Ethical AI: Tackling bias and addressing challenges

Can technology, which belongs to the seemingly impartial and objective realm, be tainted by bias? The resounding truth is both uncomfortable and illuminating yes, it can.

In the vast unknown of our digital landscape, biases have found a way to penetrate even the most advanced technological innovations. Within this realm, artificial intelligence (AI) stands as a shining example of susceptibility. While AI models rely heavily on data to shape their responses– what if the data itself is biased?

A July 2022 report by the UNDP- observed that in the healthcare sector people's lives, privacy, and equality were at grave risk due to inadequate and lack of diversity in data sets. The report emphasized the significant impact of algorithmic bias in other sectors like retail, unorganized sectors, and financial services too. 

There is a myriad of examples of AI bias in the public domain. Be it the COMPAS algorithm's discrimination against people of color, the PredPol algorithm's biased predictions in minority neighborhoods, or a major e-commerce company’s recruitment engine failure to treat women candidates fairly compared to their male counterparts; all these instances highlight the pervasive nature of bias in AI systems.  

In the Indian context, this issue takes on far greater relevance as our diverse populations, languages, colors, cultures, and traditions make the threat of AI bias even more pronounced. So, as we propel ourselves forward into the era of AI, it becomes our responsibility to ensure that the future of technology is fair, inclusive, and beneficial for all. We must confront and address the challenges posed by biased AI algorithms. 

Transparency and Explainability 

To effectively combat bias, organizations must prioritize transparency and explainability in their AI systems. We will eventually have to do away with AI algorithms that remain inscrutable. Opening up the black box of AI algorithms and providing insights into decision-making processes can help build trust among users, stakeholders, and the wider public. Explainable AI can empower individual users to grasp and question the decision-making processes, ensuring that the AI systems we create align with our values and uphold principles of equity. As we navigate the paradigm of rapid technological advancement, it's crucial that we champion transparency, unveiling the secrets of AI and working towards a future where biases are confronted and dismantled, algorithm by algorithm.  

Addressing bias at every stage through self-regulation 

Self-regulation is a critical aspect of ensuring ethical AI development and deployment. One such example is that of documenting datasets. As suggested by Gebru et al. in “Datasheets for Datasets”, documenting key information on purpose, composition, collection process, use, distribution, and maintenance of data gives dataset creators the opportunity to reflect on their building process. Thereby enabling them to refine questions and workflow through feedback from researchers, practitioners, and legal experts and provide dataset consumers with the information needed to make informed decisions. Such self-regulatory mechanisms from organizations can increase transparency, mitigate biases, and improve reproducibility. 

Rather mitigating bias in AI algorithms warrants self-regulation at every stage of development. Data pre-processing is the starting point, where organizations can meticulously examine and address biases within training datasets. As the model training phase commences, evaluating AI models using fairness metrics tools like "AI Fairness 360" aid in further analyzing datasets for potential bias. This enables the mitigation of biases during data pre-processing, model training, and prediction stages. By combining self-regulation and a robust bias mitigation strategy, organizations can build AI systems that are both transparent and accountable. 

Regulation and Compliance 

Addressing the ethical challenges of AI begs for synergy among organizations, policymakers, researchers, and society at large. By establishing cross-industry partnerships and knowledge-sharing platforms, responsible AI practices can be promoted, and best practices can be adopted across sectors. Regulatory frameworks are crucial in establishing guidelines and standards, requiring organizations to proactively comply with evolving regulations. To exemplify, the Digital India Act (2023) sets forth guardrails on high-risk AI systems, imposing strict guidelines to protect individuals' privacy, data security, and fundamental rights. It instills accountability by holding organizations responsible for any negative impacts of their AI systems. It emphasizes continuous monitoring and evaluation of AI models, ensuring that they perform fairly and equitably across diverse user groups. Such policies help align organizations and industries and inform the development of ethical guidelines as well as help mitigate potential biases. 

While some of these challenges can be met, others don’t always have a straightforward solution. The AI realm of understanding is limited by data and lacks the intuition of human experience. This is to say that AI systems, at this moment in time, lack the power of judgment based on evolutionary inferences. Alignment with human ethical values, therefore, becomes a crucial challenge that cannot be overlooked. 

AI Alignment and Human Ethical Values 

AI systems are designed to perform tasks efficiently, but they lack human understanding. This gives rise to the AI alignment problem: how can we ensure that AI systems align with human moral values? It becomes complex when multiple values need to be prioritized. For instance, in autonomous vehicles, AI alignment involves finding the right balance between safety and efficiency. If these values conflict, it is impossible to maximize both. AI alignment becomes crucial when systems operate at a scale where human evaluation is impractical. The alignment problem involves the technical aspect of encoding values reliably and the normative aspect of determining which moral principles should guide AI.  

In a paper by Betty Li Hou and Brian Patrick Green from the Markkula Institute of Applied Ethics, a solution is presented. This solution outlines a four-stage approach aimed at promoting cooperation and attaining positive results in aligning AI systems. By aligning values across individuals, organizations, nations, and the global level, favorable outcomes in AI can be achieved. However, in a world where individual values come in the way of organizational values like revenue generation, the complexity of the problem increases even further. For example, most social media platform algorithms work around engagement as the metric of success and end up producing filter bubbles where individual consumers interact with un-varying content that limits both their understanding and worldview making them suspectable to manipulation. This problem of considerable heft requires strong-willed ethical practices at all four levels of functioning. But organizations will have to act as a fulcrum between individuals, national interests, and global values in an interconnected world. By nurturing a culture of ethical practices, organizations can begin to navigate the challenges of AI alignment while upholding human values and societal responsibilities. 

Leading the Way to Ethical AI 

So, as we venture further into the AI-driven era, responsible innovation must be at the forefront of every organization's AI strategy. By prioritizing ethical AI practices, organizations can mitigate reputational risks, regulatory concerns and seize strategic opportunities as well as better align AI systems. Moreover, Ethical AI is not just a compliance requirement; it is an opportunity to shape a future that prioritizes fairness and inclusivity. By actively addressing bias, promoting transparency, and fostering collaboration, we can harness AI's potential to transform industries, empower individuals, and drive positive change globally. 

India must navigate these biases particularly well to facilitate inclusivity and ensure that AI technologies promote fairness and equal access for all. It will require a collaborative effort involving policymakers, industry leaders, researchers, and civil society to establish guidelines, prioritize diversity in data collection, and integrate ethical practices throughout the AI development process. 

Authored By: Navin Dhananjaya, Chief Solutions Officer, Merkle