5.1 Responsible AI Principles & Human Accountability

5.1 Responsible AI Principles and Human Accountability

Core principle: Responsible AI is a socio-technical discipline. It combines human rights, fairness, privacy, transparency, robustness, security, safety, accountability, and societal impact with evidence and governance appropriate to the use.

---

OECD AI Principles

The OECD principles were adopted in 2019 and updated in May 2024. They are recommendations, not a binding global statute. Their five values-based principles are:

1. Inclusive growth, sustainable development, and well-being — pursue beneficial outcomes for people, planet, and society.
2. Respect for the rule of law, human rights, and democratic values, including fairness and privacy — include safeguards and the ability for affected people to challenge outputs where appropriate.
3. Transparency and explainability — provide meaningful information about AI systems and enable people affected by an output to understand and challenge it as appropriate.
4. Robustness, security, and safety — make systems function appropriately through their life cycle, manage risks, and support safe override, repair, or decommissioning.
5. Accountability — organizations and people responsible for AI should be accountable for its proper functioning and for respecting the principles.

EU Trustworthy AI Guidelines

The European Commission's High-Level Expert Group frames trustworthy AI as lawful, ethical, and robust. Its seven requirements are:

• human agency and oversight;
• technical robustness and safety;
• privacy and data governance;
• transparency;
• diversity, non-discrimination, and fairness;
• societal and environmental well-being;
• accountability.

These guidelines are not identical to the EU AI Act and should not be presented as direct legal advice. They offer a useful assessment structure across the system life cycle.

Human Oversight Patterns

Human-in-the-loop places a person in a decision path. Human-on-the-loop supervises automation and can intervene. Human-out-of-the-loop allows operation without case-by-case review. None is automatically right or wrong. Choose according to consequences, reversibility, error detectability, scale, affected rights, and available expertise.

A formatter that converts public text to a validated schema may run without manual approval for every record. A model recommending employment, credit, medical, safety, or production-security decisions needs much stronger governance and may be inappropriate. Human review is not meaningful if the reviewer lacks time, evidence, authority, or a usable way to override the system.

Scrum accountabilities remain with people, but Scrum does not prescribe a universal manual AI workflow. Developers can use automated tests and deployment controls while remaining accountable for quality. The Product Owner may use forecasts while remaining accountable for value and effective Product Backlog management.

Meaningful Transparency

Transparency is contextual. A stakeholder may need to know that AI materially influenced a recommendation, which data and source version informed it, important limitations, and how to correct or appeal. Developers may need logs, evaluation results, model and prompt versions, and failure traces. Publishing confidential prompts or private personal data is not responsible transparency.

Explainability does not require revealing hidden chain-of-thought. Prefer concise reasons, evidence citations, feature or rule information where valid, known limitations, and outcome-level tests. A Product Owner should be able to explain the product evidence and trade-offs behind ordering decisions rather than saying “the model decided.”

Misinformation and Deepfakes

Generative systems can create convincing false text, audio, images, and video. Deepfakes may harm people through impersonation, non-consensual imagery, fraud, harassment, or false evidence. Controls include consent, provenance metadata where supported, visible labeling when appropriate, trusted-channel verification, current source checks, restricted generation and distribution, incident response, and correction.

No detector is perfect. Do not use a detector score alone to accuse someone of deception. Verify origin and context through independent evidence.

When AI Gets It Wrong

Plan for errors before release: define prohibited outcomes, representative and adversarial tests, monitoring, an accessible reporting channel, investigation ownership, rollback or safe shutdown, notification and correction, and learning. A post-incident Retrospective examines sources, controls, incentives, and the Definition of Done without transferring blame to a model.

Scrum Mapping

• Transparency: make material AI use, evidence, uncertainty, and limits visible to the right people.
• Inspection: test product behavior and verify sources with methods suited to risk.
• Adaptation: change data, prompts, model, permissions, workflow, or product when evidence shows harm.
• Values: use commitment, focus, openness, respect, and courage to resist convenient but unsafe shortcuts.

The legal obligations of vendors, employers, and practitioners depend on jurisdiction, contract, and facts. Scrum accountabilities do not replace that analysis.

Information Integrity in Team Practice

Before sharing generated media or summaries, record origin, obtain permission where people are depicted or imitated, and verify factual claims through trusted channels. Make correction easy. The possibility of synthetic content should not justify dismissing authentic evidence without investigation.