects are transferred. Aspects include description of color, intertextuality, logo, language (Arabic text), culture dress, people, modality, location, and action. In addition, the ability of the AI model to identify the image as an advertisement within the text is highlighted. However, some models tend to focus on some elements more than others. Another distinctive finding is the idea of “text hallucination”, which we explain more in detail below. In the second section, a descriptive comparison of the models in terms of accuracy in representing various cultural elements is analyzed. Data comes from 54 images generated by nine AI models from the textual input in the first stage. Similar to themes found in AI generated texts, color, logo, language (Arabic text), cultural dress, people, location, action and modality and “hallucination” these are all identified. However, intertextuality is one theme that is not found in the analysis of images due to nature of data, i.e., intertextuality is usually found within texts not images.

5.1 Representation of semiotic recourses in advertisements generate by AI models in multimodal contexts

5.1.1 Setting and location

In textual prompts, location refers to the description of setting, spatial placements of elements as objects, location of people, and their order, as well as position of people in relation to the camera. The description includes giving an overall label of the location such as “outdoor”, “kitchen”, “corporate hallway”… etc. In addition, position and location of people and objects in the image is also given; for example, “the son, seen from behind”, and “to the woman's right (viewer's left), a man in a white thobe is seated, looking towards the young girl”.

All the models provided accurate descriptions of location, except for Recraft V3, in which location was either not there or was inaccurately described. Some AI models were more detailed than others in describing description, which affected image generation later when the texts were used as the prompt to generate images. Only a few instances in the collected data describes location in relation to the camera; for example, a text generated by Qwen3-Max AI indicates that a woman “stands with her back partially to the camera”.

With regards text to image, location is important in executing the final AI image accurately. Images reflect location of people in relation to their surroundings, other people and different objects. The ability of AI models to identify the type of place including, for example, “supermarket”, “corporate”, “outdoor”… etc., is clearly reflected in the generated images. What is missing in many AI models is that description of gaze and camera angle is not indicated in the textual prompt, hence, creating a different visual output. Figure 2 shows how textual description can affect the generated visual image:

The above example shows men in a kitchen whereas the original ad is located in a corporate setting. This is due to the incorrect description of a number of elements, including location in the textual prompt: for example, “there are white kitchen cabinets”. In addition, the angle of the camera and the outlooks of individuals are different from the original image due to lack of specific information in this regard. Comparing the models, the majority [Google Imagen 4, AI, Google Gemini 2.5 Flash (Nano Banana), Meta AI, Qwen3-Max AI, Leonardo AI (Phoenix 1.0), Leonardo AI (Lucid Origin) and Recraft V3] could not produce a correct image location as described in the textual prompts. The only AI model able to produce an image with a setting similar to the textual prompts was ChatGPT 5 OpenAI.

5.1.2 People

Descriptions of people include biological features, race and gender. In the image to text data, recognition of people's demographics is a critical dimension that AI models provide in the prompts for reading the image. The majority of textual prompts generated by AI models [ChatGPT 5 OpenAI, Google (Gemini), Meta AI, Grok xAI, Qwen3-Max AI and Leonardo AI] had an accurate description of gender, age and the number of people. However, the model Recraft V3 indicated the wrong number of people in its descriptions.

Another finding that emerges is that AI models rely on descriptions of cultural dress to indicate race. For example, “thobe”, “Kandoorah” and “ghutra” are relevant to the Gulf Arabic culture. AI states racial origin based on dress as seen in the following examples: The above generated texts show that nationality is not mentioned in the first example (a.), yet correct names of traditional clothes in Saudi Arabia are given: thobe, shemagh, abaya and hijab. However, in example (b.) a contradiction occurs. The text identifies clothing as belonging to “traditional Saudi Arabian attire”. However, specific names for dress such as “checkered keffiyehs” that are mentioned in the text belong to Palestinian culture not Saudi Arabian. This contradiction offers a parallel to Hurley's (2024) explanation of “image hallucination” – distortion of reality in which cultural norms are misrepresented.

• “In the foreground, a happy family shops together — a man in a white thobe and red-and-white shemagh pushes the cart, a woman wearing a black abaya and hijab stands beside him, and two smiling children accompany them. The young girl, dressed in a checkered dress, leans playfully on the shopping cart, while the boy in a blue thobe stands near his father”. (ChatGPT 5 OpenAI)
• “A photo shows four men in traditional Saudi Arabian attire, wearing white thobes and red-and-white checkered keffiyehs, standing in a bright indoor space with a two-level building design”. (Meta AI)

Similarly, people is another main feature that is represented in the generated images from textual prompts. Relevant components such as age, gender, ethnicity and number are reflected. Based on the textual input, these features are transferred into the visual form. Interestingly, the data shows that some features mentioned in the textual prompt affect the emergence of others by association. For example, race is indicated by description of dress. On the other hand, inaccurate race representation (for example, Asian instead of Saudi) occurs in the current data due to lack of description of race and relying on dress only. Age is also problematic in many images due to lack of textual input. For example, one image shows a mother as a young girl; the text indicated the role “mother” rather than the age. Generated images also show variation in representation of the age of a boy; again, the text labels him as a “son” with no exact age indicated.

The number of people in almost all images is shown correctly; only one image added an extra person due to wrong number indication in the text (Recraft V3). Another image shows three men instead of four, even though the text states: “four men wearing traditional Middle Eastern attire” (Grok xAI). Such instances can be described as image number hallucination. Overall, people representation is provided accurately by four models [ChatGPT 5 OpenAI, Google Gemini 2.5 Flash (Nano Banana), Qwen3-Max AI, and Leonardo AI (Phoenix 1.0)]. The others [Google Imagen 4, Meta AI, Grok xAI, Leonardo AI (Lucid Origin), and Recraft V3] did not provide totally accurate people representation.

5.1.3 Action

AI textual models included description of action, including movement and activities in which people within the image are engaged. An example of action described in texts includes, “A clean, modern Almarai advertisement featuring a young woman preparing food in a bright kitchen” (ChatGPT 5 OpenAI). The comparison between the different textual prompts shows that Grok xAI, Qwen3-Max AI and Leonardo AI provided a good description of action. However, other models did not include enough details (ChatGPT 5 OpenAI, Meta AI) or had faulty action description [Google (Gemini), Recraft V3].

Actions of people are shown in most images as per the textual prompt. For example, “pushing a shopping cart” (see Figure 3) is represented in the final visual AI product. The data shows that accurate texts that provide detailed descriptions of what people are doing in the original image yields accurate action represented in the generated images. This is achieved by using action verbs such as “placing”, “smiling” and “handing”. To make these action more specific and accurate, additional adjectives and adverbs are used; for example, “gently placing his hand”.

In some cases, the AI model alternates the action in the image because the text provided an inaccurate description. Half of the AI models misinterpreted people's actions within generated images [Google Imagen 4, Google Gemini 2.5 Flash (Nano Banana), Meta AI, Qwen3-Max AI, and Recraft V3]. However, some AI models were able to produce the correct action [ChatGPT 5 OpenAI, Grok xAI, Leonardo AI (Phoenix 1.0), Leonardo AI (Lucid Origin)]. These AI models were able to produce accurate action as the generated textual prompts provided detailed descriptions.

5.1.4 Cultural artifacts

The extent to which AI generated texts reflect and accurately represents culture is another focus of this study, as well as the ability of AI models to generate texts that read and interpret cultural artifacts such as dress and identities. ChatGPT 5 OpenAI, Google (Gemini), Meta AI, Qwen3-Max AI and Recraft V3 all recognized culture accurately from the prompts. However, Grok xAI and Leonardo AI did not do so.

Cultural dress is a main dimension represented by AI models. Based on textual input, various cultural dress schemes were created and a number of cultural representation levels are displayed in the data. Cultural dress is linked to geographical locations indicated by textual prompts. One level is indicated by the larger circle, which is “Arab”, followed by “Middle Eastern”, then “Gulf region”, and finally “Saudi”. Dress is represented, or misrepresented, by the use of these words. Hence, stereotyping and collective identities are reflected in dress and race. Traditional Saudi clothes in the original images are replaced with Gulf, Palestinian, as well as Pakistani ones. This collective view lumps up all cultural varieties under one label, i.e., Arab or Muslim.

Similar to AI generated textual prompts, text to image data reflects cultural representations through dress (see Figure 2 above). Differences in AI generated images were found between the AI models, with most images missing the correct interpretation of cultural artifacts. ChatGPT 5 OpenAI, and Google Gemini 2.5 Flash (Nano Banana) provided good cultural representations, demonstrated by rich cultural cues (i.e., cultural meanings and identities) and details in the textual prompts resulting in reduced ambiguity. Google Imagen 4, Meta AI, Grok xAI, Qwen3-Max AI, Leonardo AI (Phoenix 1.0), Leonardo AI (Lucid Origin) and Recraft V3 all generated images with incorrect cultural artifacts. In some cases, the textual prompts lacked cues for cultural representation, whilst in others the cues were present, but the AI models did not produce cultural artifacts in the generated images.

5.1.5 Text/language

Analysis in this current study shows that language is problematic to some AI models. Models use a number of strategies to overcome linguistic challenges; some use “gibberish” (see Figure 4), others remove Arabic text (see Figure 5). However, English seems to be recognized by all AI models. Challenges occur only when other languages are integrated. Translation and transcription are also linguistic strategies used in AI generated texts. The original advertisement included an Arabic text in the logo (i.e., the name of the brand), as well as an Arabic slogan. The majority of textual prompts provided from ChatGPT 5 OpenAI, Google (Gemini), Meta AI, Grok xAI, Leonardo AI and Recraft V3 did not have any English translation of the Arabic text (see Figure 5). The model Qwen3-Max AI, however, not only provided an accurate translation of the Arabic text but also demonstrated the ability to identify the language as Arabic.

Manipulating language in AI generated images is another important feature in the comparison of models. Arabic text outcome is an important marker of how well AI models can interpret different cultural identities and social contexts. This is what the majority of AI models generated images lacked. ChatGPT 5 OpenAI is the only model that had a correct inclusion of Arabic textual elements in almost all generated images. ChatGPT 5 OpenAI had one instance of gibberish in the generated images. This can be explained due to the fact that textual prompts generated by ChatGPT 5 OpenAI lack Arabic texts in this instance. On the other hand, the generated images provided by Google Imagen 4, Google Gemini 2.5 Flash (Nano Banana), Meta AI, Grok xAI, Qwen3-Max AI, Leonardo AI (Phoenix 1.0), Leonardo AI (Lucid Origin), and Recraft V3 did not include Arabic texts (see Figure 5), or produced meaningless writing (gibberish) (see Figure 4).

5.1.6 Messages/ intertextuality

Intertextuality is a feature found only in AI generated texts but not images. Intertextuality within advertisements refers to shaping the text's meaning and creating interrelations of people and objects that goes beyond the advertisements. Concepts such as “family unity”, “happiness” (ChatGPT 5 OpenAI) and “togetherness” (Recraft V3) are emphasized in the text. This is relevant to the idea of intertextuality, in which texts rely and connect to other texts (Bazerman, 2004; Kristeva, 1980). Hence, AI generated texts can be described as “intertextual”. ChatGPT 5 Open AI, Qwen3-Max AI and Recraft V3 had the best description of intertextuality, in which inferences and interpretation of the type of relationship between people, objects, the brand, logo and the product are given in all texts. Leonardo AI had this feature in some texts but not all generated ones. Google (Gemini), Meta AI and Grok xAI lacked this level of in-depth textual analysis.

Some generated texts identify the photo as an advertisement, however, the majority did not and only described the logo per se. Recognizing the input image as an advertisement is an important feature that shows the AI model has realized the purpose of the image – selling a specific branded product. Comparisons between the different models showed that only ChatGPT 5 OpenAI and Qwen3-Max AI in all instances, and Meta AI in some instances stated within the text that the image was an advertisement. All other models [Google (Gemini), Grok xAI, Leonardo AI and Recraft V3] did not identify the image as an advertisement.

5.1.7 Logos

Realization and identification of logos includes description of the logo, as well as giving its correct location. The produced prompts varied with regard to understanding the logo content. Meta AI, ChatGPT 5, Google (Gemini), Grok xAI and Recraft V3 did not provide any details or descriptions and only provided the name of the brands. Textual prompts by Qwen3-Max AI and Leonardo AI explained the logo accurately, i.e., the shape of the logo, the place, and the order of the two scripts of the two languages provided (Arabic and English). None of the generated texts were able to identify nor recognize the logo as a specifically Saudi one.

In the text to image data, different logo outcomes were found in the images produced. Some models had a correct logo even when there was no detailed description of the logos in the textual prompts (ChatGPT 5 OpenAI, Grok xAI, and Qwen3-Max AI). Most models did not include a logo, even when the name or description of the logo was mentioned [Google Imagen 4, Google Gemini 2.5 Flash (Nano Banana), Meta AI, and Recraft V3]. However, in many cases the generated logo was incorrect in either color, shape, Arabic text or position of the logo [Google Imagen 4, Google Gemini 2.5 Flash (Nano Banana), Meta AI, Grok xAI, Leonardo AI (Phoenix 1.0), Leonardo AI (Lucid Origin), and Recraft V3].

The comparison of logos in the two data sets provides interesting findings. A wide range of logos were produced, ranging from the most accurate to those that were totally different in shape and color. Not all AI models were able to detect the brand and produce an accurate logo. Some AI models attempted to keep the text in the logo and did not include the symbol (see Figure 4). Others used different symbols, reflecting what the textual input described. Screen layering was also problematic for some AI models; for example, Recraft V3 failed to recognize logos and represented them as part of the background location (see Figure 3).

5.1.8 Modality

Modality is also a recurrent theme found in AI generated text and is noticeable when comparing the original and final product mediated by the textual prompt. Modality refers to the degree of realism, credibility, and truth value that an image communicates, seen in the description of color saturation, lighting, sharpness and the use of filters. Description of general color, light and photo atmosphere is given in some prompts, i.e., “blue gradient overlay at the bottom”. The study data shows that if aspects of modality are explained in the text, the final images reflect that, and vice versa. Only a few examples represent modality, such as the image in Figure 4:

Both images in Figure 4 include a “blue gradient overlay at the bottom”, which other AI models fail to indicate in the textual prompt. However, the blue gradient color is not shown in the AI generated image in Figure 5, nor the white one in the generated image in Figure 3 due to lack of textual prompt describing modality in Meta AI and Recraft V3. Overall, all models showed great accuracy in depicting real life to some extent, with the exception of image hallucinations of color, people, and objects, which are representations of reality that mirrors the original image. None of the generated images show other levels of modality such as cartoons, drawings, comics, animation or black and white photos.

ChatGPT 5 OpenAI, Grok xAI, Qwen3-Max AI and Recraft V3 provided a good description of modality in AI generated texts, whereas Meta AI and Leonardo AI did not include any description of modality in their textual prompts. With regard to the AI generated images, most of the models lacked an understanding of modality elements, such as lighting, filters, and gradient colors (Google Imagen 4, Google Gemini 2.5 Flash (Nano Banana), Google Gemini 2.5 Flash (Nano Banana), Meta AI, Grok xAI, Qwen3-Max AI, Leonardo AI (Phoenix 1.0), Leonardo AI (Lucid Origin), and Recraft V3). This can be attributed to the fact that neither textual prompts include sufficient details nor could generated AI images interpret the modality elements. These models focused more on semantic content over aesthetic attributes, and the implicit and subjective nature of such visual cues in textual prompts added difficulty. ChatGPT 5 OpenAI was the only model that had the right modality elements in the generated images. This shows ChatGPT 5 OpenAI has the ability both to interpret modality features and produce them.

5.1.9 Color

The ability to describe the color in the generated textual description of the original advertisement is an important feature that allows the AI tool to provide a multimodal understanding of the visual image. Google (Gemini), Meta AI, Grok xAI, Qwen3-Max AI and Leonardo AI all demonstrated the ability to provide a good description and accurate understanding of the different colors in the original advertisements. Chat GPT 5 Open AI provided such descriptions of colors in the images in the majority of cases. However, in one image there was insufficient description of people's clothing. Recraft V3 only provided color description in two advertisements.

Further examination of AI images shows that color representation is found on multiple levels: color of atmosphere, objects, clothes, skin tone, hair, logo…etc. The overall analysis shows that AI models are successful in transferring accurate colors. This is clear in depicting white walls, red Ghutras, blue shirts and orange bottles. However, in some cases color is misrepresented; for example, in Figure 3, the original advertisement shows a little girl wearing a black and white dress whereas the image generated by Recraft V3 shows the girl wearing a red checkered dress.

The comparison between different AI models shows the degree of accuracy in interpreting color representation, saturation and contrast based on AI generated texts. ChatGPT 5 OpenAI, Google Imagen 4, Google Gemini 2.5 Flash (Nano Banana), Meta AI, Grok xAI, Qwen3-Max AI, Leonardo AI (Phoenix 1.0) and Leonardo AI (Lucid Origin) all produced correct colors in the images. Only with Recraft V3 was the color in the image not aligned with textual prompts.

5.1.10 Hallucination

This study identifies three types of hallucination: textual, perspective (cultural and visual) and spatial. An interesting finding is what we term “text hallucination”, derived from Hurley's (2024) “image hallucination”, which falls under “technical errors”. Text hallucination is defined in this study as inaccurate generated textual description of objects or actions from images uploaded to AI models. This can be linked to “Intrinsic hallucinations” in which the generated content is contradicting the original content (Ji et al., 2023). In addition, our text hallucination can be seen as what Hurley (2024) terms as “perceived errors” in which Hallucination in text in the collected data can be seen in incorrect descriptions of objects (as in Figure 2 above). In this example, the generated text by Recraft V3 erroneously described a man “holding a hard yellow hat”, which does not exist in the original advertisement. This resulted in an AI generated photo with a man wearing a yellow hard hat that is different from the original.

In the visual AI generated images, cultural perspective hallucination is found in generated cultural clothes. For example, in the current data, a woman is wearing a red and white head piece (shemagh) that is usually worn by men resulting in a distortion of reality. This is an example of what Hurley called “hallucinations as a distorted view of reality” in her hallucination taxonomy (2024, p. 1,249). With regard to spatial hallucination, objects are usually misplaced or mispositioned. Figure 3 (below) shows spatial hallucination in the AI generated image by Recraft V3. This can be seen in the way the little girl is positioned in the supermarket's trolly. In addition, biological features such as face are not in normal proportion. Dress and colors are not consistent with the text nor high quality images. Also, this example shows visual perspective hallucination in which faulty layering of screens in which an Almarai sign is shown in the supermarket instead of a logo added to a visual advertisement. This type of hallucination, spatial and visual perspective, corresponds to Ji et al.'s (2023) intrinsic hallucination in which the generated output contradicts the source information.

Hallucinations can be found in the majority of AI models, even when there is no misinformation in the textual prompts. Google Gemini 2.5 Flash (Nano Banana), Meta AI, Grok xAI, Leonardo AI (Phoenix 1.0), Leonardo AI (Lucid Origin) and Recraft V3 all produced hallucinations. This shows how AI models, even with the right textual prompts, are sometimes unable to produce an image without adding, altering, or misrepresenting visual content. Only ChatGPT 5 OpenAI and Google Imagen produced images without any hallucinations.

Some examples from both generated text and generated images might first look like hallucinations, yet are not. Another interesting finding with regard to generated text is linguistic ambiguity associated with meaning. This will later affect AI generated images. For example, a text generated from Google (Gemini), “All four men are wearing traditional white thobes and red and white ghutras”, resulted in an image that is different from the original. The ambiguity is created by the word “and” which links the four men with different colored clothes. Some men were shown in the generated image wearing either red or white head piece whereas the original shows all men wearing a head piece that has two colors, white and red. Moreover, there are a few examples in which the image shows incorrect actions that are not the result of wrong textual description but rather due to language ambiguity. For example, in an image generated by Recraft V3 a man raises his hand as if he is waving or greeting someone; the text includes the word “welcoming” used as an adjective, whereas the image interpreted it as a verb.

Overall, the identified themes found in AI generated text and images indicate strengths as well as weaknesses in relation to semiotic and cultural representation mediated by technology. A general comparison of all examined AI models is now presented.

5.2 Comparison of AI models in generating images and texts

5.2.1 Image to text: comparison

This study compared different AI models based on the text and image generated. This involved a comparison between the original advertisements and the output generated advertisements. A comparison is made in which all AI models are brought together in a (see Supplementary Appendix Table S2) to evaluate the textually generated models. This comparison is based on the following common themes: color, intertextuality, logo description, translation (Arabic text), cultural description, location of the image, action of people, and the model's ability to recognize the advertisement as an advertisement (see Figure 6):

Figure 6 shows the score of semiotic features across AI models. The figure shows that location and people are the highest scoring features followed by culture and color. The lowest scoring feature is “Arabic text” indicating that AI models do not use the Arabic script in their produced textual interpretations.

Figure 7 shows a comparison of AI models with regard to their representation of themes analyzed that are identified as semiotic resources:

The above figure shows the total score for each AI model in relation to textual prompt features. When comparing textual prompts from each of the AI models, similarities and differences emerge. Although ChatGPT 5 is by far the most popular AI chatbot (Fischer, 2025), was lacking in some textual interpretations of texts. In comparison, Qwen3-Max AI gave an accurate interpretation of all features of the input image, including recognizing the image as an advertisement. This could be due to the number of words used. The average range of AI model texts is between 63 and 187 words, whilst Qwen3-Max AI textual prompts were between 159 and 303. Furthermore, one of Qwen3-Max AI standout features is its ability to understand and describe the Arabic texts and the logo provided in advertisements. Other models provided acceptable interpretations of color, people, location and action. However, the majority lacked an understanding of the Arabic texts and failed to identify the image as an advertisement.

5.2.2 Text to image: comparison

The second stage of analysis relied on the production of AI images from the AI generated textual prompts. A comparison is made in which all AI models are brought together in a (see Supplementary Appendix Table S3) to evaluate the generated images across AI models. Generated images were compared with regard to their capability to produce: color, logo (right logo, wrong logo, or missing logo), Arabic text, culture, people, modality, hallucination, location and action (see Figure 8):

The above Figure shows points of strength and weakness with respect to AI models' representation of semiotic features in generated images. Color, people and action are represented in most AI generated models accurately whereas logos, Arabic script, culture and location are not reflected. This realization of how AI works in real life situations has a number of implications for many sectors.

As shown in Figure 9, the text-to-image generated features of the selected AI models are compared. The figure shows the total score for image-generated features across AI models. From the overall comparison, the AI model with the best image generation is ChatGPT 5 OpenAI. This is surprising, as ChatGPT 5 OpenAI did not produce the best textual prompts. One explanation is that ChatGPT 5 OpenAI stores the image in its database, even though the image produced was generated from another account. Textual prompts generated from ChatGPT 5 OpenAI lacked detail: for example, it does not include the textual details of the brand logo, but in the image the Almarai logo was correct. Another explanation is that ChatGPT 5 OpenAI utilizes additional input from the internet. Despite the fact that Qwen3-Max AI produced the best textual prompts, it lacked the ability to produce accurate images. This figure shows to what extent different AI models vary in terms of accuracy in representing semiotic features.

5.2.3 Semiotic features in AI texts and images: a holistic view

When considering different semiotic features found across the models, it can be said that there are variations between generated textual prompts (see Figure 6) and generated AI images (see Figure 8) in terms of representation of semiotic resources. While color functioned as a salient semiotic feature in both the textual prompts and the AI-generated images, it demonstrated greater semiotic coherence within the visual outputs, where five AI models were correct in textual prompts as opposed to the eight AI models that were correctly displayed in the generated images.

The majority of AI models lacked logo descriptions in the textual prompts. However, in the generated AI images none of the AI models included the logo correctly or missed it out completely. Two AI models had it correct in the textual prompts but there were none correct in the AI models generated images. Arabic text as a feature was featured correctly in only one AI model in the textual prompts. However, it was not present at all within the image generated across the different AI models. Cultural representation in textual prompts was described correctly in five out of seven AI model generated textual prompts but it was only found in two of the AI generated images.

Six of the AI models had an accurate description of people textual prompts, while only four AI models had an acceptable interpretation in the generated AI images. Modality as a semiotic feature was found in four of the textual prompts, while in the generated images only one AI model produced it appropriately. Even though location was described correctly in six of the AI textual prompts, this was not the case for the locations in the generated AI images, where only one was correct. Action was produced effectively in textual prompts by three AI models, and by four AI models generated images.

It can be concluded from this overall comparison that the AI models' textual prompts had a valid interpretation of the different semiotic features – color, culture, people, and location. However, a true interpretation of Arabic texts and logo descriptions was lacking across the different AI models' textual prompts. Images generated by the different AI models had the highest level of satisfaction for color, followed by people and action. The generated images lacked accurate location and modality. Finally, the images generated by the different AI models did not have any satisfactory representation of Arabic text and logo.

6 Discussion and conclusion

This study examines the interrelation between text input and visual, and vice versa, with an attempt to understand how inter-semiotic relations manifest in AI algorithms to represent linguistic, social, and cultural aspects. The study also aims at answering the following questions: “What cultural elements in advertisements are represented when generated by AI models in multimodal contexts?” and “What are the strengths and weaknesses of different AI tools regarding multimodal representation when generating advertisements?”

In relation to the first research question, the study shows that AI attempts to capture most semiotic resources found within given prompts when remediating input: texts, symbols, color, objects, and people are described in the AI generated texts and portrayed in AI generated images. This is in line with Alikhani et al. (2023), who refer to the cooperation of modes in AI as “image-text coherence”. Yet, this is true on a surface understanding of semiotic features within a multimodal examination. In addition, analysis shows that dress and race are among the most evident social elements that are transmitted. This finding is in line with Assis and Moura (2025) who conducted a study of images created by AI and discuss how race and other cultural aspects are integrated in AI.

The data shows that text, as well as images, operate on multiple geographical circles to represent different cultures: there is the wider Muslim circle followed by Arab, then the Gulf Area, and finally the Saudi one. Although the original image is a Saudi advertisement with Saudi subjects, the final produced image in many cases does not present Saudi traditional clothing. Intermediated by textual input, variation occurs due to semantics surrounding word choice. For example, Middle Eastern dress yields Emarati traditional thobe. Different names used in the texts such as “keffiyya”, “kandoorah”, “shemgh” belong to different countries in the Arab world. This is similar to the concerns raised by Boutier (2025) who discusses the idea of AI exclusion in fashion that threatens cultural differences by producing output that is colored with the concepts of “pluralism”.

There were a number of setbacks when it comes to transfer of some semiotic aspects which affect the quality of a produced advertisement. This can be seen in problems with Arabic language inclusion, screen layering and AI hallucinations. As explained in the analysis, many AI models failed to recognize or include Arabic text in both AI generated texts and AI generated images. Instead, gibberish was used to fill in the linguistic void. Moreover, models failed to realize screens that belong to different perspectives, i.e., advertisement screen layered on top of the supermarket space. Hence, the technology faces challenges to realize multiple “sites of engagements”, spaces and semiotic “landscapes” (Jaworski and Thurlow, 2010; Jones, 2005). AI hallucinations can be seen on many levels: on the level of language, action, culture and physics. This is in line with Hurley's (2024) explanation of image hallucination, which highlights alteration of reality. The current study extends the idea of image hallucination to textual, as well as perspective and spatial hallucination.

With regard to the second research question, a number of key semiotic features are identified: color, logo, Arabic text, culture, people, modality, hallucination, location, and action. These features vary in their emergence within the AI models examined. The study identified a number of the most repeated features that appear in AI, as well as the lowkey features. The study also compared different AI models with regard to their inclusion/exclusion of key semiotic features and yields the following results: for example, color, people and location are among the recurrent semiotic features whereas Arabic text and logos are marginalized.

This exploratory study evaluated modern AI models when remediating advertisements in order to assess to what extent cultural and semiotic resources are adapted to produce successful advertisements. This study has an innovative methodological design in terms of data procedure (i.e., image to AI text, AI text to AI image). The findings contribute to theories of multimodality, AI and cultural studies. The study illustrates the extent to which AI models realize and reproduce cultural and semiotic resources (Kress and Van Leeuwen, 2006). The study has practical implications for marketing and advertising sectors which seek to brand for a specific culture. Implication for AI programmers include paying more attention to various cultures, script and languages. However, this study is limited in terms of the size of data that is extracted via continuously evolving AI models. More research is needed to address challenges facing brands that plan to integrate AI, as well as preserve specific culture and ideologies.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Seven AI models including ChatGPT, Gemini, Meta AI, Grok, Qwen, Leonardo AI, and Recraft V3, were used to create images for this article. More information is explained in the methodology and can be found in Supplementary Appendix Table S1.

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.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcomm.2026.1959385/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Sabine Tan, Singapore University of Social Sciences, Singapore

Reviewed by: Jun Liu, Cheongju University, Republic of Korea

Valerio Perna, Epoka University, Albania