Open photo in lightbox

With each day, artificial intelligence ("AI") becomes more embedded into the fabric of our society. The number of decisions we make incorporating AI has increased exponentially in recent years. People use AI to help find a physician or diagnose an illness, discover new recipes and meal prep, plan vacations, and so much more. What happens when the very systems designed to optimize efficiency in our lives also have the potential to cause real harm? This boosted dependence on AI prompts a closer evaluation of the unintended consequences that occur when biased data impacts automated decision-making.

AI bias can be harmful because it amplifies issues of inequality instead of resolving them.¹ Many AI systems review past events and look for patterns, and this creates a serious problem when the past is filled with historic injustices and discrimination. Thus, these AI systems often continue to perpetuate unfair outcomes in new ways that are harder to identify.

AI bias is notably problematic in areas like hiring, criminal justice, and lending as it further hinders people who are already disadvantaged and treated unfairly by society. AI looks at what has happened or what decisions have previously been made and makes current recommendations for what should happen at present.² When companies are hiring, they often use AI to help screen thousands of resumes and applications. Based on previous hiring practices, AI tends to select applicants who resemble the people the company has hired before. Using this method, women and people of color might not get hired.³ This unfair practice is not an accurate indicator of who is most qualified, nor does it account for additional factors when decision-makers try to reconcile past injustices in modern hiring practices.⁴

Facial recognition technology is another avenue where bias shows up. These AI systems have flaws that create errors when attempting to recognize women and people of color. This can cause unnecessary security concerns leading to the denial of service at places like banks or when applying for a rental application where facial recognition is necessary to confirm a person’s identity.

Another concern is that oftentimes people rely fully on AI tools, believing AI is always truthful, accurate, and fair. This means there is minimal fact checking of AI systems. Without intentional vetting, human monitoring, and systems designed to detect and correct their own biases, AI risks embedding historic prejudices into decision-making processes at scale — causing real-world harm.

Even prior to its impact on real-world outcomes, bias in AI can manifest in different phases of production: data collection, data labeling, model training, and deployment.⁵ AI bias is not formed spontaneously and generally originates in the data collection process. AI algorithms learn through the process of inputting data, and when the data does not represent a diverse demographic of individuals, any AI outputs will reflect those biases.⁶

The introduction of bias can also occur when labeling the AI training data into subsets. Different human annotators often interpret the same data in multiple ways based on their varying lived experiences.⁷ If data categories are labeled subjectively, the resulting outcome can exhibit personal and cultural biases. When model training and developing, AI systems often reflect historical injustices because these AI tools are often trained on large collections of online texts and images, essentially real-world data containing patterns of inequality. This causes AI models to inherit cultural biases that mirror discriminatory practices such as racism, sexism, gender stereotyping, and ableism.⁸

In deployment, biases in AI content emerge in different ways such as exclusionary AI-generated images and inaccurate summaries of historic events, even if the bias seemingly did not appear in training.⁹ For example, the lack of diversity in fields such as computer science or computer engineering promote the practice of current AI hiring tools trained based on previous hiring data to favor white male applicants over Black female applicants, especially when the historical dataset reflects gender and racial imbalances in leadership roles. Additionally, many facial recognition software tends to perform poorly on darker skin tones due to the underrepresentation of subjects in the AI training data. These biases can result in discriminatory consequences in employment, lending, policing, and criminal justice, reinforcing systemic disparities rather than mitigating them.¹⁰

How does one ensure biases in AI are limited throughout the various stages? There are several strategies AI developers can utilize to mitigate generative bias in AI tools. This includes ensuring there is thorough documentation of the AI data generation process and confirming the AI input data used resembles reality as much as possible by consistently measuring synthetic data against actual datasets.¹¹ Another practice is maintaining traceability by documenting data sources and making modifications to correct errors or biases.¹² Further, it is important to incorporate a wide variety of individuals from different demographics and cultures in the AI input to enhance inclusivity in AI outputs.¹³

It is equally important to involve human experts to monitor, detect, and mitigate bias. Finally, AI developers must implement periodic monitoring schedules to recognize bias and minimize those biases by adjusting the data so existing biases are not fortified.¹⁴ As long as AI developers and consumers remain diligent in identifying bias and working to eliminate bias, the general public can collectively minimize the risk of perpetuating injustice and discrimination.

¹ Grillo, M. (2025, June 10). AI bias: Understanding and mitigating unfair outcomes in your AI systems. MyMobileLyfe. https://www.mymobilelyfe.com/artificial-intelligence/ai-bias-understanding-and-mitigating-unfair-outcomes-in-your-ai-systems/

² Smith, Genevieve, and Ishita Rustagi. Mitigating Bias in Artificial Intelligence: An Equity Fluent Leadership Playbook. Center for Equity, Gender and Leadership, University of California, Berkeley Haas School of Business, July 2020. https://haas.berkeley.edu/wp-content/uploads/UCB_Playbook_R10_V2_spreads2.pdf

³ Id.

⁴ Rivero, Nicolas. "How to Use AI Hiring Tools to Reduce Bias in Recruiting." World Economic Forum, Oct. 13, 2020, www.weforum.org/stories/2020/10/ai-hiring-tools-bias-recruiting-hiring-diversity-fairness-equality/

⁵ Chapman University, "Bias in AI," Chapman University AI Hub, https://www.chapman.edu/ai/bias-in-ai.aspx

⁶ Cano, Y. M., Venuti, F., & Martinez, R. H. (2023). ChatGPT and AI text generators: Should academia adapt or resist? Harvard Business Publishing. https://hbsp.harvard.edu/inspiring-minds/chatgpt-and-ai-text-generatorsshould-academia-adapt-or-resist

⁷ Chapman University, "Bias in AI," Chapman University AI Hub, https://www.chapman.edu/ai/bias-in-ai.aspx

⁸ Sheridan Libraries. (2025, September). Bias in AI: How to spot it, why it matters, and what you can do. Johns Hopkins University. https://www.library.jhu.edu/news/2025/09/bias-inai-how-to-spot-it-why-it-matters-and-whatyou-can-do/

⁹ Tuhin, M. (n.d.). The dark side of AI: Bias, surveillance, and control. Science News Today. https://www.sciencenewstoday.org/the-darkside-of-ai-bias-surveillance-and-control

¹⁰ Chapman University, "Bias in AI," Chapman University AI Hub, https://www.chapman.edu/ai/bias-in-ai.aspx

¹¹ Bano, Muneera. (2024, August 27). AI model collapse: Why diversity and inclusion in AI matter? LinkedIn. https://www.linkedin.com/pulse/ai-model-collapse-why-diversity-inclusion-matter-muneera-bano-7zzvc/

¹² Id.

¹³ Id.

¹⁴ Id. Smith, Genevieve, and Ishita Rustagi. Mitigating Bias in Artificial Intelligence: An Equity Fluent Leadership Playbook. Center for Equity, Gender and Leadership, University of California, Berkeley Haas School of Business, July 2020. https://haas.berkeley.edu/wp-content/uploads/UCB_Playbook_R10_V2_spreads2.pdf

Copyright @ USFN 2026

Winter 2026 USFN Report - Jan. 2026