Introduction

Artificial intelligence (AI) has become a pervasive force, capable of writing emails, diagnosing illnesses, and approving loan applications at speeds and scales beyond human capacity. However, as these systems grow more powerful, they generate complex ethical dilemmas regarding responsibility, harmful content generation, and resource consumption. These are not merely technical engineering problems but profound ethical questions about the kind of world society chooses to create. The current challenge is not whether AI will shape the world, but determining the kind of world it will shape.

Building on solid ground

This framework builds upon the 2018 publication, Ethical by Design: Principles for Good Technology, which established eight core principles for responsible development. While foundational truths such as the imperative to maximise human freedom and design for vulnerable users remain vital, the technological landscape has shifted dramatically. The rise of generative AI has introduced capabilities that even experts struggle to predict or explain. This new reality requires operationalising ethics within systems that learn and adapt independently, often under commercial pressures that threaten to outweigh the imperative for responsible deployment.

What this document offers

The updated framework retains the spirit of the original principles while addressing unique AI challenges. Key considerations include ensuring AI respects human dignity when influencing behaviour, achieving algorithmic explainability, and mitigating the risks of synthetic content and misinformation. It also addresses the potential for AI to amplify societal biases and the significant environmental costs associated with training and deploying systems at scale. Each principle includes a philosophical rationale and concrete practices for practical implementation within organisations.

Principles for responsible AI technology

The document outlines eight principles, starting with Ought before can, which dictates that the purpose of an AI system must be evaluated for legitimacy before development begins. For example, the release of deepfake technology illustrates the consequences of prioritizing technical capability over ethical review, leading to foreseeable harms such as identity fraud. Recommended practices include conducting ethical impact assessments and establishing an AI ethics review body to evaluate necessity and risk.

Non-instrumentalism requires that AI respect the intrinsic dignity of individuals and ecosystems, ensuring people are never treated merely as data points or ‘means’ to an end. This principle highlights the exploitative working conditions of data labellers in low-income countries who process disturbing material for a few dollars per hour. Organisations are urged to prohibit non-permissible uses and ensure fully informed consent for all data collection and usage.

Self-determination ensures AI enables users with transparent choices and explainable outputs. In automated credit approval systems, applicants often face life-changing financial decisions without the ability to interrogation or contest the algorithmic outcome. Practices for implementation include providing opt-in mechanisms, avoiding dark patterns, and establishing redress mechanisms for AI-driven outcomes.

Responsibility mandates that developers anticipate multi-use risks, such as surveillance or misinformation. The rapid misuse of large language models for propaganda was foreseeable but inadequately planned for. Suggested actions include implementing risk governance for the full AI lifecycle and maintaining transparent provenance of models and datasets.

Net benefit requires that AI deliver measurable social and environmental gains while actively mitigating harms. The energy demands of AI are significant; training a single large-scale language model can produce carbon emissions equivalent to several long-haul flights. Ethical obligations include optimising models for energy efficiency through pruning or quantisation and using harm metrics to monitor impact.

Fairness demands equitable performance across demographics. Facial recognition tools have historically shown error rates for darker-skinned women that exceed those for lighter-skinned men by more than 30%. Organisations should audit algorithms for bias, use diverse training datasets, and build inclusive design teams to ensure equitable access.

Accessibility focuses on universal design and low-resource settings. While healthcare diagnostic tools have the potential to assist under-resourced regions, they are often trained on Western data, making them less accurate for the populations they were intended to serve. Recommended practices include designing lightweight models and providing offline or low-bandwidth functionality.

Purpose requires AI systems to have clear ethical goals aligned with human flourishing. Social media recommendation algorithms offer a cautionary example, as many drifted from their stated purpose of connection to an effective purpose of maximising time on the platform, contributing to mental health deterioration. Developers must clearly communicate intended use and use minimalist design to reduce unnecessary computation.

Conclusion

These principles are presented as a continuous discipline rather than a final destination. Adhering to them requires the courage to slow down, ask uncomfortable questions, and collaborate with affected communities. Ultimately, AI is shaped by the choices of its creators, and this framework serves as an invitation to align those choices with values that support a just and sustainable future.