AI in Practice: How to Create Fair and Trustworthy AI Systems

June 16, 2026

Artificial intelligence is increasingly finding its way into healthcare, public services, human resources processes, educational institutions, and many other areas of daily life. As a result, questions of fairness, transparency, and accountability are coming into sharper focus. In an exclusive interview with Bayern Innovativ, Alexandra Wudel, founder of FemAI and “AI Person of the Year 2024,” explains what “Responsible AI” means in practice, why biases in AI systems are far more than just a technical problem, and how organizations can develop and deploy AI solutions that are more inclusive, trustworthy, and beneficial to society.

Ms. Wudel, would you like to briefly introduce yourself? What inspired you to found FemAI?

Alexandra Wudel: My name is Alexandra Wudel, and before I founded FemAI, I worked at the German Bundestag, where I helped implement the administration’s digital strategy. Prior to that, my research focused on analyzing the impact of AI on society, with a particular emphasis on documenting the patterns of social discrimination that are reproduced and reinforced by AI systems. As long as AI isn’t developed, managed, and monitored responsibly, it cannot be designed to be fair. This realization was the catalyst for founding FemAI in 2023—we’ve been on the market as an AI startup since the launch of ChatGPT. For FemAI, fairness is not a “nice-to-have,” but a prerequisite for high-quality AI products—especially in sensitive areas such as healthcare.

What gaps in the AI ecosystem do you hope to fill with FemAI?

Alexandra Wudel: Based on our research, we develop consulting offerings and services that address specific market needs. At the same time, we consistently pursue the vision that AI should serve people, especially women and other underrepresented groups.
Initially, we worked primarily with governments and public institutions, for example on regulatory issues surrounding the ethical boundaries of autonomous weapons systems. Since then, our focus has shifted significantly toward AI applications that are already available on the market. Technological development is advancing so rapidly that regulation alone is not sufficient to effectively protect human rights. This is evident in the developments surrounding the EU AI Act. Therefore, those who make decisions about the use of AI currently bear the greatest responsibility for ensuring an ethically sound approach. But how can we assess which solutions are truly trustworthy when transparency and certification have so far been limited?
FemAI positions itself between the developers of AI systems and their users. Our work focuses on groups that are often underrepresented, such as women, children, people with disabilities, and other marginalized groups. Our work encompasses research, professional development, and consulting. In the long term, we are also exploring approaches to the certification and systematic evaluation of AI systems. Because, let’s be honest, many of the AI tools available today fall far short of the trustworthiness requirements that are particularly necessary in sensitive areas.

The term “feminist AI” may lead to misconceptions or misunderstandings among people outside the field. How can feminist AI be explained in simple terms?

Alexandra Wudel: Feminist AI is a subset of Responsible AI. It brings a specific perspective to the field by utilizing methods from intersectional feminist research. In doing so, it helps identify, understand, and reduce biases in AI systems. At its core, feminist AI asks questions such as:

• Who benefits from an AI system?
• Who might be disadvantaged by it?
• Who was taken into account during development, data selection, testing, and implementation—and who was not?

From our perspective, AI has enormous potential. It can improve healthcare, facilitate access to knowledge, automate routine tasks, and create more space for interpersonal relationships and care. At the same time, real-world experience shows that AI can reinforce existing inequalities—especially when systems are based on biased datasets or developed without sufficient oversight.
For us, feminist AI is therefore both a vision and a practical approach. The goal is to develop systems that are not only efficient or economically successful but also make a positive contribution to society.

When we talk about fair AI, what does fairness actually mean, and who decides whether a system is fair?

Alexandra Wudel: Currently, fairness is often defined by those who develop AI systems or make strategic product decisions. In many cases, there is no independent review to assess what data was used, how biased a system is, or whether it works equally reliably for different population groups.
From our perspective, fairness has three key dimensions: data, design, and context.
Data fairness means that training and test data are sufficiently representative. A system should function reliably for different population groups.
Design fairness means integrating mechanisms that can reveal and address performance differences between various groups—even when the available data is not perfect.
Contextual fairness considers the broader context: Who is developing the system, and who is affected by it? Does the team have the necessary understanding of societal and regulatory risks?
A fair AI system should ensure comparable accuracy, robustness, and reliability for all affected groups. This is particularly crucial in healthcare. Here, differences can have direct implications for diagnoses, treatments, and people’s trust.

Can you give a specific example of how AI systems can unintentionally discriminate against women or other marginalized groups?

Alexandra Wudel: A simple but very illustrative example is AI image generation. For example, if you ask an image generation system to create an image of a doctor, you often get a depiction of a white, middle-aged man with Western features. This doesn’t necessarily happen every time and also depends on the language and prompt, but it remains a common pattern. This doesn’t mean that the system consciously favors men; rather, it reflects the data used to train the model. Historically, many datasets have contained significantly more depictions of white men in certain professional roles than of women or people from the Global South.
If, for example, a hospital uses such an AI-generated image for a job posting, male applicants may feel more drawn to it. Women or members of marginalized groups, on the other hand, might feel less represented. In this way, seemingly insignificant design decisions can reinforce existing inequalities. And this example is representative of a broader problem: Biases can occur in images, language, recommendation systems, clinical decision support systems, or recruiting tools. Therefore, the first step is to become aware of this issue, and the second is to actively challenge and question biased default settings.
When used correctly, AI can even help create more inclusive outcomes—for example, by specifically highlighting diversity or supporting inclusive communication. This means we can promote societal goals rather than reproducing existing inequalities.

In light of the EU AI Act and the growing requirements for responsible AI: In your opinion, what should companies consider before implementing an AI system?

Alexandra Wudel: Companies should address early on at what point AI governance becomes relevant. Those who wait until after a system has been implemented to start addressing this issue usually face significantly higher costs and greater effort. Responsible AI should not be added at the end of the process, but rather considered from the very first idea. The EU AI Act provides an important framework, but it is not the only relevant requirement, as there are additional horizontal regulations and cybersecurity requirements, such as those under the NIS 2 Directive. In the healthcare sector, additional industry-specific regulations also apply.
It can be particularly difficult for startups to meet all requirements at once. That is why it is important to discuss AI governance early on. Even a single-day workshop can help identify risks, responsibilities, and regulatory requirements. It is crucial to determine early on how issues such as governance, fairness, transparency, and compliance will be addressed throughout the product development process.

Looking to the future: What would need to change so that AI systems become not only more powerful, but also more equitable, inclusive, and trustworthy?

Alexandra Wudel: In my view, the most important factor is transparency. Providers should openly communicate in which application contexts and for which population groups their systems function reliably, and they should also highlight where there is still room for improvement. This builds trust, which is particularly important in the healthcare sector. Hospitals, doctors, patients, and policymakers must be able to determine whether an AI system performs equally reliably for different population groups. If there are differences in performance, these should be openly identified and addressed. This would create positive incentives in the market. Companies that offer responsibly developed AI could leverage trustworthiness as a competitive advantage. At the same time, pressure on other market participants to further develop their systems would increase.
Ultimately, AI systems should not be evaluated solely on the basis of their performance, as reliability, fairness, inclusivity, and accountability are just as important. Especially in sensitive areas such as healthcare, trust is not an option but a fundamental prerequisite for a technology to be accepted and to create genuine social value.