1-2 | Purpose, audience and legal footing of the Guidelines | 0 |
3 | AI in the public service: task types, benefits and risks | 0 |
4.1a | Assess fundamental rights impact before development and let affected people give feedback | 0 |
4.1b | Equip users to understand and contest AI systems, with opt-out from solely automated decisions where warranted | 0 |
4.1c | Build human oversight into every AI process and keep final decisions and reviews with people | 0 |
4.2a | Hold internally built and third-party AI systems to the body's IT security policies and national guidance | 0 |
4.2b | Prepare and prove a fallback plan, and complete a full risk assessment, before deployment | 0 |
4.2c | Test and monitor across the lifecycle for accuracy, reliability and reproducibility, and escalate shortfalls | 0 |
4.3a | Classify the dataset first: personal data brings the GDPR into play, anonymised or synthetic data may not | 0 |
4.3b | Settle the legal basis for personal data processing, consulting the Data Protection Commission where required | 0 |
4.3c | Respect privacy and data protection across the lifecycle, run DPIAs and obtain vendor compliance confirmations | 0 |
4.3d | Check data quality and integrity before training and follow data management policies and documentation | 0 |
4.3e | Restrict access to data to those with a legitimate need | 0 |
4.4a | Keep traceable documentation of model development | 0 |
4.4b | Explain the AI system's data and decision-making in terms end-users understand | 0 |
4.4c | Tell people when they are dealing with AI or AI-assisted outputs and offer alternative service options | 0 |
4.5a | Detect and mitigate unfair bias before and after deployment, with diverse perspectives and inclusive design | 0 |
4.5b | Design AI systems for accessibility and universal design | 0 |
4.5c | Consult affected stakeholders throughout the lifecycle and keep seeking feedback after deployment | 0 |
4.6a | Deploy AI in an environmentally conscious way, mitigating its environmental harm | 0 |
4.6b | Weigh and monitor social impact, and never deploy AI that risks the democratic process | 0 |
4.7a | Document who is responsible at each lifecycle stage and name an ethics contact in user guides | 0 |
4.7b | Enable auditability, including independent audit of systems affecting fundamental rights or safety | 0 |
4.7c | Hold risk assessments and safe-usage policies, protect those who raise concerns, and tell affected people how to seek redress | 0 |
5.1 | Decide with a cross-functional team whether AI is the best solution | 0 |
5.2 | Choose the type of AI solution that fits the use | 0 |
5.3 | Evaluate build, buy or free-of-charge options and review benefits after the project | 0 |
5.3.1 | Adopt GenAI only on an approved business case, with access restricted by default until policies and awareness are in place | 0 |
5.3.3 | Analyse value for money from the start, cost AI across its lifecycle and plan evaluation early | 0 |
5.4 | Bring diverse perspectives into the AI team from the start | 0 |
6 | Complete the Responsible AI Canvas with a cross-functional team at the planning stage | 0 |
7.2.1 | Planning and design: embed the seven principles in the project baseline | 0 |
7.2.2 | Data collection and processing: collect lawful, representative, secure and documented data | 0 |
7.2.3 | Model building: train securely, explainably and with continuous bias testing | 0 |
7.2.4 | Verification and validation: test the model against real-world conditions before release | 0 |
7.2.5 | Deployment: implement oversight, inform users, monitor drift and keep the fallback ready | 0 |
7.2.6 | Operation and monitoring: keep the system under human control and under review | 0 |
7.2.7 | Retirement, decommissioning or refresh: disengage safely, handle data lawfully and document the decision | 0 |
8a | Assure the quality of data relied on for AI: the right data and accurate data, with senior independent quality assurance | 0 |
8b | Do not use free public GenAI tools for public service work and never enter non-public information into them | 0 |
8c | GenAI end-users: disclose AI-generated content, protect data, validate outputs, keep human judgement and check for bias | 0 |
9 | Use cases, related strategies and regulation | 0 |