QCon London 2026: Ethical AI Is an Engineering Problem
At QCon London 2026, a critical discussion emerged surrounding the notion that many risks associated with AI systems are fundamentally engineering challenges rather than merely governance or policy issues. As AI technology becomes more embedded in critical products and decision-making processes, the potential for failures escalates, necessitating that engineers treat the ethical properties of AI systems with the same rigor applied to reliability, performance, or security.
The Importance of Ethical AI
The session opened with a striking example: Robert Williams, a man wrongfully arrested due to a facial recognition system’s misidentification. This incident underscores how algorithmic errors can have dire consequences for individuals and communities alike.
Such failures often stem from technical choices made during development. For instance, training datasets may not accurately represent the affected populations, or model architectures may lack the necessary explainability. Moreover, evaluation pipelines might fail to detect bias before deployment.
Engineering as the Foundation
Rather than viewing these issues as external policy concerns, the discussion emphasized that they originate from the engineering process itself. AI systems encode the values embedded in their design. Decisions regarding data collection, feature engineering, model architecture, and evaluation metrics significantly influence system behavior in production settings.
For example, biased outcomes in areas such as loan approvals, hiring processes, or medical diagnostics can arise from unrepresentative training data or poorly defined optimization objectives. Without explicit checks, models risk reinforcing historical biases present in datasets.
Integrating Ethical Principles
Integrating ethical principles into the AI lifecycle necessitates that engineers ask pertinent questions throughout development, rather than post-deployment. This includes:
- Evaluating datasets for representativeness
- Measuring model behavior across demographic groups
- Ensuring systems remain observable once deployed
Key dimensions guiding AI system design include:
- Fairness: Evaluate model performance across different groups to avoid disadvantaging specific populations.
- Transparency: Enhance interpretability and explainability so stakeholders understand decision-making processes.
- Security: Address emerging concerns like prompt injection and model extraction.
- Sustainability: Consider the computational costs associated with training and deploying large models.
Challenges in Implementation
Organizations frequently struggle to translate high-level ethical concepts into practical engineering workflows. While teams may recognize the importance of fairness or transparency, they often lack clear methods for implementation.
The presentation proposed embedding ethical checks throughout the development lifecycle, including:
- Fairness evaluation during model training
- Explainability analysis before deployment
- Security testing against adversarial attacks
- Monitoring systems for unexpected behavior in production
By incorporating these practices early in system architecture, organizations can mitigate ethical issues discovered after systems are in use.
A Comparison with Other Industries
The talk drew parallels between the current state of AI development and earlier technological transitions in industries like aviation, electricity, and automotive engineering. These industries initially advanced rapidly, often outpacing the safety standards required to govern them. Over time, they developed new engineering practices, standards, and regulatory frameworks to ensure reliability at scale.
Similarly, as AI systems transition from experimental tools to critical infrastructure, it is anticipated that engineering practices will evolve to include safety, reliability, and ethical considerations as core requirements. Software architects and engineering leaders play a pivotal role in shaping these practices, particularly since technology often evolves faster than corresponding regulations.
Conclusion
The presentation concluded by urging developers to treat ethical properties of AI systems as measurable engineering requirements. By embedding fairness evaluation, explainability checks, security testing, and resource efficiency into the development lifecycle, organizations can ensure that AI systems remain both technically robust and socially responsible. As AI continues to be integrated into products, platforms, and infrastructure, the engineering decisions made during development will increasingly shape their societal impact.
