1-12 | Purpose, policy basis and characterisation of AI, software and security in scope | 0 |
100-101 | Appoint a RAISO who assures responsible AI across all projects in scope and sets the culture | 0 |
102 | On transfer of AI, establish accountability in the receiving organisation and have its RAISO accept the risk | 0 |
103-104 | Put context-appropriate roles in place; every project identifies its RAISO and medium or high risk projects inform them | 0 |
105-106 | Provide an Ethics Manager and independent ethics assurance for large, high-ethics-risk or production-line AI work | 0 |
107-108 | Consult the DAU before external communications on ethically sensitive work and legal advisers on legal aspects | 0 |
109-111 | Non-sovereign AI and allied interoperability | 0 |
112-114 | Satisfy the RAISO that shared, exported or imported AI and data meet ASR policy and respect host-nation considerations | 0 |
115-118 | Human/AI teaming and Human Centred Design | 0 |
119-121 | Apply a recognised, resourced Human Centred Design approach across the lifecycle, and HFI under JSP 912 for military capability | 0 |
122-123 | Identify users and their needs, involve them in development and testing including after retraining, and apply human sciences to the design | 0 |
125 | Analyse the allocation of functions between humans and AI across all modes and autonomy levels | 0 |
127-128 | Assess and plan for workforce implications from the outset and keep re-evaluating | 0 |
129-131 | Train users to understand behaviour, limits and calibrated trust, include edge cases, and assure collaborative learning teams case by case | 0 |
13-16 | Apply the directive from the outset of capability development, across the whole life cycle and all stakeholders, in proportion to maturity | 0 |
132-134 | Lifecycle models: DevOps, MLOps and the directive's lifecycle-agnostic stance | 0 |
135-137 | Plan the AI lifecycle with version-controlled plans that embed the ethical principles | 0 |
140-142 | State performance requirements, run a hazard analysis including data failure modes and feed new hazards to system safety | 0 |
143-148 | Keep requirements traceable, demonstrate learned behaviours meet high-level intent, specify the operating context and data, and state how each requirement is verified | 0 |
150 | Make the AI architecture traceable to requirements and protective against identified failure modes | 0 |
151-153 | Justify and document algorithm design choices and performance measures, and optimise for the operational context | 0 |
154-155 | Maintain traceability in implementation and justify and check every framework and library | 0 |
156-160 | Demonstrate data is relevant, complete, balanced and accurate, and keep training, test and verification data separate and representative | 0 |
161-163 | Assure data provenance, keep all influencing data under configuration management and validate synthetic data | 0 |
165-168 | Document each model's assumptions, dependencies, errors and ODD; check this before re-use; design for interpretability; classify models with their data | 0 |
17-18 | Tailor only through a recorded risk assessment managed under Section 3 | 0 |
170-173 | Verify and validate across the ODD and its boundaries, demonstrate safe behaviour outside it, regression test updates and evidence sufficiency | 0 |
174-177 | Confirm performance in the real operating context, analyse environment differences, re-demonstrate after substantive host changes and define the approved ODD | 0 |
178-181 | Write proportionate Security Operating Procedures, monitor environment and performance, add expert checks for risky decisions and control modification | 0 |
183-186 | Leadership maintains AI competence: analyse roles, train and supervise staff, and bring in external support where in-house competence falls short | 0 |
187-189 | Apply JSP 940 quality policy to AI with planning and assurance scaled to risk and maturity | 0 |
19-22 | TLB Holders and Chief Executives direct, resource and delegate to competent people | 0 |
190-194 | Evidence-based safety management: analyse unique AI safety risks early, include them in safety cases, meet Def Stan 00-055 and 00-056 and manage risk under JSP 892 at AI pace | 0 |
195-197 | Secure by Design with continuous assurance: analyse AI-specific security risks such as data poisoning and adversarial inputs from the earliest opportunity | 0 |
198-200 | Require externally acquired AI to meet the directive with equal confidence, and require suppliers to understand the Defence context and demonstrate competence | 0 |
201-202 | Require arguments for third-party software and assured, documented provenance of all training and test data | 0 |
203-204 | Contract for through-life access to data, algorithms and models with clear IP ownership, and assess export control restrictions on freedom of action | 0 |
205-208 | AI assurance as an overlay on traditional software assurance, assurance risk and the system context | 0 |
206 | Do not adopt AI where assurance risk is intolerable; otherwise assure in proportion to the risk of incorrect outputs | 0 |
209 | Assure behaviour on excursions from the ODD, including adversarial causes, with failure mode behaviour chosen and confidence matched to risk | 0 |
23-26 | Apply Def Stan 00-055 and 00-056, review current standards and prefer civil standards with a military delta | 0 |
28 | Record and agree the risk balance for choosing AI over existing technology for major and higher risk components | 0 |
29-37 | Robotic and autonomous systems and digital systems: definitions and the continuum of autonomy | 0 |
35-36 | Account for the level of autonomy and test the military delta of dual-use systems | 0 |
38-39 | Define the ODD of AI in digital systems and identify every in-scope digital system as AI-based | 0 |
40-41 | State AI use in products and official statistics and give risk owners the assurance information | 0 |
42-43 | Identify alternative information sources and the influence of AI advice in system risk analysis | 0 |
44-49 | The ASR perspectives and the five MOD AI Ethical Principles | 0 |
47-48 | Take legal advice from the start and communicate the controls that keep AI use lawful | 0 |
50-51 | Consider the ethical principles as early as practicable and signpost redress | 0 |
52-55 | Human-centricity: identify everyone affected and assess impact across all modes using Human Security factors | 0 |
56-58 | Responsibility: name responsible persons through the governance chain and articulate human involvement for every context of use | 0 |
61-65 | Understanding: give each stakeholder group the level of understanding it needs, within security limits and in the system-of-systems context | 0 |
67 | Set an open, inclusive culture for raising AI issues and ethical risk assessment | 0 |
69-71 | Bias and harm mitigation: analyse data, algorithms and models for unwanted bias, and monitor and mitigate harms including from intentional bias | 0 |
75 | Define and communicate the operating context and bound the system's behaviour | 0 |
76-78 | Set and demonstrate performance targets for the operating context, reconsider inherited targets, and define responses to out-of-design inputs | 0 |
79-81 | Analyse security threats across the development environment and supply and data chains, install guards and minimise the attack surface | 0 |
82 | Analyse failure modes, including inter-system emergent effects and cascading failures | 0 |
83-85 | Research ethics: comply with JSP 536, declare AI used in research and apply the principles at every readiness level | 0 |
86-88 | Run an AI ethical risk assessment at the outset and on material change, review it at senior level and rate each use case on residual risk to set the approvals route | 0 |
89-90 | Refer special-attention and high-ethical-risk applications to the DAU | 0 |
91 | Halt development or deployment safely when ethical risks become unacceptable | 0 |
92-93 | Manage ethical risk throughout the lifecycle and have the RAISO manage trade-offs with operational effectiveness | 0 |
95 | Develop and communicate evidence of ethical development and use, as transparently as security allows | 0 |
96-98 | Governance of AI builds on existing software governance | 0 |
99 | TLB Executive Boards give 2PUS an annual Statement of AI Ethical Assurance backed by auditable evidence | 0 |