Abstract

This study examines how well distinct indigenous cultures are represented and how artificial intelligence views ethnicity. The researchers used a quantitative approach, analyzing various datasets to identify patterns AI uses to classify human facial structures. The results show a lack of diversity in representation, highlighting the need to incorporate new datasets and ethnicity-related data into AI systems. These findings have important implications for educators, parents, and policymakers, as they demonstrate the necessity of including members from all ethnic groups and communities in this rapidly growing platform.

Introduction

Artificial intelligence (AI) technologies have the potential to transform many industries, but there are still concerns that they may worsen existing inequalities and encourage racial prejudice. This literature review aims to provide a thorough analysis of research findings and differing perspectives on how AI systems can discriminate against ethnic minorities. Researchers and academics have extensively studied how AI systems may reinforce racial biases present in the training data. For instance, Safiya Umoja Noble's groundbreaking study, "Algorithms of Oppression," reveals how search engine algorithms perpetuate negative stereotypes by favoring some content over others, which in turn reinforces racial biases (Noble, 2018). According to research on facial recognition algorithms by Joy Buolamwini and Timnit Gebru, the algorithm classifies women and people with darker skin tones with more inaccuracies, which may be due to biases in the training set (Buolamwini & Gebru, 2018). Furthermore, research has demonstrated that models of natural language processing trained on biased text data can provide discriminating results, therefore solidifying racial disparities (Bender et al., 2021). Significant prejudice and discrimination have been found in AI systems, according to studies (Element AI and Global AI Talent Report 2019). This is especially true for platforms like Google's search results and Facebook's ad distribution (Simonite, T. 2018). “Facial recognition has been independently assessed and proven to be less accurate on certain populations—even if the algorithm does not explicitly include race as an attribute, the outcomes still favor one race over another”(Thompson 2020).

The Effect on Minorities of Race

For ethnic minorities, the biased results of AI systems have real-world consequences in a variety of fields. AI systems applied in healthcare to make medical decisions may cause differences in how patients from underserved communities are diagnosed and treated (Obermeyer et al., 2019). Furthermore, it has been argued that predictive policing algorithms unfairly target minority populations, which has resulted in heightened police and surveillance in these regions (Ensign et al., 2018). Furthermore, it has been discovered that AI-powered hiring algorithms reinforce prejudices against ethnic minorities, making it more difficult for them to find work prospects (Chouldechova, 2017). “

The Amazon resume scanning example is just one of many that show how the functional logics of a given technology echo the gender and racial dynamics of the industry that produced it Amazon’s Rekognition facial analysis service previously demonstrated gender and racial biases worse than those of comparable tools, biases that took the form of literally failing to “see” dark-skinned women while being most proficient at detecting light-skinned men.”(West, S.M., Whittaker, M., Crawford, K. (2019). “Relying primarily on survey-based research conducted in educational settings, pipeline studies seek to understand the factors that lead to gender-based discrimination in computer science, more precisely by interrogating what drives women and people of color away from the field, and implicitly, what might make them stay”(West, S.M., Whittaker, M. and Crawford, K. (2019). AI has a significant and wide-ranging influence on ethnic minorities, with ramifications in many different fields. Artificial intelligence (AI) algorithms utilized in healthcare, for instance, may lead to differences in the diagnosis and course of treatment for patients from underserved populations (Obermeyer et al., 2019). Studies have indicated that minority populations are disproportionately targeted by predictive police algorithms, which results in more cops and monitoring in these regions (Ensign et al., 2018). Furthermore, it has been discovered that AI-powered recruiting algorithms reinforce prejudices against ethnic minorities, making it more difficult for them to find work prospects (Chouldechova, 2017).Furthermore, discriminating results in different circumstances may be sustained by racial prejudices ingrained in AI systems. For example, it has been demonstrated that face recognition algorithms make more mistakes when applied to women and people with darker skin tones, which may be due to biases in the training set (Buolamwini & Gebru, 2018). Similarly, discriminating outputs from natural language processing algorithms trained on biased text input may reinforce racial disparities (Bender et al., 2021).

Racial minorities are directly impacted by these biased results, which exacerbate already-existing disparities and restrict their access to opportunities and resources. Therefore, it is essential to address the racial prejudices present in AI systems to guarantee justice, equity, and fairness for all societal members.

Requests for AI Ethics:

The creation and application of ethical AI systems have come under increasing pressure due to the discriminatory effects of AI technologies. Academics push for enhanced responsibility and openness in developing and applying AI algorithms and diversity in the teams creating these technologies (Crawford et al., 2019). Moreover, legal frameworks that address the ethical implications of AI and guarantee that these technologies respect justice and fairness are desperately needed (Eubanks, 2018). To guarantee that decision-making processes are comprehensible and traceable, there is an increasing need for accountability and transparency in AI systems (Jobin et al., 2019). Transparency makes developers responsible for the results of these systems and helps consumers understand how AI algorithms make judgments. To ensure that AI systems do not reinforce or worsen prejudices against marginalized groups, ethical AI development demands a dedication to justice and non-discrimination (Barocas & Selbst, 2016). It is imperative to incorporate fairness measures and approaches into AI development processes to address algorithmic bias and prejudice.

In AI ethics, maintaining data rights and privacy is crucial (Floridi et al., 2018). To protect people's private information from abuse or exploitation, AI systems should abide by privacy-preserving principles such as data reduction, anonymization, and user permission. Human control and autonomy should be given priority in ethical AI, ensuring that AI systems support human decision-making rather than supplant them (Bostrom & Yudkowsky, 2014). To avoid unforeseen repercussions or injury, AI systems should include human monitoring and intervention tools. This would indicate that there is a dire need to study the possible effects of AI deployment on diverse stakeholders and communities, through societal impact evaluations (Jobin et al., 2019). These evaluations have to take into account aspects related to ethics, society, economy, and culture to guide the appropriate development and use of AI. To solve cross-border ethical issues and guarantee compliance with international human rights norms, ethical AI necessitates international cooperation and governance structures (Allen et al., 2019). Multi-stakeholder projects, legal frameworks, and moral standards can aid the global development of ethical AI. It is crucial to incorporate ethical issues from the beginning when designing and developing AI systems (Floridi et al., 2018). Throughout the AI lifecycle, ethical design techniques like value-sensitive design and participatory design can aid in identifying and addressing ethical challenges. To address new ethical issues and gauge the long-term effects of AI systems, ethical AI necessitates ongoing observation and assessment (Jobin et al., 2019). Continuous assessments, audits, and feedback systems can assist in identifying and reducing moral hazards and unintended consequences.

Images and references are now created using artificial intelligence on a variety of platforms, including Kaggle, Flickr, etc. Since these platforms are so widely used, visuals can be produced for practically every statement. Although this has a lot of advantages, questions have been raised regarding how much picture is being generated and how insufficient the datasets are.

Given the rising significance of artificial intelligence across many platforms, it is critical to look at how different people from across the world are represented on this ever-expanding platform. This study examines multiple datasets and AI pattern usage to close this gap.

In AI ethics, maintaining data rights and privacy is crucial (Floridi et al., 2018). To protect people's private information from abuse or exploitation, AI systems should abide by privacy-preserving principles such as data reduction, anonymization, and user permission. Human control and autonomy should be given priority in ethical AI, ensuring that AI systems support human decision-making rather than supplant them (Bostrom & Yudkowsky, 2014).

To avoid unforeseen repercussions or injury, AI systems should include human monitoring and intervention tools. This would indicate that there is a dire need to study the possible effects of AI deployment on diverse stakeholders and communities, through societal impact evaluations (Jobin et al., 2019). These evaluations have to take into account aspects related to ethics, society, economy, and culture to guide the appropriate development and use of AI. To solve cross-border ethical issues and guarantee compliance with international human rights norms, ethical AI necessitates international cooperation and governance structures (Allen et al., 2019). Multi-stakeholder projects, legal frameworks, and moral standards can aid the global development of ethical AI. It is crucial to incorporate ethical issues from the beginning when designing and developing AI systems (Floridi et al., 2018). Throughout the AI lifecycle, ethical design techniques like value-sensitive design and participatory design can aid in identifying and addressing ethical challenges. To address new ethical issues and gauge the long-term effects of AI systems, ethical AI necessitates ongoing observation and assessment (Jobin et al., 2019). Continuous assessments, audits, and feedback systems can assist in identifying and reducing moral hazards and unintended consequences.

Leftward images: Actual Indigenous communities
Rightward images: AI-generated images of the same communities

Methodology:

1. Research Design:

1.1 Data Collection:

This study will have 100 pairs of pictures. Each pair has one AI-made image and one search result image. We'll select pictures that summarize the search well. We'll avoid problematic representations. We'll document every image's source, date, and metadata. This is to be open and allow others to repeat our study.

1.2 AI Model Selection:

A. The images that are selected are based on the following factors:
Covering different demographic regions to ensure enough data from different communities.
‍
B. Covering different indigenous communities from the same region to check for any demographic biases in the image generation.

2. Data Analysis:

2.1 Image Comparison:

Develop a systematic process for comparing each image pair based on the following factors:

Accuracy

This criterion assesses if photos accurately depict the physical and cultural traits of studied indigenous populations. It compares features in the images with known characteristics of the tribes for qualitative accuracy evaluation.

Cultural Representation

This criterion assesses whether traditional attire, symbols, artifacts, and architectural designs that are symbolic of indigenous cultures are present and are depicted. To find this out, each image will be qualitatively assessed.

Facial Expressions

Facial expressions convey emotions and cultural context. Evaluation of facial expressions in images considers their emotional authenticity and cultural relevance, assessed qualitatively.

Clothing

In indigenous societies, traditional attire frequently represents cultural identity and ancestry. The existence and correctness of traditional clothing will be evaluated qualitatively by contrasting it with established depictions of indigenous wear and taking into account elements like design, color, and decorations.

Demographic Factors

The surroundings of the image offer context and understanding of the cultural and environmental contexts of the indigenous communities. An evaluation will be carried out to determine how accurate and pertinent the surrounding environment is concerning the selected indigenous community.

Skin Color

We meticulously analyze the portrayal of skin tone in images of indigenous communities, recognizing its significance in depicting their diverse cultural identity. Through this examination, we aim to capture celebrate the rich variety and cultural uniqueness within these communities.

Image Generation Prompts

Evaluating how prompts can change the results.

3. Data Interpretation:

- Rate each Google image with the AI-generated image out of 10.
- There is a certain point to keep in mind.

If a particular image is not rated or there is blank space it means that either of the images is in black and white, or the AI-generated images are unclear with no proper face structures.

5. Discussion:

The research yielded significant insights into the portrayal of indigenous communities by Artificial Intelligence. The key findings are summarised as follows:

5.1 Accuracy

With an accuracy rate of 70% for the majority, most AI-generated pictures appear to be quite faithful to the attributes they are meant to depict. There is space for development, though, as the produced photos still include a considerable amount of mistakes, despite the accuracy rate.

5.2 Facial markings:

The majority is 80 to 100% accurate, this is true as a good number of indigenous communities do not have facial markings, however, those who do were correctly depicted by a good count of the AI-generated images.

5.3 Facial Expressions

Facial expressions also get communicated within indigenous contexts through emotions represented in pictures generated by AI at a satisfactory level which is 70%. Nonetheless, refinement needs to be done here too because sometimes slight differences may occur leading to wrong interpretations concerning subtle feeling states among people belonging to such communities.

5.4 Clothing

As far as clothing is concerned, it achieves 80% accuracy for the majority thereby showing traditional dresses correctly while reflecting on rich tapestries woven around different cultures worldwide including those specific regions where these clothes were worn. However, still should work more on this algorithm so that fidelity towards various historical backgrounds or even local variations within certain areas can be enhanced further.

5.5 Demographic factors

In demographic aspects, there is only 60% correctness recorded for the majority by this system due to various challenges faced when dealing with such data types; although capturing them generally right, more precision should be shown towards illustrating finer details accurately about indigenous demographics.

5.6 Skin color 100%

The AI performs excellently well in terms of skin color with a perfect 100% accuracy for the majority that guarantees true representation and respects the global diversity among different tribes or nations.

5.7 Ornaments 85%

AI-generated images for indigenous ornaments have an accuracy rate of 85% for the majority and remarkably represent their intricate details and cultural value. However, there are some inaccuracies that pass through which can still be improved.

References:

1. Bender, E. M., et al. (2021). Dangers and Biases in Language Models. Communications of the ACM.
2. Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of the 1st Conference on Fairness, Accountability and Transparency.
3. Chouldechova, A. (2017). Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments. Big Data.
4. Crawford, K., et al. (2019). AI Now 2019 Report. AI Now Institute.
5. Eubanks, V. (2018). Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin's Press.
6. Ensign, D., et al. (2018). Runaway Feedback Loops in Predictive Policing. Science.
7. Noble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press.
8. Obermeyer, Z., et al. (2019). Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations. Science.
9. Allen, G. C., et al. (2019). Towards Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims. Harvard Kennedy School.
10. Barocas, S., & Selbst, A. D. (2016). Big Data’s Disparate Impact. California Law Review, 104(3), 671–732.
11. Bostrom, N., & Yudkowsky, E. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
12. Floridi, L., et al. (2018). Ethics of AI and Robotics. Stanford Encyclopedia of Philosophy.
13. Jobin, A., et al. (2019). The Global Landscape of AI Ethics Guidelines. Nature Machine Intelligence.

------
‍
1. https://www.kaggle.com/datasets/nipunarora8/age-gender-and-ethnicity-face-data-csvArora, N. (n.d.). AGE, GENDER AND ETHNICITY (FACE DATA) CSV.
2. Image Classification on ImageNet
3. MF2 Training Dataset
4. https://arxiv.org/pdf/2112.03109.pdf Zheng, Y., Yang, H., Zhang, T., Bao, J., Huang, Y., Yuan, L., Chen, D., Zeng, M., & Wen, F. (1AD). General Facial Representation Learning in a Visual-Linguistic Manner (thesis). GitHub.
5. https://www.researchgate.net/publication/354858400_The_Santhals_Their_Culture_and_Traditions Soren, Priyanka & Jamir, Waluneba. (2021). The Santhals: Their Culture and Traditions.
6. https://www.grida.no/resources/7125. Nhttps://www.npr.org/2020/06/24/882683463/the-computer-got-it-wrong-how-facial-recognition-led-to-a-false-arrest-in-michig Allyn, B. (2020, June 24). “The computer got it wrong”: how facial recognition led to false arrest of black man. NPR. oble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press.

------
‍
Image links:
1. image 1
2. https://i.ytimg.com/vi/knVV9Jg26uU/maxresdefault.jpg
3. https://tse1.mm.bing.net/th?id=OIP.Ag9r2DF2Xmh0AAAdDOZEkgHaFj&pid=Api&P=0&h=180
4.

Indigenous Representation in AI: 21st Century Implication