MosAIC: Multi-Agent AI Framework for Cross-Cultural Image Captioning

Introduction

Cross-cultural image captioning in AI is a crucial area of research aimed at improving the cultural sensitivity and diversity of image captions. This concept addresses the need for AI models to generate captions that are not only contextually accurate but also reflect the cultural nuances of the images they describe. As AI technologies become more widespread, it is essential that these systems are trained on data that represent the diverse cultural contexts in which images are viewed. Without such considerations, AI models might perpetuate stereotypes or fail to recognize culturally significant elements of an image, ultimately leading to biased outputs.

The importance of cross-cultural image captioning lies in its potential to enhance the accessibility and inclusivity of AI-generated content. In particular, diverse captioning models can better serve users from different cultural backgrounds by providing more accurate, relevant, and respectful descriptions of images. For example, an image depicting a traditional cultural practice in one country could be misunderstood or misrepresented by a model not sensitive to that culture’s values and practices. As a result, developing AI models capable of recognizing and incorporating cultural diversity can prevent the reinforcement of harmful stereotypes and promote fairer and more inclusive representation.

Additionally, such models contribute to the ethical development of AI by ensuring that AI systems do not perpetuate biases. AI models trained on diverse data sources are better equipped to produce balanced outputs that represent a broad range of cultural perspectives. This is particularly important in fields like content creation, education, and social media, where image captions influence how information is perceived and shared globally. Moreover, AI systems with a deep understanding of cultural diversity can help break down cultural barriers, fostering better cross-cultural communication and collaboration.

Recent advancements, such as the development of frameworks like the Culturally-aware Image Captioning (CIC) model, highlight the ongoing efforts to address these challenges. The CIC model aims to incorporate cultural visual elements alongside traditional caption prompts, ensuring that the generated captions are not only contextually accurate but also culturally relevant. This approach has shown promising results in human evaluations, underscoring the potential for AI to create more inclusive and representative captions for diverse audiences.

By improving cultural awareness in image captioning, AI systems can move towards more equitable and representative content generation, benefiting users from various cultural backgrounds and promoting a more inclusive digital world.

MosAIC (Multi-Agent Image Captioning) is an innovative framework designed to enhance the cultural awareness of AI-generated image captions. Its primary objective is to address the growing need for more accurate and culturally sensitive image captioning in AI systems. Traditional AI models, especially vision-language models (VLMs), often fall short in recognizing and interpreting cultural nuances within images. MosAIC seeks to rectify this by leveraging a multi-agent approach that involves collaboration between various agents, each specializing in different cultural domains and contexts.

The challenge of ensuring that AI-generated captions reflect the deep cultural context behind images is significant, as current AI systems may describe images accurately in terms of visual elements, but they often fail to grasp the cultural significance embedded within those elements. For example, an image of a cultural artifact might be captioned correctly in terms of its appearance but without any recognition of its historical or symbolic meaning. This lack of cultural awareness is problematic, especially when AI systems are deployed in sensitive contexts where cultural respect and understanding are paramount.

MosAIC addresses this gap by creating a multi-agent system where each agent is fine-tuned to recognize and describe specific cultural aspects. These agents are not only trained on diverse image datasets but also on the cultural contexts that make those images meaningful. By using this approach, MosAIC aims to produce image captions that are both visually accurate and culturally sensitive. For example, when captioning an image of a Native American artifact, the model would not only describe its appearance but also provide context about its cultural significance, such as its role in spiritual practices or its symbolic meaning in the community.

Additionally, MosAIC integrates a new evaluation metric, the Cultural Awareness Score (CAS), which quantifies how well the generated caption reflects the cultural dimensions of the image. This score helps assess the effectiveness of the AI in capturing culturally significant elements, ensuring that the caption isn't merely a generic description but one that resonates with the cultural importance of the image.

Through its innovative use of multi-agent collaboration and the introduction of the CAS, MosAIC represents a significant step forward in making AI more attuned to the rich diversity of human cultures, ensuring that AI technologies are not only functional but also culturally aware and respectful.

What is MosAIC?

The MosAIC framework for cross-cultural image captioning utilizes a multi-agent model consisting of three distinct agents: the Social, Moderator, and Summarizer agents. These agents work together in a collaborative process to generate accurate and contextually relevant image captions, which is particularly valuable in diverse cultural contexts.

1. Social Agent: This agent is responsible for interpreting the social and cultural context of the image. It understands how different societies perceive the content of the image and ensures that the captions produced are sensitive to these differences. By factoring in social cues, relationships, and cultural nuances, the Social agent tailors the description of the image to resonate with specific target audiences, respecting varying cultural perspectives. For example, an image of a public gathering may be described differently depending on the social norms and values of the culture interpreting it.

2. Moderator Agent: The Moderator acts as a central overseer, ensuring that the captions generated by the other agents are consistent with the overall goals of the framework. It resolves any conflicts that arise between the Social and Summarizer agents, and helps maintain coherence and relevance in the captioning process. This agent helps to mitigate biases and ensures that the captions remain inclusive and accurate across different cultural representations.

3. Summarizer Agent: This agent focuses on distilling the essential elements of the image into concise and clear captions. It takes into account the features identified by the other agents and synthesizes them into a narrative that is both informative and culturally appropriate. This is where the actual caption generation takes place, drawing from the data and contextual insights provided by the Social and Moderator agents.

Together, these agents form an integrated system that enhances the image captioning process, ensuring that captions are not only contextually accurate but also sensitive to cultural diversity. This multi-agent model provides a dynamic and flexible approach to captioning, making it adaptable to various social and cultural contexts. Additionally, such a framework can significantly improve the accuracy of automated systems for content moderation, cross-cultural communication, and social media platforms, where understanding cultural nuances is critical​.

To achieve a more nuanced and culturally diverse description of an image through multi-agent systems, various agents can be designed to emulate collaboration in a culturally diverse group. Each agent in this system is modeled to reflect a distinct cultural perspective, leveraging its unique experiences and knowledge to create richer, more varied interpretations of the same visual data. For instance, in a system developed for cultural image captioning, agents can represent different cultural personas, such as those from China, India, and Romania, each bringing their own historical, social, and aesthetic influences to the image captioning task.

The interaction between these agents allows for a more comprehensive and culturally sensitive description of the image. Instead of a single agent relying solely on generalized training data, these multiple agents can cross-check, expand, and balance their interpretations, ensuring that cultural nuances are considered. For example, an image of a traditional festival could be described differently depending on whether the agent's cultural background is based in Eastern or Western traditions, enriching the caption with context-specific details. This form of collaboration mimics real-world interactions among individuals from different backgrounds, each contributing their own insights to form a more holistic understanding.

Recent work in this field has shown that multi-agent systems outperform single-agent models in tasks that require cultural sensitivity. These systems can be particularly effective in creating more inclusive and diverse content in applications like image captioning, where understanding the cultural context is key. The MosAIC framework, for example, demonstrates how multi-agent systems can work together to improve cross-cultural image captioning. This system uses distinct cultural personas in its agents, leading to more culturally enriched captions across various datasets.

By emulating the collaborative process within culturally diverse groups, multi-agent models can generate more representative and sensitive content, enhancing the inclusivity of AI-generated descriptions. This approach moves beyond Western-centric perspectives, addressing the need for more globalized and context-aware AI systems.

How the Multi-Agent Model Works

In the MosAIC framework, the roles of the different agents are pivotal in generating a nuanced and culturally diverse caption for images. Here's a deeper dive into these agents and how they contribute to the process:

1. Moderator: The moderator is central to the operation of MosAIC, acting as the orchestrator of the process. It sets the context for the image captioning task and generates a series of questions or prompts that guide the social agents. The moderator's role ensures that all agents work within a structured framework, adhering to a shared understanding of the task and cultural perspectives involved. This controlled environment allows for meaningful interactions between agents while maintaining the focus of the output.

2. Social Agents: These agents are responsible for bringing various cultural perspectives into the captioning process. Each social agent represents a different cultural background, such as China, India, or Romania, and is tasked with describing the image from the lens of its specific cultural context. For instance, an agent representing Chinese culture might focus on symbolic elements, values, or historical references prevalent in Chinese society, while an agent from India might incorporate regional aesthetics or social narratives into their description. These agents ensure that the caption produced is not only accurate but also culturally resonant, adding layers of meaning that would be missed in a more generic approach.

3. Summarizer: After the social agents have each provided their descriptions, the summarizer agent takes over. Its role is to integrate the diverse insights from the social agents and synthesize them into a coherent and comprehensive final caption. The summarizer needs to ensure that the varied cultural nuances are balanced effectively, avoiding biases or contradictions. This final output is the culmination of the inputs from the moderator and the social agents, providing a rich and multifaceted description that reflects a global perspective. The summarizer's task is to blend these inputs into a seamless narrative that is both informative and inclusive.

This multi-agent approach to image captioning mirrors the growing trend in AI development towards more interactive and modular systems. By incorporating diverse cultural viewpoints, MosAIC aims to overcome the limitations of conventional image captioning systems, which often fail to account for the complexities of cultural context. This method not only makes image captions more inclusive but also opens new possibilities for cross-cultural AI applications, such as in global marketing, education, and content creation​.

These roles, when combined, create a system that is not only efficient in its technical execution but also sensitive to the cultural nuances that shape human understanding and communication. This makes MosAIC a powerful tool for generating culturally aware and contextually rich image captions in an increasingly globalized world.

The iterative learning process in multi-agent systems, particularly in tasks like image captioning, involves agents that collaborate over multiple rounds to enhance the quality and relevance of their outputs. Each agent works to improve its predictions by learning from both its own experiences and those of its peers.

In the context of image captioning, such systems often employ multi-agent reinforcement learning (MARL) frameworks. Here, each agent is responsible for producing a word or phrase in the description of an image. Through iterative rounds, agents share insights and refine their predictions based on the feedback from their counterparts. The agents' actions are evaluated based on a shared reward, which encourages them to generate more accurate and contextually rich captions.

The process begins with each agent observing the visual context from the image and making an initial prediction about the caption. The agents are not isolated; they interact with each other through communication layers, exchanging feedback and refining their individual models. This learning loop allows them to correct missteps and learn how their individual contributions impact the overall caption quality. By performing multiple iterations, agents gradually optimize their strategies and converge toward generating more precise and coherent captions that better reflect the image's content.

The use of counterfactual learning plays a key role in addressing challenges like decoding inconsistencies. In some multi-agent systems, each agent is provided with a counterfactual baseline, which represents what would happen if different actions had been taken. This helps agents better understand their contribution to the team reward, enabling more targeted improvements in future rounds. By comparing the actual reward to the counterfactual baseline, agents learn which actions (words or phrases) are most effective and which should be avoided.

Over successive iterations, the agents' collective learning process enables them to produce increasingly nuanced and accurate image captions, enhancing the overall performance of the system. This iterative process also allows the system to adapt to a variety of different images, learning the unique characteristics of each and refining its ability to describe them in contextually appropriate ways.

Cultural Benchmark and Evaluation Metrics

The cultural benchmark dataset used in MosAIC, a framework for culturally-aware image captioning, includes captions that represent diverse cultures such as China, India, and Romania. This dataset draws from multiple cultural categories, including architecture, clothing, food and drink, dance and music, and religion, allowing for a rich and diverse understanding of cultural representation in visual content. The dataset plays a key role in ensuring that the generated captions are not only visually accurate but also culturally sensitive and appropriate for different cultural contexts.

To generate these culturally-aware captions, the framework incorporates a multi-step process that starts by formulating specific cultural questions aimed at extracting visual elements tied to the different cultural categories. For example, questions might inquire about the architectural style or the clothing worn by people in an image, enabling the system to extract culturally relevant features from the visuals. This information is then passed to a large language model (LLM) like ChatGPT, which refines and generates captions that reflect the cultural nuances of the image content, ensuring they align with the visual elements extracted through a Visual Question Answering (VQA) model.

The dataset's significance lies in its ability to handle images from diverse regions such as East Asia, South Asia, and Eastern Europe, representing a variety of global cultural norms. The goal is to create captions that are contextually accurate and sensitive to the culture portrayed, while avoiding a one-size-fits-all approach that could lead to cultural misinterpretations. As such, the MosAIC framework has set a benchmark for improving the cultural awareness of image captioning technologies, making it a valuable tool for applications in areas like AI-driven content creation, education, and accessibility.

The introduction of new metrics for evaluating the quality of generated image captions, with a particular focus on alignment, completeness, and cultural relevance, has sparked significant advancements in AI-based captioning systems. These new metrics aim to ensure that captions are not only technically accurate but also culturally sensitive and comprehensive, reflecting the diverse contexts in which images may be interpreted.

1. Alignment refers to the extent to which the generated caption accurately matches the visual content of the image. This metric addresses the issue of hallucination, where an AI might generate captions that include elements not visible in the image, a problem that occurs especially when dealing with complex cultural categories. For example, when an image contains multiple cultural cues, it is crucial that the caption aligns with specific visual elements tied to those cues, such as architecture, clothing, or food. Misalignment can lead to captions that are culturally insensitive or irrelevant​.

   
   

2. Completeness evaluates whether the caption encapsulates all relevant information from the image. In the context of culturally-aware captions, this means that the model should integrate all cultural elements represented in the image without omitting significant details. For instance, a caption should reflect not only the visible people and objects but also cultural aspects such as regional architecture or traditional attire, which are vital for a holistic interpretation of the image. The use of Visual Question Answering (VQA) frameworks, which extract detailed cultural context from images, is integral in ensuring that the generated captions cover all necessary aspects​.

   
   

3. Cultural Relevance is perhaps the most innovative metric introduced, focusing on ensuring that the caption respects and accurately represents the cultural context of the image. This involves using a specialized set of questions to extract culturally-specific elements from the image. For example, in an image depicting a traditional Asian festival, the caption must recognize and correctly reference cultural practices such as dance or food, which can vary greatly across different cultural backgrounds. This metric aims to prevent the generation of captions that are either too generalized or culturally biased​.

   
   

These new evaluation criteria align with the broader goal of developing AI models that can generate captions that are not only accurate but also culturally nuanced. By addressing these three core metrics—alignment, completeness, and cultural relevance—AI systems can generate captions that are more reflective of the true essence of the images they describe, especially in the context of diverse global cultures.

The Role of AI in Cross-Cultural Understanding

The ethical and practical implications of using AI for cross-cultural tasks are vast and critical, especially as AI systems become increasingly global in their application. One of the key concerns is ensuring fairness in AI models, which is central to creating AI that can work across diverse cultural contexts without perpetuating biases. Traditional AI models often suffer from biases that reflect the cultural and societal influences of their creators, which can lead to discriminatory or unfair outcomes, particularly in cross-cultural settings.

One significant challenge is the integration of ethics into AI systems. As AI continues to be deployed across various industries worldwide, the question arises: whose ethical framework should guide AI decision-making? This issue is especially pertinent when considering the global diversity of cultural practices and values. A practice that may be considered ethical in one culture can be seen as problematic or even unethical in another. For example, the practice of gift-giving in business is widely accepted in Chinese culture but may raise red flags in other parts of the world, where it could be viewed as bribery. This cultural divide in ethical norms becomes even more problematic when AI systems are tasked with interpreting or responding to such practices.

AI's inherent biases, often shaped by the data it is trained on, can also reflect and reinforce harmful stereotypes. For instance, AI models trained on biased data may misinterpret visual data, leading to mislabeling or wrongful profiling of individuals based on race, gender, or other characteristics. A well-known example of this is AI-powered facial recognition systems that have been shown to be less accurate for people of color, particularly in high-stakes situations like law enforcement. Such biases are not only unethical but also potentially dangerous, as they can lead to unjust decisions being made by AI systems without human oversight.

Addressing these challenges requires a multi-faceted approach. First, it is essential to prioritize diversity in the development of AI models. AI systems should be designed and trained by interdisciplinary teams that bring a variety of cultural perspectives and ethical viewpoints. This approach would help ensure that the AI systems are better equipped to handle cross-cultural tasks in a fair and unbiased manner. Moreover, creating global standards for ethical AI is crucial. Institutions like the European Union are already taking steps in this direction by planning regulations that enforce ethical AI practices and hold companies accountable for any harms caused by their AI systems.

Furthermore, initiatives like MosAIC, which explicitly tackle biases within conventional AI models, are part of the solution. These systems actively work to mitigate the cultural biases inherent in traditional models by incorporating diverse data sources and ethical considerations into their training processes. MosAIC’s approach is a promising model for developing AI that is both fair and culturally aware.

Ultimately, creating AI systems that are ethical, culturally sensitive, and free of bias requires continuous reflection on the part of developers, businesses, and policymakers. It also requires active collaboration on a global scale to ensure that AI can be used responsibly across different cultures and contexts.

Applications of MosAIC

Culturally aware image captioning has immense potential across a range of practical applications, particularly in fields like international marketing, accessibility, and education. Here's how such applications can create significant value:

1. International Marketing: Marketing campaigns that utilize AI for image captioning can significantly benefit from understanding cultural nuances. Accurate, culturally sensitive captions can ensure that advertisements resonate with diverse audiences without inadvertently offending or misrepresenting them. For example, a campaign featuring a global product can automatically adjust its imagery and descriptions to reflect culturally appropriate messages for each market. This can prevent costly missteps and improve the brand’s image worldwide, fostering deeper connections with consumers.

2. Accessibility: In accessibility, particularly for people with visual impairments, AI-generated captions that take cultural context into account can provide richer, more meaningful descriptions. For instance, a visually impaired user in one country may need different explanations or references in an image than a user in another country, based on their unique cultural experiences. Incorporating culturally relevant captions enhances the inclusivity of technology, ensuring that all users, regardless of location, can understand the visual content in a way that feels personalized and accurate​.

3. Education: Culturally-aware captioning can transform education, especially in language learning and cross-cultural communication. AI-powered tools can generate descriptions for educational videos or images that not only describe the content accurately but also include cultural context. This is particularly useful in online education, where students from different backgrounds may need tailored learning materials. For example, a history lesson using imagery might generate captions that explain cultural references, symbols, and historical significance in a way that resonates with a diverse student body. Additionally, in virtual classrooms, students can gain insights into different cultures and perspectives, fostering empathy and cross-cultural competence​.

   
   

In all these fields, leveraging AI to generate culturally-aware captions ensures that content is accessible, accurate, and relevant across different regions and user groups. This approach also mitigates biases that may arise in traditional, less context-aware models, making it a critical tool for fostering inclusivity and understanding in global communication​.

Improving AI's cultural sensitivity can have a profound impact on various technologies, especially those designed to engage and assist people across different cultures and regions. At the core, AI’s ability to recognize, respect, and adapt to cultural differences is crucial for building trust and increasing adoption, particularly in sectors like healthcare, education, and everyday user interfaces. When AI systems are culturally sensitive, they become more effective at meeting the diverse needs of their users, enhancing both user satisfaction and the accessibility of services.

In the context of healthcare, for instance, culturally sensitive AI can prevent biases that might otherwise lead to misdiagnoses or unequal treatment. Research has shown that AI models trained on homogeneous data can produce errors when applied to underrepresented groups. For example, AI systems used in diagnosing skin conditions might perform poorly for individuals with darker skin tones, leading to a disparity in healthcare outcomes. Similarly, gender bias in medical AI models can skew results, disproportionately affecting women. By ensuring that AI systems are trained on diverse datasets that include varied ethnic, cultural, and socioeconomic backgrounds, we can reduce these disparities and offer more equitable solutions.

Beyond healthcare, AI's cultural sensitivity is also vital in creating more inclusive and user-friendly technology. User-centered design, which involves testing systems with people from diverse cultural backgrounds, ensures that interfaces are intuitive and effective across different cultural contexts. By allowing customization or personalization—where users can tailor interfaces to align with their cultural preferences—we can foster a sense of ownership and comfort. For instance, in AI-powered apps, users may be able to modify language settings or adjust content to better suit their cultural norms and values.

Another significant area is ethical engagement with users. Ethical AI development involves transparent algorithms and informed consent processes that respect cultural norms. This means considering how data is collected, processed, and used, with attention to how different cultures view data privacy and ownership. Providing multilingual support, culturally tailored content, and continuous community engagement ensures that AI technologies are both accessible and respectful of cultural diversity.

The broader implications of this cultural sensitivity extend to fostering global cooperation and enhancing the overall fairness of AI systems. By taking cultural differences into account, AI can serve as a bridge rather than a barrier, creating more inclusive and accessible technology. This approach not only benefits individuals from diverse backgrounds but also contributes to societal harmony, making AI a powerful tool for positive change.

Challenges and Future Directions

Implementing a culturally sensitive AI system like MosAIC poses several challenges, especially when it comes to achieving accurate cultural representation without perpetuating stereotypes. A primary concern in this process is the potential for AI-generated content to reinforce harmful biases or inaccuracies. For example, AI models can sometimes generate outputs that reflect cultural stereotypes—such as associating certain cultures with violence, as seen in some controversial AI-generated images like the South Sudan Barbie from Buzzfeed’s viral “Barbies of the World” blog.

Cultural bias is particularly concerning in AI systems because it can amplify existing societal prejudices, leading to negative consequences like discrimination in areas such as hiring, healthcare, or even law enforcement. Racial or cultural biases in algorithms can result in skewed predictions, such as misdiagnosing patients based on racial assumptions or misclassifying individuals in predictive policing systems.

Addressing these biases requires careful attention to the training data used for AI models. Inaccurate or insufficient data sets can exacerbate these issues, making it crucial to ensure that AI systems are trained on diverse, representative datasets. However, even with comprehensive datasets, cultural nuance can be hard to capture, and there’s a fine line between portraying a culture accurately and perpetuating oversimplified or offensive stereotypes.

Moreover, mitigating cultural biases involves not just improving datasets but also adopting practices like continuous monitoring and feedback loops where AI outputs are regularly reviewed and refined. This ongoing process helps to identify and correct cultural misrepresentations before they reach the end user.

The complexities involved in developing MosAIC or similar AI models highlight the need for sensitivity and inclusivity in AI design. It’s not just about avoiding stereotypes but also about creating a system that respects and authentically represents the diverse cultures it aims to portray.

In terms of improving evaluation techniques and enhancing model diversity in the AI space, several key developments are worth considering. First, as AI models evolve, current evaluation methods often fail to capture the nuances of diversity in model performance. Traditional benchmarks may not fully reflect how models handle diverse contexts, languages, or less-represented demographic groups. Future research should prioritize the development of more comprehensive evaluation frameworks that go beyond accuracy metrics to include fairness, adaptability, and inclusivity. For instance, assessing AI models across different cultural or linguistic settings could provide valuable insights into their generalizability.

Moreover, improving model diversity is crucial. Many current AI systems suffer from biases due to the lack of diverse training data. Expanding the datasets used for training, ensuring they include a broad range of cultural, geographical, and socio-economic contexts, could lead to models that are more equitable and capable of serving diverse populations. Alongside this, fostering collaborations between researchers from different backgrounds can bring in diverse perspectives that contribute to the design and implementation of AI systems that cater to a broader spectrum of needs.

Additionally, AI’s role in addressing diversity, equity, and inclusion (DEI) should not be understated. AI can help identify patterns of exclusion in existing systems, providing opportunities to rectify imbalances. However, AI should also be viewed as a tool that can help organizations develop better DEI strategies, not just as a potential source of bias. For instance, AI's ability to analyze large datasets can uncover hidden biases in hiring or customer service processes, facilitating more inclusive decision-making.

In conclusion, the future of AI hinges on not just refining technical performance but ensuring that these systems serve everyone fairly. Incorporating diversity into both model design and evaluation processes, as well as continuously testing models for inclusivity, will be crucial in shaping an equitable AI landscape. These advancements will be key to fostering trust and maximizing the positive impact of AI in diverse sectors.

Conclusion

The importance of the MOSAIC project lies in its ability to infuse AI models with cultural sensitivity, particularly in contexts where image captioning and other AI-generated outputs may be inherently biased or lack relevance to specific cultural nuances. The MOSAIC-1.5k dataset and the Cultural Awareness Score (CAS), developed by machine learning engineer Vinija Jain, highlight a critical step in ensuring that AI systems not only recognize universal patterns but also understand cultural specifics.

This project underscores how necessary it is for AI systems to be designed with cultural context in mind, especially when these models are used across diverse populations. For example, while traditional AI models might excel at understanding broad visual or linguistic patterns, they often fall short when it comes to interpreting cultural symbols, traditions, and societal norms. By leveraging datasets like MOSAIC-1.5k, AI models can be better equipped to respond to culturally-specific inputs in ways that are more accurate, meaningful, and respectful.

The MOSAIC initiative’s focus on Indian culture illustrates a larger movement towards culturally informed AI, where global perspectives can be integrated into technological design. This represents a significant shift from creating AI systems based purely on Western frameworks to developing models that can better serve the diverse, multicultural world in which they operate.

Such advancements are essential not only for the technical progress of AI but also for its ethical deployment, ensuring that AI technologies are inclusive and representative of all cultures. Thus, MOSAIC plays a pivotal role in fostering a future where AI is truly global, reflecting a wide spectrum of cultural identities and realities.