Decoding Injustice: How Hidden AI Bias Proposes Fatal Outcomes for Black Individuals

In the labyrinth of technological advancements where artificial intelligence (AI) promises a future of unparalleled convenience and efficiency, a shadow looms, stark and troubling. It’s the shadow of bias, a specter that haunts the algorithms designed to make life easier but instead perpetuates age-old prejudices.

The impact of AI bias, especially against Black individuals, is not just a statistical anomaly—it’s a lived reality, illustrated by instances where AI-driven decisions lead to unfair treatment and discrimination. The presence of language biases in AI and machine learning algorithms further complicates this issue and perpetuates systemic discrimination.

The Cornell University study that first identified this issue holds immense significance in this context. The study peeled back the layers of Black individuals AI’s operations, revealing how large language models (LLMs) like GPT-4 and LLaMA2 are not just neutral tools but are in fact imbued with covert racism.

This investigation into the “covert racism” of LLMs exposed a troubling aspect: the propensity of these models to criminalize African American English (AAE), casting a stark light on the implications of unchecked AI biases.

This article seeks to explore these complexities, with the goal of sparking a discussion on the importance of inclusive AI development practices and strong ethical frameworks.  By unpacking the layers of AI bias, particularly in language models, and exploring the findings of Black individuals pivotal studies like that of Cornell University, the goal is to contribute to a broader understanding of and action towards combating bias in AI.

Understanding AI and Language Bias of Black Individuals

AI bias, particularly in language models, roots itself in the very foundation of AI development: data and algorithm design. These biases aren’t merely reflections of individual prejudices, but are systemic, woven into the vast datasets that train AI models and the algorithms that process this data. Advanced AI language models, such as GPT-4 and LLaMA2, use data patterns to make predictions and decisions. But when the data contains societal biases, these models may unintentionally spread discrimination.

The Cornell University study serves as a critical lens, providing a detailed analysis of how LLMs process language and identifying potential sources of bias. One of its key findings is the disproportionate manner in which LLMs criminalize Black individuals like African American English, attributing negative stereotypes and outcomes more frequently to this dialect. This revelation is pivotal, highlighting not just the existence of bias, but its active contribution to systemic discrimination.

Unraveling the mechanics of language models and AI bias requires defining complex terms such as “matched guise probing” and “covert racism.” Matched guise probing is a technique used to study biases by presenting identical content with varying language variations and analyzing how responses differ. Covert racism of Black individuals in AI is the unseen and subtle prejudices embedded within AI systems, which can have substantial societal consequences such as unfair hiring practices and biased legal decisions.

The Depth of the Problem

AI bias isn’t just a research topic; it has real-life effects. In-depth analyses, like those studying AI use in hiring and legal decisions, show the seriousness of the problem. These studies show that instances of bias are not isolated, but an ongoing concern that has a significant impact on disadvantaged people.

AI-powered hiring systems, meant to make hiring more efficient, show biases that have serious effects. Studies have found that AI algorithms used in job screening may favor candidates from some racial groups over others.

AI systems are up to 50% less likely to shortlist candidates with names linked to African American or Hispanic communities, compared to Black individuals names perceived as “White.” This unfairness undermines equal opportunity in hiring and keeps historical biases in the job market alive.

AI bias affects law enforcement, particularly in predictive policing tools used to identify possible crime areas. However, research conducted by the AI Now Institute found that predictive policing tools are 3.5 times more likely to attribute higher risk scores to minority neighborhoods, irrespective of actual crime rates.

A study showed that in some cities, predictive policing software was more likely to label areas with a higher percentage of Black residents as potential crime hotspots, even if the actual crime rates were not higher than other areas with fewer Black individuals residents. This highlights the importance of carefully reviewing the data used in AI systems to ensure they don’t contribute to negative consequences in the real world.

Personal stories from people who have been affected by AI bias make the issue more relatable. These include job seekers unfairly rejected by automated systems and individuals wrongly accused because of flawed facial recognition software. Areport by the ACLU shared the story of a man mistakenly arrested based on a faulty facial recognition match, Black individuals demonstrating the serious consequences of unchecked AI biases.

Another alarming case involved a facial recognition system misidentifying a Black individual as a criminal suspect, leading to a wrongful arrest. The NAACP Legal Defense Fund found this exact scenario occurs on average at a rate five times higher among African Americans than their Caucasian counterparts. This highlights the need for solutions that tackle both the technical and human aspects of this growing issue.

Analyzing the Causes

To address AI bias effectively, it’s crucial to understand its roots. A closer look at the data sources used to train AI models reveals a reflection of societal biases. The content that feeds into AI systems—whether it’s text from the internet, historical documents, or media archives—carries the prejudices and stereotypes of the societies that produced them. If biased data is used to build AI models and decisions are based on these models, it will reinforce these existing biases.

AI algorithms, particularly those powered by machine learning, scrutinize data to uncover patterns and make inferences. When these Black individuals patterns are distorted, the algorithms not only reflect the inherent biases but also intensify them. This leads to more pronounced and impactful biased results.

A review of how AI is developed shows it’s easy for biases to enter or go unnoticed at any stage, from gathering data to using models. This demonstrates the need for AI development to be more cautious, actively identifying and addressing biases at every stage.

Solutions and Mitigation Strategies

Fixing bias in AI involves using several methods, from changing technology to making new laws. The tech industry realizes how important this problem is, and companies and scientists are trying to find and fix the problem.

They are doing things like checking algorithms, where AI systems are closely analyzed looking for any biases, and using bias correction methods, which adjust AI models to make up for any biases that have been found. While these practices are a step in the right direction, they often correct Black individuals biases after they have already affected software. Instead, we need to take steps before bias gets into AI systems.

AI advancements offer new ways to fight bias. One method, counterfactual data augmentation, supplements fake data with different results, which reduces the need for old biased data. AI systems can now explain their decisions, making it easier for people to find any biases. Moreover, involving a wide range of people in the design of AI through participatory design processes, allows for different viewpoints and helps reduce biases from the start.

There’s an increasing demand for ethical standards in AI development. Supporters lobby for strong rules and principles to guide the development and use of AI. These standards should value fairness, clarity, and taking responsibility. They can help AI developers build systems that respect and protect human rights and dignity. The Institute of Electrical and Electronics Engineers (IEEE)’s Ethically Aligned Design and the EU’s Ethics Guidelines for Trustworthy AI are successful examples providing detailed guidance for developing AI ethically.

The Role of Stakeholders

To eliminate AI bias, collaboration is crucial, extending beyond AI creators and researchers to policymakers, regulators, and the public. Technology companies, playing a significant role in AI development, must prioritize ethical usage. These companies must be responsible for testing AI systems for bias, ensuring transparency in their AI operations. Furthermore, companies must embrace transparency, accountability, and independent verification, sharing the results to foster trust.

Government policies and regulations play a key role in shaping the development and use of AI technologies. By setting guidelines and requirements, they can provide a structure that AI companies must adhere to. Policymakers have the power to guide the direction of AI development by creating laws that focus on promoting ethical practices, reducing bias, and ensuring transparency. These regulations can draw inspiration from existing frameworks for data protection and consumer rights, while tailoring them to the specific challenges posed by AI.

To build AI systems that are welcoming and non-discriminatory, it’s crucial to involve the community. By including people from all walks of life in the development of AI, we make sure that it reflects a broad spectrum of opinions and experiences. By doing this, we can prevent the creation of biased or unfair AI systems. This inclusive approach, sometimes known as participatory design, brings people with various perspectives together to help design, test, and provide feedback on AI systems, resulting in AI systems that are fair and beneficial for everyone.

Conclusion

Building unbiased artificial intelligence (AI) is a complex task with both obstacles and opportunities. It’s crucial to tackle AI bias because ignoring it can lead to unfairness and harmful discrimination. Encouragingly, the AI community is recognizing the importance of responsible development and taking steps to reduce bias.

Right now, we have a special chance to change how AI fits into our society. We can develop AI technology in a way that helps all, and makes systems more fair, instead of using it to alienate or target specific groups. As we move ahead, we must ask ourselves if we will use AI’s power to make the world a better place, or if we will let bias and unfairness control how technology develops in the future.

The discussions about bias in artificial intelligence have gained momentum thanks to academic research, like the Cornell University study, and ongoing conversations with AI ethics experts. It’s positive to see growing awareness and commitment to tackling AI bias. This progress paves the way for the development of AI technologies that prioritize inclusivity, justice, and the common good.

References

Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (n.d.). Machine Bias. ProPublica. Retrieved March 22, 2024, from https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

Beaulieu, A., & Leonelli, S. (2022). Data and Society: A Critical Introduction. SAGE Publications Ltd. https://ore.exeter.ac.uk/repository/bitstream/handle/10871/127993/Data%20and%20Society_Preprint.pdf?sequence=2

Berkman Klein Center. Ethics and Governance of AI Initiative. (n.d.). Retrieved March 22, 2024, from https://cyber.harvard.edu/projects/ethics-and-governance-ai-initiative

Blodgett, S. L., Barocas, S., Daumé III, H., & Wallach, H. (2020, July). Language (Technology) is Power: A Critical Survey of “Bias” in NLP. ACL Anthology. https://aclanthology.org/2020.acl-main.485/

Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf

Chugh, H. (2023, July 11). Predictive Policing — A Double-Edged Sword. Medium. https://medium.com/@harshaan.chugh/predictive-policing-a-double-edged-sword-a43cb9e7686

Desmarais, A. (2024, March 9). AI models found to show language bias by recommending Black defendents be “sentenced to death.” Euronews. https://www.euronews.com/next/2024/03/09/ai-models-found-to-show-language-bias-by-recommending-black-defendents-be-sentenced-to-dea

European Commission, Directorate-General for Communications Networks, Content and Technology (2019). Ethics guidelines for trustworthy AI, Publications Office. https://data.europa.eu/doi/10.2759/346720

FP Staff. (2024, March 11). Racist AI: ChatGPT, Copilot, more likely to sentence African-American defendants to death, finds Cornell study. Firstpost. https://www.firstpost.com/tech/racist-ai-chatgpt-copilot-more-likely-to-sentence-african-american-defendants-to-death-cornell-study-13747636.html

Future of Privacy Forum. (2024, February 29). Future of Privacy Forum Awarded National Science Foundation and Department of Energy Grants to Advance White House Executive Order on Artificial Intelligence. https://fpf.org/blog/future-of-privacy-forum-awarded-national-science-foundation-and-department-of-energy-grants-to-advance-white-house-executive-order-on-artificial-intelligence/

Hofmann, V., Kalluri, P. R., Jurafsky, D., & King, S. (2024, March 1). Dialect prejudice predicts AI decisions about people’s character, employability, and criminality. ArXiv. https://arxiv.org/abs/2403.00742

Hossain, S. Q., & Ahmed, S. I. (2021, May). Towards a New Participatory Approach for Designing Artificial Intelligence and Data-Driven Technologies. ArXiv. https://arxiv.org/pdf/2104.04072

Hsu, J. (2024, March 7). AI chatbots use racist stereotypes even after anti-racism training. New Scientist. https://www.newscientist.com/article/2421067-ai-chatbots-use-racist-stereotypes-even-after-anti-racism-training/

IEEE Standards Association. (n.d.). Ethically Aligned Design – A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems. Retrieved March 22, 2024, from https://standards.ieee.org/ieee-whitepaper/ethically-aligned-design-a-vision-for-prioritizing-human-well-being-with-autonomous-and-intelligent-systems/

Kamensky, J. M. (2013, November 3). Fighting Crime in a New Era of Predictive Policing. Governing. https://www.governing.com/archive/col-crime-fighting-predictive-policing-data-tools.html

Lee, N. T., Resnick, P., & Barton, G. (2019, May 22). Algorithmic bias detection and mitigation: Best practices and policies to reduce consumer harms. Brookings. https://www.brookings.edu/articles/algorithmic-bias-detection-and-mitigation-best-practices-and-policies-to-reduce-consumer-harms/

Müller, V. C. (2020, April 30). Ethics of Artificial Intelligence and Robotics. Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/ethics-ai/

Najibi, A. (2020, October 24). Racial Discrimination in Face Recognition Technology. Science in the News. https://sitn.hms.harvard.edu/flash/2020/racial-discrimination-in-face-recognition-technology/#:~:text=This%20result%20corroborated%20an%20earlier,incorrectly%20matched%20with%20mugshot%20images.

Omwanda, E. (2024, March 9). Uncovering Language Bias: AI Models Implicated In Covert Racism Study. Cryptopolitan. https://www.cryptopolitan.com/language-bias-ai-in-covert-racism-study/

Radford, A., Wu, J., Amodei, D., Amodei, D., Clark, J., Brundage, M., & Sutskever, I. (2019, February 14). Better language models and their implications. OpenAI. https://openai.com/research/better-language-models

Ruder, S. (2017, July 25). Deep Learning for NLP Best Practices. Ruder.Io. https://www.ruder.io/deep-learning-nlp-best-practices/

Selbst, A. D., Boyd, D. M., Friedler, S. A., Venkatasubramanian, S., & Vertesi, J. (2019, January 29). Fairness and Abstraction in Sociotechnical Systems. ACM Digital Library. https://dl.acm.org/doi/10.1145/3287560.3287598

Thong, J. L. K. (2024, January 23). Explaining the Crosswalk Between Singapore’s AI Verify Testing Framework and The U.S. NIST AI Risk Management Framework. Future of Privacy Forum. https://fpf.org/blog/explaining-the-crosswalk-between-singapores-ai-verify-testing-framework-and-the-u-s-nist-ai-risk-management-framework/

West, S. M. (2019, April 1). Discriminating Systems: Gender, Race, and Power in AI – Report. AI Now Institute. https://ainowinstitute.org/publication/discriminating-systems-gender-race-and-power-in-ai-2