What comes to mind when you hear “AI Fairness and Causality”? For many, these terms represent the future of artificial intelligence, one where technology serves humanity ethically and equitably. Leading this transformative field is Ruta Binkyte, a postdoctoral researcher at the Helmholtz Centre for Information Security (CISPA).

Ruta’s innovative research in Machine Learning Ethics focuses on making AI systems fair, ethical, and trustworthy. Her work uniquely blends cultural anthropology, history, and data science, allowing her to address the complex interdisciplinary challenges of AI ethics from a holistic perspective. This approach has earned her recognition in top journals and conferences worldwide.

Ruta Binkyte

“AI ethics cannot be dictated solely by politicians or embedded by engineers into AI systems; it is something that we, as a society, must collectively discover and develop.”

“AI ethics cannot be dictated solely by politicians or embedded by engineers into AI systems; it is something that we, as a society, must collectively discover and develop.”

In this interview, Ruta shares her expertise on fairness in machine learning, the critical importance of causal analysis, and the broader ethical considerations of AI.

Q & A

Q. For our readers who might be new to this field, can you briefly explain the concept of fairness in machine learning and why it’s important?

A. Fairness in ML has two aspects. One is procedural and requires that everyone should be treated the same despite their ethnicity, gender, disability, and other sensitive features as defined by the non-discrimination law. Another requirement is to ensure that no community is affected disparately by the AI decisions. Sometimes those two requirements are hard to satisfy simultaneously, nevertheless, people in the AI Fairness community work to ensure that the benefits of technology are as equally distributed as possible. This also means the acknowledgment of the historical and psychological biases that are reflected in the training data and that technology has the potential to exacerbate them at an unprecedented scale.

Q. What does it mean to use causality in assessing fairness in machine learning?

A. The field of statistical causality aims to distinguish the causal relationships from the spurious correlations from the data, that do not necessarily come from ideal conditions. Such ideal conditions would a randomized controlled trials studies, where all the potential non-causal factors are controlled. In ML most often observational data is used, which means, that it can be biased in multiple ways. Causality helps to disentangle different pathways that connect sensitive attributes and the decision. Sometimes, in the situation, that is called Simpson’s paradox, the causal evaluation can show the opposite result to the purely statistical and find discrimination that would be otherwise undetected.

Q. Why is having a causal model important for ensuring fairness in AI?

A. Building a causal model helps to be more explicit about the relationships in the data. This helps not only to evaluate and mitigate bias but also to leave space for future scrutiny.

Q. With your unique blend of humanities and data science expertise, how do you approach the study of fairness and causality differently from others in the field?

A. I try to turn to history and sociology to inform about the possible disparities in the data. For example, cultural specifics, and historical circumstances can have a huge impact on how different communities are represented in the data. If some groups are not in the data set or are misrepresented because of stereotypes and discrimination of the past, the ML model will not be able to learn accurate predictions for them. Another thing that I try to keep in mind is cultural diversity in the understanding of fairness. Different communities and different domains of application might require different approach to fairness. One person or group cannot make those decisions. Being explicit about the ethical implications of the approaches and tools you use empowers end users to evaluate and scrutinize those decisions according to their own value system.

Image Source: Envato

Q. What inspired you to focus your research on this field?

A. Re-iterating the idea on being as explicit and transparent as possible, I think that causality gives good tools to achieve this goal. Causality also allows to connect ML with scientific discovery and gives a powerful tool to better understand the world.

Q. In your research, you reviewed different methods for discovering causal relationships. How do these methods vary and why is this variation significant?

A. The methods vary from purely statistical, that aims to discover the causal relationships from the data to expert-informed models. The combination of prior knowledge and data-driven methods is probably the reliable approach.

Q. How do you test or validate the effectiveness of various causal discovery methods?

A. Data-driven methods are prone to discovering spurious links and directions, that contradict common sense. For example, education is the cause of age. The results should always be critically assessed and if possible, combined with expert knowledge.

Q. Could you give an example of how small differences in understanding cause and effect might impact decisions about fairness in automated systems?

A. One example would be a direct link between the sensitive attribute and the decision. The presence or absence of one arrow can determine if there is direct discrimination.

Reveal Quote

“Fairness in machine learning requires that everyone should be treated the same despite their ethnicity, gender, disability, and other sensitive features, ensuring no community is affected disparately by AI decisions.”

Reveal Quote

“Working in AI ethics allows me to connect my current interests in AI with my prior experience in cultural anthropology and history. I hope it makes me more sensitive to the implicit value systems that are in-built into the models.”

Q. What are some of the practical difficulties you’ve encountered when applying theories of causality to real-world data?

A. One of the main limitations of causality is the availability of a reliable causal graph. The assumption of prior knowledge is both the strength and limitation of the causal approach. On one hand, it limits the applications and full automation of the process. On the other hand, it allows incorporating the expert knowledge and brings transparency to decision-making.

Q. How does your research contribute to the broader effort to create ethical and unbiased artificial intelligence systems?

A. I am interested in approaches and tools that allow to connect existing causal knowledge that comes from science and experiments data with ML and data-driven solutions. I believe it can bring the best of two worlds together and contribute to better AI.

Q. Can you share some insights on how fairness is measured in AI systems, and what are the challenges in ensuring it?

A. There are multiple formal fairness notions, that correspond to different understandings of fairness. For example, one approach requires an equal distribution of positive decisions across the sensitive attribute. Another notion of fairness, allows for disparity, given it is based on some quality or need, for example, priority medical care is given more often to certain ethnic groups if people of that group tend to have poorer health. The challenge is in determining which approach is more suitable for which situation. More often, than not, different fairness notions contradict each other and cannot be applied simultaneously.

Image Source: Envato

Q. How would you describe the impact of AI in everyday life and its ethical implications?

A. AI is being used in multiple decision-making such as loan granting, health care, job hiring. Biased decisions can have huge implications on who gets the benefit and who is discriminated. More subtly, AI is also a gatekeeper to the information online and on social media. That means that implicit values of the systems have the power to ultimately shape the views and opinions of the society.

Q. What motivated you to delve into AI ethics, and how does your diverse academic background shape your approach in this field?

A. Working in AI ethics allows me to connect my current interests in AI with my prior experience in cultural anthropology and history. I hope it makes me more sensitive to the implicit value systems that are in-built into the models. I also turn to history and sociology to look for possible sources of bias in the data.

Q.  As a closing note, what is one key takeaway you’d like our readers to remember about the importance of ethics in AI?

A. The key takeaway is that AI ethics cannot be dictated solely by politicians or embedded by engineers into AI systems. Instead, it is something that we, as a society, must collectively discover and develop. This includes contributions from people of all backgrounds and disciplines as we integrate this powerful and pervasive technology. Achieving this requires significant responsibility and mutual respect, along with a willingness to learn from one another and recognize the limits of our knowledge and understanding.

About The Author

Branislava Lovre

Branislava is a Media Expert, Journalist, and AI Ethicist who leverages her expansive knowledge and experience across various media outlets and digital landscapes.

Branislava Lovre

Branislava is a Media Expert, Journalist, and AI Ethicist who leverages her expansive knowledge and experience across various media outlets and digital landscapes.