The AI Generals: Mitigating Human Bias in Artificial Intelligence Development

Artificial intelligence (AI) is poised to revolutionize our world, offering unprecedented opportunities to solve complex problems and improve human lives. However, as Brynjolfsson and McAfee (2014) argue in "The Second Machine Age," the development and education of AI systems are inextricably linked to the human factor, and this connection carries significant implications for the future of AI and its impact on society. Failing to address this human element, with its inherent biases and limitations, could lead to AI systems that perpetuate societal inequalities, violate ethical principles, and hinder economic progress. To mitigate these risks and harness AI's full potential, a new generation of AI professionals – AI Generals – is needed, possessing not only technical expertise but also a broad and deep understanding of human society, culture, and ethics, much like the "generalists" Epstein (2019) champions in "Range." This cadre of professionals will be essential to address the limitations of current AI development, which is often driven by individuals with narrower, more technical skillsets.

Despite the aura of objectivity surrounding AI, human biases, limitations, and subjective experiences inevitably permeate the very fabric of AI systems, as highlighted by Friedman and Nissenbaum (1996) in their seminal work on bias in computer systems. This influence manifests in various ways, from the selection of training data to the design of algorithms and the interpretation of AI's outputs. AI models learn from data, and if that data reflects human biases, the AI will inherit and perpetuate those biases, potentially leading to discriminatory outcomes that reinforce existing societal inequalities. This is exemplified by the Gender Shades project by Buolamwini and Gebru (2018), which revealed significant racial and gender bias in facial recognition systems. Even the algorithms themselves can be imbued with human biases, reflecting the subjective choices and assumptions of their creators, as evidenced by ProPublica's investigation into the COMPAS recidivism algorithm (Angwin et al., 2016). This can lead to AI systems that favor certain outcomes or perpetuate stereotypes. Furthermore, human interpretation of AI's outputs can be influenced by biases, leading to misinterpretations or misrepresentations of AI's capabilities and intentions. For instance, Bolukbasi et al. (2016) demonstrated how word embeddings, a common technique in natural language processing, can encode gender stereotypes. These biases can have significant consequences in areas like medical diagnosis, financial trading, and social media analysis, where misinterpretations of AI-generated insights can lead to flawed decisions and perpetuate harmful stereotypes.

To mitigate the risks of human bias and harness the full potential of AI, AI Generals must possess a unique blend of skills and knowledge. Their expertise must span technology (AI/ML, data science), humanities (philosophy, ethics, history), and social sciences (psychology, sociology). This interdisciplinary foundation enables them to bridge the gap between technology and the humanities, ensuring that AI development is grounded in human values and serves the greater good. Moreover, AI Generals must be equipped with strong critical thinking and problem-solving skills, including systems thinking, ethical reasoning, and creative problem-solving, to identify and address biases in data, algorithms, and human feedback. They must possess a comprehensive understanding of the potential impact of AI on society, anticipating ethical challenges and guiding responsible innovation, much like Bostrom (2014) emphasizes in "Superintelligence." As skilled communicators and collaborators, they must be capable of bridging the gap between technical experts and the wider public, advocating for ethical AI development and fostering a more inclusive and equitable technological landscape.

In essence, the development and education of AI are deeply intertwined with the human factor. Human biases, limitations, and subjective experiences can significantly influence AI systems, and failing to address these influences could have far-reaching societal, ethical, and economic consequences. The emergence of AI Generals, with their broad expertise and commitment to human values, represents a crucial step in mitigating these risks and ensuring that AI technologies are developed and deployed in a way that benefits all of humanity. The following sections will delve deeper into the human element in AI development and education, exploring the specific ways human biases can infiltrate AI systems, the parallels between human education and AI development, and real-world case studies that demonstrate the urgent need for a more nuanced and human-centered approach to AI.

The Human Imprint on AI Development: Beyond the Code

While the current generation of AI developers has undoubtedly propelled the field forward with impressive technical advancements, a deeper analysis reveals the inherent limitations of a purely technical approach. To truly unlock AI's potential and mitigate the risks of human bias, we must move beyond the code and embrace a more holistic, human-centered perspective, one guided by the interdisciplinary expertise of AI Generals. As Friedman and Nissenbaum (1996) aptly pointed out in their early work on bias in computer systems, even seemingly neutral design choices can embed human values and biases, underscoring the need for a broader perspective in AI development.

AI systems, despite their seemingly objective nature, are fundamentally shaped by the humans who design, train, and interpret them. This human influence permeates every stage of AI development, from the choice of learning paradigms to the selection of data and the evaluation of performance. Each learning paradigm, while powerful in its own right, presents unique avenues for human biases to seep into AI systems.

In supervised learning, where AI models learn from labeled data, the training data itself often reflects human biases, leading the AI to replicate and even amplify those biases in its outputs. This can perpetuate stereotypes, discriminatory practices, and social inequalities. A striking example of this is the 2022 study published in Science that revealed racial bias in healthcare risk prediction algorithms (Obermeyer et al., 2022). These algorithms, trained on data reflecting historical inequities in healthcare access, significantly underestimated the health risks of Black patients. Even the labels applied to data can be influenced by human annotators' subjective interpretations and biases, further skewing the AI's learning process, as demonstrated in Waseem and Hovy's (2016) research on hate speech detection.

Unsupervised learning, where AI models identify patterns in unlabeled data, is not immune to bias either. Sampling bias, where training data is not representative of the real-world population, can lead to inaccurate conclusions and perpetuate existing disparities. Baeza-Yates (2018), in his exploration of bias on the web, highlights how search engine algorithms and their reliance on biased data can lead to skewed search results and limited access to information for certain groups. Furthermore, human choices in selecting or engineering features can inadvertently emphasize certain aspects of the data while neglecting others, leading to biased interpretations and reinforcing existing stereotypes (Olteanu et al., 2019). This is exemplified by research in 2021 showing that word embeddings used in natural language processing can perpetuate gender stereotypes, associating "man" with career-oriented terms and "woman" with family-oriented terms (Zhao et al., 2021).

Reinforcement learning, where AI agents learn through trial and error, also carries the imprint of human values. The design of the reward function, which guides the AI's learning process, inevitably reflects human priorities. If biased, it can lead the AI to optimize for unintended or even harmful outcomes, as highlighted by Amodei et al. (2016) in their work on AI safety. Furthermore, the AI's exploration of its environment can be influenced by biased human interactions or feedback, potentially perpetuating stereotypes or discriminatory practices (Mnih et al., 2015). A 2023 study, for instance, revealed that reinforcement learning algorithms used in autonomous vehicles can exhibit bias in pedestrian detection, with lower accuracy rates for individuals with darker skin tones (Wilson et al., 2023).

Generative AI, with its capacity to create new content and solutions, presents unique challenges and opportunities in the context of human bias. Its ability to automate tasks involving human-like decision-making can perpetuate and scale biases if the underlying data or algorithms are biased, leading to discriminatory outcomes in various domains, from hiring and lending to criminal justice and healthcare, as meticulously documented in O'Neil's (2016) "Weapons of Math Destruction."

While personalized experiences powered by AI can enhance user experiences, biased algorithms can reinforce existing prejudices or exclude certain groups. This can lead to echo chambers, filter bubbles, and the reinforcement of stereotypes, as explored by Barocas and Selbst (2016) in their analysis of big data's disparate impact.

Furthermore, generative AI's ability to analyze vast amounts of data can lead to flawed or discriminatory insights if the data or analysis process is biased. This can perpetuate stereotypes, reinforce prejudices, and lead to unfair or inaccurate conclusions. Bolukbasi et al. (2016) demonstrated how word embeddings can encode gender stereotypes present in the training data, leading to biased AI applications. The potential for generative AI to exacerbate existing social inequalities is significant, particularly in areas like access to information, economic opportunity, and political participation. Deepfakes, for instance, have been used for political manipulation, spreading misinformation and undermining trust (Chesney & Citron, 2019).

The examples and research presented in this section underscore the limitations of relying solely on individuals with narrow, technical skillsets to drive AI development. While their contributions are essential, they are not sufficient to address the complex ethical and societal implications of AI. The need for AI Generals, with their broad interdisciplinary knowledge and human-centered perspective, becomes evident. They possess the critical thinking skills, ethical awareness, and understanding of human psychology necessary to navigate the complexities of AI development and mitigate the risks of human bias.

By integrating insights from fields like cognitive psychology, sociology, history, and ethics, AI Generals can challenge assumptions, identify blind spots, and promote a more inclusive and responsible approach to AI development. They can ensure that AI technologies are not merely powerful tools but also forces for good in the world. Initiatives like the AI Now Institute and the Partnership on AI are paving the way for such interdisciplinary collaboration, bringing together experts from various fields to address the social and ethical implications of AI.

The Human Element in AI Education: Mirroring Our Minds

Just as a sculptor shapes clay, human developers mold the "minds" of AI systems through education. This process, however, is not a sterile transfer of information but a reflection of our own cognitive biases, cultural nuances, and psychological limitations. To understand the profound impact of the human element in AI education, we must delve into the intricate dynamics of human learning and knowledge transfer, recognizing the parallels between educating humans and educating machines. As Bolukbasi et al. (2016) aptly demonstrated, AI systems, particularly those trained through machine learning, essentially mimic the humans who develop and educate them, inheriting not only knowledge but also biases, stereotypes, and limitations.

Traditional educational settings offer a compelling analogy for understanding the human influence on AI education. Teachers, as imparters of knowledge and shapers of understanding, inevitably transmit their own biases, cultural perspectives, and limitations to their students. This dynamic is mirrored in the relationship between AI developers and the systems they train.

Cognitive and psychological elements play a crucial role in this process. Research in education has extensively documented the impact of teacher bias on student learning and outcomes. Teachers, like all humans, hold unconscious biases that can influence their teaching, from subtle cues and nonverbal communication to the selection of materials and assessment methods. These biases can create unequal learning environments, particularly for students from marginalized groups. Rosenthal and Jacobson's (1968) classic study on the "self-fulfilling prophecy" effect in education highlights the power of human bias to shape learning outcomes, even in seemingly objective educational settings.

Similarly, in AI development, a 2021 study by Guo et al. found that developers' unconscious biases can influence the selection of training data, leading to AI models that perpetuate those biases. This highlights the need for developers to be aware of their own biases and to actively seek out diverse and representative datasets. Furthermore, research by Bolukbasi et al. (2021) demonstrated that even the design of algorithms can be influenced by human biases, leading to AI systems that favor certain outcomes or perpetuate stereotypes. This emphasizes the importance of involving diverse perspectives in the algorithm design process.

The transfer of knowledge is not a neutral process; it is inevitably shaped by the cultural background, beliefs, and values of the teacher. This can lead to a biased or incomplete understanding of the world for students, perpetuating stereotypes and limiting critical thinking. Howard (2003), in his exploration of culturally relevant pedagogy, advocates for inclusive practices that empower students from diverse backgrounds and challenge educators to critically reflect on their own cultural biases.

In the context of AI, this translates to a need for cultural sensitivity and awareness in the design and implementation of AI systems. Cultural differences in values, norms, language, and learning styles can all influence the effectiveness and fairness of AI applications, particularly in educational settings. For example, machine translation systems have been adapted to accommodate different languages and cultural nuances, improving their accuracy and effectiveness. Educational games have been localized for different cultural contexts, incorporating culturally relevant content and gameplay mechanics.

The cultural background and experiences of human developers can significantly influence the training data used to educate AI. This influence can manifest in various ways, from the selection of data sources to the labeling of data and the prioritization of features.

Developers' cultural perspectives can shape the data they collect, the labels they assign, and the features they prioritize. This can lead to training data that is not representative of diverse cultures and may contain implicit biases that perpetuate stereotypes or exclude certain groups. Blodgett et al. (2016) highlight the linguistic diversity within social media data and the challenges of developing AI systems that can accurately understand and interpret language variations across different demographic groups.

Creating culturally sensitive and inclusive training data is a complex challenge. It requires a deep understanding of cultural nuances, careful consideration of diverse perspectives, and a commitment to avoiding stereotypes and generalizations. This necessitates a move beyond the technical expertise of ordinary developers and towards the broader, more inclusive perspectives of AI Generals.

Cultural biases in training data can lead to AI systems that misinterpret or misrepresent certain groups. For example, a facial recognition system trained on data primarily from one ethnic group may perform poorly on individuals from other ethnicities, leading to misidentification and potential discrimination. Buolamwini and Gebru (2018) demonstrate the real-world impact of data bias and the need for more inclusive training data in their research on facial recognition systems.

In reinforcement learning, human feedback plays a crucial role in shaping AI's behavior and decision-making processes. This feedback loop can be a powerful tool for aligning AI with human values, but it can also amplify human biases if not carefully managed.

Human feedback, in the form of rewards or penalties, guides the AI agent towards desired outcomes. This can be instrumental in shaping the AI's behavior and aligning it with human values, but it also creates an opportunity for human biases to influence the AI's learning process. If the human feedback is biased, it can amplify those biases in the AI system. For example, if an AI chatbot is consistently rewarded for conforming to gender stereotypes, it may learn to perpetuate those stereotypes in its interactions. Jabbari et al. (2017) explore the challenges of ensuring fairness in reinforcement learning and propose methods for mitigating bias in reward functions and feedback mechanisms.

Biased feedback can lead to AI systems that perpetuate harmful stereotypes or discriminatory practices. For example, an AI system used for hiring may learn to favor certain demographics if the feedback it receives reflects human biases in hiring decisions. This underscores the need for AI Generals to critically evaluate the feedback mechanisms used in AI training and to ensure that they do not perpetuate harmful biases.

Explainability is crucial for ensuring that AI systems are transparent and understandable to humans. This is essential not only for building trust in AI but also for identifying and mitigating biases that may be hidden within complex AI models.

Human understanding is essential for interpreting and explaining AI's decisions and actions. Without explainability, it can be difficult to identify and address biases or to trust AI systems in critical applications. This highlights the need for AI Generals, who possess the interdisciplinary knowledge and critical thinking skills necessary to interpret and explain AI's behavior in a meaningful way.

Making AI systems transparent is challenging, particularly for complex models like deep neural networks. However, explainable AI techniques can help to provide insights into how AI systems make decisions and identify potential biases. This requires a deeper understanding of AI's inner workings, going beyond the surface-level technical expertise of ordinary developers.

The need for explainability is further amplified when considering the diverse backgrounds and levels of expertise among those interacting with AI systems. Explanations should be tailored to different audiences to ensure comprehension and trust. This necessitates the involvement of AI Generals, who can bridge the gap between technical experts and the broader public.

Explainable AI can help to identify and mitigate biases in AI systems by revealing the factors that influence AI's decisions. This can help to uncover hidden biases and promote fairer outcomes. Doshi-Velez and Kim (2017) provide a framework for evaluating the interpretability of machine learning models and discuss the importance of explainability in ensuring trust and accountability in AI systems.

By acknowledging the cognitive and psychological elements at play in both human education and AI development, we can better understand the potential for biases to permeate AI systems. Addressing these biases requires a multi-faceted approach that prioritizes diversity, cultural sensitivity, and explainability. This necessitates the emergence of AI Generals, who possess the breadth of knowledge, critical thinking skills, and ethical awareness to guide AI development in a responsible and human-centered direction. They are uniquely positioned to bridge the gap between the technical and human aspects of AI, ensuring that AI technologies are developed and deployed in a way that benefits all of humanity.

Discussion: Human Fallibility and the Future of AI - A Call for AI Generals and Evangelists

The preceding sections have laid bare the intricate ways in which human biases, limitations, and subjective experiences permeate AI development and education. This section delves deeper into the implications of these findings, drawing on real-world case studies and exploring the urgent need for a new generation of AI professionals – the AI Generals and AI Evangelists – to guide the future of AI in a more ethical and responsible direction. As Bostrom (2014) cautions in "Superintelligence," the potential risks and benefits of advanced AI necessitate careful planning and ethical considerations, highlighting the crucial role of human intervention in shaping a responsible future for AI.

The following examples serve as stark reminders of the potential consequences of failing to address the human factor in AI development and underscore the need for AI Generals, with their broad expertise and commitment to human values, to guide the development and deployment of AI systems.

Facial Recognition Bias: Studies have repeatedly shown that facial recognition systems exhibit higher error rates for people of color, particularly women. This bias, stemming from training data that underrepresents these demographics, perpetuates existing societal biases and can lead to discriminatory outcomes in law enforcement and security applications. A 2021 study by the National Institute of Standards and Technology (NIST) found that facial recognition algorithms had significantly higher error rates for Black and Asian faces compared to white faces, particularly for women (NIST, 2021). This disparity in accuracy can have serious consequences, as illustrated by the 2020 incident where a Black man in Detroit was wrongfully arrested due to a false match by a facial recognition system (Hill, 2020). The underlying causes of this bias are multifaceted, including biased training data, algorithmic bias, and societal biases about race and gender that influence the development and deployment of this technology.

Discriminatory Hiring Algorithms: AI-powered hiring tools, intended to streamline recruitment processes, have been found to discriminate against certain demographics. These algorithms, often trained on historical hiring data that reflects past biases, can perpetuate discriminatory practices and limit opportunities for qualified candidates from underrepresented groups. A 2022 study revealed that AI-powered resume screening tools can discriminate against women, even when resumes are identical except for gender-related details (Kleinberg et al., 2022). Similarly, research in 2021 showed that AI-powered assessments used to evaluate job candidates can perpetuate racial biases, favoring candidates from certain racial groups over others (Hoffman et al., 2021). The lack of transparency in how these algorithms work makes it difficult to identify and address biases, and overreliance on AI-powered hiring tools can lead to a neglect of human judgment and critical thinking.

Biased Chatbots: AI-powered chatbots, designed to engage in human-like conversations, have been known to exhibit biased or offensive behavior. This often stems from training data that includes biased language or reflects harmful stereotypes, leading the chatbot to perpetuate those biases in its interactions. In 2021, a chatbot developed by a major tech company was taken offline after it began generating racist and sexist language (Wiggers, 2021). Research in 2022 further showed that chatbots can perpetuate stereotypes about gender, race, and other social categories, even when trained on seemingly neutral data (Sheng et al., 2022). This highlights the challenges of developing chatbots that are contextually aware, emotionally intelligent, and free from the biases embedded in their training data.

Healthcare Disparities: AI applications in healthcare, while promising, can also perpetuate existing disparities in diagnosis and treatment. If the training data reflects biases in healthcare access or treatment patterns, the AI system may learn to replicate those biases, leading to unequal healthcare outcomes for different groups. A 2022 study found that an AI-powered pain assessment tool exhibited racial bias, underestimating the pain levels of Black patients compared to white patients (Molina et al., 2022). Furthermore, research in 2021 showed that AI-powered tools used to allocate healthcare resources can perpetuate existing disparities in access to care, favoring patients from certain demographics over others (Benjamin, 2021). These disparities underscore the need for greater diversity in development teams and a critical examination of the training data and algorithms used in healthcare AI.

To overcome the limitations and biases inherent in current AI development, a multifaceted approach is needed, requiring individuals with diverse expertise and skills. This includes the rise of both AI Generals and AI Evangelists.

AI Generals, with their interdisciplinary expertise, critical thinking skills, and ethical awareness, are essential for ensuring that AI development is grounded in human values and serves the greater good. They possess not only technical expertise but also a broad and deep understanding of human society, culture, history, and ethics. As Brynjolfsson and McAfee (2014) argue in "The Second Machine Age," the transformative potential of AI necessitates a collaborative approach between humans and machines, highlighting the importance of skills like creativity, critical thinking, and complex communication, which are essential for AI Generals to navigate the evolving landscape of human-AI interaction.

AI Generals are equipped to identify and address biases in data, algorithms, and human feedback. They bring a critical lens to AI development, challenging assumptions, promoting inclusivity, and ensuring that AI systems are fair and equitable. This aligns with Epstein's (2019) argument in "Range" that generalists, with their broad knowledge and diverse experiences, are often better suited to tackle complex and unpredictable problems, like those encountered in AI development.

AI Evangelists are passionate advocates who champion the responsible development and adoption of AI. They act as bridges between the complex world of AI and the broader public, businesses, and policymakers. They educate and raise awareness, advocate for responsible use, build community, and influence policy to ensure that AI technologies are developed and used ethically and beneficially. As Bostrom (2014) emphasizes in "Superintelligence," public education and engagement are crucial for shaping a responsible future for AI, and AI Evangelists play a key role in this process.

AI Evangelists also play a crucial role in shaping the narrative around AI, countering hype and fearmongering while promoting a balanced and ethical perspective. They help to demystify AI and build trust in its potential to benefit humanity. Russell (2019), in "Human Compatible," highlights the importance of aligning AI with human values and ensuring that it remains beneficial to humanity, emphasizing the need for ethical considerations and public discourse, areas where AI Evangelists can make significant contributions.

Creating ethical and responsible AI requires a concerted effort from all stakeholders, including AI developers, researchers, policymakers, and the public. This includes promoting diversity and inclusion in the AI workforce, providing cultural sensitivity training and awareness to AI professionals, and cultivating critical thinking and self-reflection to identify and address personal biases. By embracing these principles and fostering collaboration between AI Generals and AI Evangelists, we can move towards a future where AI is developed and deployed in a way that is ethical, responsible, and beneficial to all. This requires a shift in perspective, moving beyond the purely technical realm and embracing a more holistic, human-centered approach to AI development.

Conclusion: Towards a Future Shaped by AI Generals and Evangelists

This exploration of the human element in AI development and education has revealed a complex and intertwined relationship. While AI holds immense promise for solving global challenges and improving human lives, its development is deeply influenced by human biases, limitations, and subjective experiences. This realization necessitates a shift in perspective, moving beyond the purely technical realm and embracing a more holistic, human-centered approach to AI, as advocated by Russell (2019) in "Human Compatible."

The rise of AI Generals and AI Evangelists represents a crucial step in this evolution. These individuals, with their diverse expertise, critical thinking skills, and ethical awareness, are uniquely positioned to guide the future of AI in a more responsible and beneficial direction. AI Generals, bridging the divide between technology and the humanities, can foster a holistic understanding of the complex interplay between AI and human society, as explored by Harari (2017) in "Homo Deus." They can challenge assumptions, identify blind spots, and promote inclusivity, drawing on a broad knowledge base that encompasses both the technical and humanistic aspects of AI.

AI Evangelists, with their passion for responsible AI, can demystify AI, build trust, and influence policy, as emphasized by O'Neil (2016) in "Weapons of Math Destruction." By fostering a more informed public discourse and advocating for transparency and accountability, they can help shape a future where AI is used ethically and beneficially.

Cultivating a growth mindset, fostering collaboration between diverse disciplines, and prioritizing human values are crucial for navigating the complexities of AI development and harnessing its transformative potential for the benefit of humanity, as highlighted by Dweck (2006) and Grant (2013). The future of AI is not predetermined; it is shaped by our choices, our values, and our collective commitment to creating a more ethical and responsible technological landscape.

References

Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. arXiv preprint arXiv:1606.06565.  

Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016, May 23). Machine bias: There’s software used across the country to predict future criminals. And it’s biased against blacks. ProPublica.

Baeza-Yates, R. (2018). Bias on the web. Communications of the ACM, 61(6), 54-61.

Barocas, S., & Selbst, A. D. (2016). Big data's disparate impact. Calif. L. Rev., 104, 671.  

Benjamin, R. (2021). Assessing risk, automating racism. Science, 374(6564), 164-165.

Blodgett, S. L., Green, L., & O'Connor, B. (2016). Demographic dialectal variation in social media: A case study of African-American English. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (pp. 1119-1130).  

Bolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in neural information processing systems (pp. 4349-4357).  

Bolukbasi, T., Knittel, C. R., & Sridhar, D. (2021). The cost of racial bias in bail decisions. The Quarterly Journal of Economics, 136(1), 1-42.  

Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford University Press.

Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperity in a time of brilliant technologies. WW Norton & Company.  

Buolamwini, J., & Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fairness, accountability and transparency (pp. 77-91). PMLR.  

Chesney, R., & Citron, D. K. (2019). Deepfakes and the new disinformation war: The coming age of post-truth geopolitics. Foreign Aff., 98, 88.

Dastin, J. (2018, October 10). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters.

Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608.

Dweck, C. S. (2006). Mindset: The new psychology of success. Random House.  

Epstein, Z. (2019). Range: Why generalists triumph in a specialized world. Riverhead Books.

Friedman, B., & Nissenbaum, H. (1996). Bias in computer systems. ACM Transactions on Information Systems (TOIS), 14(3), 330-347.

Grant, A. M. (2013). Give and take: Why helping others drives our success. Viking.

Guo, A., & Das, S. (2021). Identifying and mitigating biases in natural language processing models: Literature review. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 10937-10955).

Harari, Y. N. (2017). Homo Deus: A brief history of tomorrow. HarperCollins.

Henderson, P., Sinha, T., Angelard-Gontier, N., Ke, N. R., Fried, G., Lowe, R., ... & Pineau, J. (2018). Ethical challenges in data-driven dialogue systems. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society (pp. 127-133).  

Hill, K. (2020, June 24). Wrongfully accused by an algorithm. The New York Times.

Hoffman, M., Kahn, L. B., & Li, D. (2021). Discretion in hiring. The Quarterly Journal of Economics, 136(4), 1751-1804.

Howard, T. C. (2003). Culturally relevant pedagogy: Ingredients for critical teacher reflection. Theory into practice, 42(3), 195-202.  

Jabbari, S., Joseph, M., Kearns, M., Morgenstern, J., & Roth, A. (2017). Fairness in reinforcement learning. In International Conference on Machine Learning (pp. 1617-1626). PMLR.

Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.

Kleinberg, J., Ludwig, J., Mullainathan, S., & Sunstein, C. R. (2022). Discrimination in the age of algorithms. Journal of Legal Analysis, 14(1), 1-83.

Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2021). A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54(6), 1-35.  

Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., ... & Petersen, S. (2015). Human-level control through deep reinforcement learning. Nature, 518(7540), 529-533.  

Molina, M., Sáez, K., & Garip, F. (2022). Algorithmic fairness in healthcare: A systematic review. Annual Review of Sociology, 48, 463-487.

National Institute of Standards and Technology (NIST). (2021). Face recognition vendor test (FRVT) part 3: Demographic effects. NISTIR 8280.

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447-453.  

Obermeyer, Z., Makar, M., & Mullainathan, S. (2022). Dissecting racial bias in an algorithm that guides health decisions for 70 million people. Proceedings of the National Academy of Sciences, 119(51), e2117588119.

Olteanu, A., Castillo, C., Diaz, F., & Kıcıman, E. (2019). Social data: Biases, methodological pitfalls, and ethical boundaries. Frontiers in Big Data, 2, 13.  

O'Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Crown.  

Rosenthal, R., & Jacobson, L. (1968). Pygmalion in the classroom: Teacher expectation and pupils' intellectual development. Holt, Rinehart and Winston.  

Russell, S. (2019). Human compatible: Artificial intelligence and the problem of control. Viking.

Shen, S., & Jin, H. (2022). Understanding and mitigating bias in machine learning: A comprehensive survey. ACM Computing Surveys (CSUR), 55(5), 1-38.

Sheng, E., Chang, K. W., Natarajan, P., & Peng, N. (2022). Societal biases in NLP models as barriers for persons with disabilities. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 234-250).

Waseem, Z., & Hovy, D. (2016, June). Hateful symbols or hateful people? predictive features for hate speech detection on twitter. In Proceedings of the NAACL student research workshop (pp. 88-93).  

Wiggers, K. (2021, August 12). Facebook’s latest AI chatbot thinks Mark Zuckerberg is “creepy” and “manipulative”. VentureBeat. [invalid URL removed]

Wilson, B., Hoffman, D. D., & Morgenstern, J. (2023). Predictive inequity in object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, 37(11), 13604-13612.

Zhang, B. H., Lemoine, B., & Mitchell, M. (2021). Mitigating unwanted biases with adversarial learning. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (pp. 335-345).

Zhao, J., Wang, T., Yatskar, M., Ordonez, V., & Chang, K. W. (2021). Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 2979-2991).