Responsible AI Practices

This content was generated with the assistance of AI and should be verified against the official documentation of each tool and framework before being relied on in production.

This section’s bibliography lists the reference material consulted while preparing these pages.

Written on 2026-10-01 against NIST AI 600-1 (July 2024), Regulation (EU) 2024/1689 and WCAG 2.2; see the version baseline. The practices are the author’s synthesis; no fairness or accessibility measurements of a live DocsAssistant are reported.

Security keeps an attacker from misusing the system; responsible AI asks whether the system treats legitimate users fairly, honestly and safely. The two overlap, since the NIST categories human-AI configuration and harmful bias are as much product concerns as security ones (see NIST AI RMF).

The RAISE framework

Steve Wilson’s The Developer’s Playbook for Large Language Model Security proposes RAISE, a framework for building AI responsibly and securely, as a checklist of engineering practices that a team applies through the project rather than a single gate before release. This page does not reproduce the book’s text; it paraphrases the idea and ties it to the controls already in this section: narrow the system’s domain, treat data and models as supply-chain inputs, keep humans able to intervene, test adversarially, and monitor after release. Consult the book for the framework’s own wording and ordering.

Fairness and bias testing

  • Define who the users are, including languages, proficiency and assistive-technology users.

  • Build evaluation sets with the same questions phrased by different groups and in different languages, and compare answer quality, refusal rate and tone (see Building Eval Sets).

  • Compare per-group results rather than only the average: \( \Delta = \max_g m_g - \min_g m_g \), where \( m_g \) is a quality metric for group \( g \); set a tolerated gap and fail the release check when it is exceeded.

  • Review retrieval sources for coverage gaps, since missing documentation produces worse answers for the affected users.

  • Re-run on every model or prompt change in Evaluation in CI.

Transparency and disclosure

Users should know they are talking to an AI, what it can and cannot do, and where answers come from. This site shows an example: every page in this section opens with a disclaimer that the content was generated with AI assistance and should be verified against official documentation. For the assistant itself, display an AI label, cite sources, state the knowledge date, and say how to reach a person. These also support the Article 50 duties on the EU AI Act page.

Human oversight and escalation

Albada’s Building Applications with AI Agents (Chapter 13) treats human oversight as a design problem: decide which actions need approval, make the approval informative, and give the human a real alternative. For the DocsAssistant:

  • read-only answers need no approval; write or external actions pass an approval gate (Human in the Loop);

  • low-confidence or sensitive questions escalate to a person, with the conversation context passed along;

  • reviewers are shown the evidence and the proposed action, not just an approve button, to limit automation bias;

  • escalation outcomes feed the evaluation set (Online Evaluation and Feedback).

See also Agent Guardrails and Safety.

Accessibility

Chat interfaces fail accessibility tests often: live regions that are not announced, focus lost after a response, answers with only visual structure. Meet WCAG 2.2 for the widget (keyboard operation, announced updates, sufficient contrast, plain language), offer a non-chat route to the same information, and include assistive-technology users in testing. Voice channels need transcripts and alternatives to speech.

Final checklist

  • Purpose, users and out-of-scope uses are written down.

  • AI disclosure and source citations are visible to users.

  • Fairness checks run per group and language, with a tolerated gap.

  • Approval gates and escalation paths exist and are tested.

  • Accessibility was tested with assistive technology.

  • Security controls and red-team results are current (AI Red Teaming).

  • The risk mapping to NIST and the EU tier assessment are up to date.

  • An incident and rollback procedure exists (LLMSecOps).

Where to go next

References