AI in 2025 with Watsonx: Overcoming Bias and Building Trust

The potential for AI to change lives is immense, but its success hinges on one key factor—trust. Building that trust begins with acknowledging and addressing bias, a challenge deeply rooted in all existing AI models.

So, why is bias so persistent in AI?

Contrary to popular belief, bias doesn’t arise because AI systems are inherently flawed or programmed to do evil. It emerges from the training data we feed into them, which often reflects the historical inequities ingrained in the larger society.

Systemic disparities in education, employment, and healthcare are frequently encoded in the data collected from these sectors. A report by the National Academies of Sciences highlights how underfunded schools in marginalized communities create data that underrepresents student success, perpetuating inequity in educational predictions and AI models. Similar trends exist in employment and healthcare, where historical inequalities shape the datasets used in these sectors, leading to bias in AI models.

Stepping back, we can logically recognize that schools in underprivileged areas often lack resources, leading to data that underrepresents success rates, which will naturally skew AI predictions and lead to false assumptions about the abilities of the students, themselves.

Apply this reasoning to healthcare, where historical gaps in treatment data for minority populations have resulted in algorithms that fail to address their specific medical needs. Core medical training data often contains bias because it reflects historical inequalities, societal norms, and incomplete representations of diverse groups.

In the workforce domain, job descriptions historically contained gendered language that discouraged women from applying to certain roles. A study by researchers at the University of Edinburgh highlighted how language patterns in job ads, such as "dominant" or "competitive," significantly reduced applications from women. When AI systems used these datasets, they inadvertently perpetuated these biases in hiring recommendations.

When datasets are predominantly sourced from regions or groups with similar demographics, they fail to capture the nuances of a broader population, resulting in skewed outcomes. AI, after all, reflects the world as we’ve documented it—with all its imperfections, inequities, and systemic gaps.

Begin with building equity in data.

To tackle bias, we must begin with equity in data. If the data used to train AI systems is unbalanced or skewed, the outputs will mirror those inequities. One example is facial recognition software’s struggle with diversity. Passport photo verification systems have rejected images based on skin tone alone.

A study conducted by the National Institute of Standards and Technology (NIST) found that facial recognition systems had higher error rates for individuals with darker skin tones, caused by an evident gap in training data diversity. The issue wasn’t the algorithm itself, but rather, the data used to train it. These training sets lacked sufficient representation of diverse populations, leading to systems that worked well for some but failed others.

A more recent anecdote involves loan approval systems. AI models trained on historical financial data were found to deny loans to minority applicants at higher rates, perpetuating existing disparities in access to credit.

A 2020 study by the National Bureau of Economic Research showcased that algorithms used by some lenders were more likely to flag minority applicants as high-risk due to biased historical data. Some of these datasets trace back to post-Civil War reconstruction and Jim Crow laws, encoding systemic disparities in lending practices. This research highlights the critical need for equitable data approaches going forward.

Fully addressing these issues requires a complete overhaul of how financial datasets are curated and evaluated, but a good start is recognizing why the problem exists in the first place.

Bias in AI systems is often traced back to human nature.

Data reflects our collective knowledge, values, and priorities but also blind spots. Historical records, scientific research, and societal norms have all been shaped by unequal power dynamics. In the 20th century, medical trials often excluded women from studies. The U.S. Food and Drug Administration’s 1977 guideline recommended excluding women of childbearing age from drug trials, which left gaps in understanding how medications affected women.

Similarly, census data historically undercounted minority populations, resulting in systemic inequities in resource allocation and representation. These dynamics often manifest in datasets that overrepresent dominant groups while neglecting marginalized communities. Studies such as the Pew Research Center's findings on economic disparities in digital access highlight how these biases are embedded in data.

Voice recognition software illustrates this issue well. Early systems struggled to recognize accents, regional dialects, or the voices of women and minorities. Research by Stanford University revealed these limitations, emphasizing the need for more inclusive training datasets.

But even knowing that the issue exists, we still haven't done enough to combat it. With the prevalence of Large Language Models (LLMs) trained on broad, open datasets, people are unwittingly bringing bias into their organizations in everything from email drafting to requirements development, HR, customer service, and IT. Basically, unintended bias is a broad-spanning problem, and if you use an LLM, you are very likely directly contributing to the propagation of it.

Overcoming bias requires deliberate choices.

This includes committing to diversity in training datasets and scrutinizing how data is collected, labeled, and processed. Fairness also involves reflecting critically on the purpose and impact of AI applications. Are systems designed to benefit all users, or is convenience for the majority prioritized? Have you studied how the model you want to use was trained in the first place? Have you considered the incorrect correlations that could be drawn from the data that it was trained on?

In Chicago, AI tools flagged certain neighborhoods as high-risk based on historical crime data. This data often reflected over-policing of minority communities rather than actual crime rates, exacerbating inequities in law enforcement practices and leading to more over-policing of minority communities. This is how bias becomes a self-feeding problem if not considered from the onset.

Community advocates have criticized these types of systems for perpetuating harmful stereotypes, leading to increased scrutiny and calls for more transparent and equitable approaches to policing, but the practice has not stopped.

Being informed is not enough. We have to act on it. 

Fairness is essential. Whether you’re a developer, policymaker, or business leader, ask if the data is equitable. Are systems serving everyone? By taking intentional steps now, AI can begin to fulfill its potential as a tool for equity and progress, rather than continue as a tool for spreading stereotypes and amplifying societal bias. 

To be successful in action, we must address equity of data. This means addressing bias at the source by focusing on diverse, representative data rather than merely adjusting model outputs. We have to use our human, innate ability to understand complex data correlation issues, and build both policies and systems to support more just outcomes from AI models.

By prioritizing equity from the ground up, businesses can set a new standard for building AI systems that truly reflect fairness and inspire trust, moving us forward as a society, rather than anchoring us on the biases of our past. 

What can we do about bias in AI?

To address bias in AI, we must take a more rigorous and transparent approach. IBM defines bias in AI as any unfair or prejudiced outcomes that can arise from training data, models, or systems, leading to disproportionate effects on specific groups. Bias can manifest in many ways, from skewed training data that underrepresents certain communities to the model's design itself. The key to tackling this issue lies in identifying and addressing these biases early in the AI development process. IBM's AI Fairness 360 Toolkit is a comprehensive solution that helps organizations assess, detect, and mitigate bias in their AI models, promoting more equitable outcomes.

IBM's Watsonx suite includes built-in tools for ethical AI development, such as automated fairness assessments and continuous monitoring for potential biases during model deployment. With these capabilities, Watsonx allows enterprises to create AI systems that are both innovative and aligned with fairness principles. This tool is a critical asset for businesses that want to ensure their AI models reflect equity and transparency throughout their lifecycle.

For organizations looking to understand the role of bias in their AI systems, C4G Enterprises, in partnership with IBM, is committed to advancing ethical AI solutions. We strongly believe that Watsonx provides the most robust, ethical, and responsible AI tools available today, helping businesses navigate the complexities of bias detection and mitigation. Enterprises that prioritize unbiased AI solutions are better equipped to foster trust, inclusivity, and long-term success in their operations.

If you'd like to learn more about how bias impacts AI and explore strategies for creating more equitable AI systems, we invite you to reach out to the C4G team for a conversation. Together, we can build a future where AI serves everyone, fairly and responsibly.