Building Ethical AI

Artificial intelligence is rapidly becoming part of everyday life. As AI systems increasingly influence decisions that affect people's lives, building them responsibly is no longer optional.

Building ethical AI is an engineering and governance challenge grounded in established research, internationally recognized frameworks, and continuous oversight. Rather than relying on good intentions alone, organizations can draw on evidence-based guidance to design AI systems that are fair, transparent, accountable, and safe.

There's no single definition of ethical AI, but the most influential reference point is the OECD AI Principles: the first intergovernmental standard for AI, adopted in 2019 and updated in 2024. They are composed of five values-based principles and five recommendations that provide practical and flexible guidance for policymakers and AI actors, and they shape what customers, investors, and partners expect, while helping organizations mitigate reputational and legal risk.

The five core principles for trustworthy AI:

• Inclusive growth, sustainable development, and well-being: AI should benefit the planet and the people, such as augmenting human capabilities.
• Human-centred values and fairness: AI should respect human rights, diversity, democratic values, and the rule of law while helping prevent unfair outcomes.
• Transparency and explainability: People affected by AI should receive meaningful information about how systems function and how decisions are made.
• Robustness, security, and safety: AI systems should remain reliable, secure, and resilient throughout their lifecycle.
• Accountability: Organizations and individuals developing and deploying AI must remain responsible for its behaviour and outcomes.

These principles have shaped AI governance worldwide. By 2023, governments across more than 70 jurisdictions had introduced over 1,000 AI policy initiatives aligned with the OECD framework.

In the United States, the NIST AI Risk Management Framework (AI RMF) has become another influential reference. Although voluntary, it provides organizations with a structured process for identifying, assessing, managing, and monitoring AI risks throughout a system's lifecycle. In 2024, NIST published additional guidance specifically for generative AI, highlighting risks such as bias, privacy concerns, cybersecurity vulnerabilities, intellectual property issues, and the production of misleading content.

Together, these frameworks emphasize that responsible AI is not achieved through a single feature or policy. It requires governance, documentation, ongoing evaluation, and clear accountability from design through deployment.

Generative AI, including large language models (LLMs), differs from traditional software because it generates new content rather than following predefined rules. This enables powerful applications such as drafting clinical documentation, summarizing information, or providing conversational assistance.

However, these capabilities introduce risks that require careful management.

Research shows that generative AI systems can reproduce or amplify societal biases present in their training data. Bias may affect outputs related to gender, race, culture, age, disability, or other characteristics, potentially leading to unfair or inequitable outcomes. In healthcare, these biases can contribute to disparities in the quality of information or recommendations provided.

Another distinctive challenge is that large language models can produce confidently inaccurate outputs, often referred to as hallucinations. These responses may sound plausible while containing factual errors, fabricated information, or misleading conclusions. In high-impact settings, even occasional inaccuracies can have significant consequences if they are accepted without verification.

The research therefore emphasizes that organizations should not assume these risks disappear once a model is deployed. Bias, accuracy, and performance require continuous monitoring, evaluation, and human oversight.

Research consistently concludes that AI should not replace human judgment, particularly in sectors where mistakes can directly affect people's health, safety, or well-being.

A 2024 systematic review examining the ethics of large language models in medicine and healthcare identified recurring concerns across hundreds of studies, including fairness, transparency, privacy, potential harm, and inaccurate outputs. One of its central conclusions was that ethical AI requires meaningful human oversight, tailored to the specific context in which the technology is used.

The appropriate level of oversight depends on the potential consequences of an error. In high-impact environments such as healthcare, human professionals remain responsible for interpreting AI-generated information, validating important outputs, and making final decisions.

This reflects a broader principle found across international AI governance frameworks: accountability cannot be delegated to an AI system. Organizations must establish clear processes for human review, define responsibilities, and ensure that people remain accountable for decisions that affect others.

Trustworthy AI is not created by adding a single safety feature or publishing an ethics statement. It emerges from a disciplined process built on research, governance, and continuous improvement.

A science-backed approach includes:

• Following internationally recognized frameworks such as the OECD AI Principles and the NIST AI Risk Management Framework.
• Assessing and documenting risks throughout the AI lifecycle.
• Designing systems that prioritize fairness, transparency, security, and accountability.
• Continuously monitoring models for bias, errors, and changing performance after deployment.
• Maintaining meaningful human oversight, especially in high-impact applications such as healthcare.

To be ethical, AI requires documented risk assessment, meaningful transparency, ongoing monitoring after deployment, and clear lines of human oversight and accountability.

Ultimately, ethical AI is about earning trust. By grounding development in established research and internationally recognized best practices, organizations can build systems that are not only technically capable but also responsible, transparent, and worthy of the confidence people place in them.