compliance · the Dubai AI Principles
Wytness and the Dubai AI Principles
Guideline-by-guideline evidence to sit beside your Dubai AI ethics self-assessment.
Dubai's AI Principles and Ethics Guidelines (Smart Dubai, January 2019) set four principles — ethics, security, humanity and inclusiveness — and seventeen guidelines on making AI fair, accountable, transparent and explainable, with a self-assessment tool that scores a system against them. They are non-binding and the self-assessment is not audited. Wytness evidences 8 items that rest on a record of what your agents did and how they were governed, numbered as the guidelines number them so the pack can sit beside your self-assessment answers; the other 17 are listed below with the reason. This is a first mapping; the pack evidences your governance record, not the models themselves.
Sources: Smart Dubai AI Ethics Principles & Guidelines (January 2019, PDF) · AI System Ethics Self-Assessment Tool
the artefact
What's in the pack
A Dubai AI Principles pack contains the agent and tool inventories, evidence and a measured metric for each mapped guideline or principle, signed event samples, the LLM provider and model mix, and the list of items it does not evidence, with the reason for each.
Every claim in the pack is traceable: each control cites the events behind it, and each event carries the signature your own key produced. The pack is sealed, so an auditor can confirm it reached them unaltered.
A real sealed pack is published — nothing here is a mock-up. Download the specimen and check it yourself on the verify page, in your browser, without uploading it. It is a SOC 2 (Trust Services Criteria) pack — the seal and the per-event signatures work identically whichever framework you generate.
// What the pack asserts about itself:
// control_mapping[] — 8 criteria, each carrying
// its evidence and a metric measured over your period:
…// + 4 more, in the table below
// Every control cites events an auditor can re-verify:
the mapping
What we map to what
| Item | Guideline or principle | What Wytness provides |
|---|---|---|
| 1.1.3 | Significant decisions informed by AI should be fair | Guideline 1.1.3.1 suggests formal procedures such as discrimination impact assessments. The register's AI risk assessment includes a bias and fairness section, and the responsible AI impact assessment covers affected people and potential harms; each is answered by the customer against a register row, scored on submit, and its gaps become remediations with an owner and a due date, with each submission and each remediation's creation, status change and deletion written to the audit log. Wytness records the procedure and its outcome; it does not measure the fairness of a model or a decision. |
| 1.2.1 | Accountability for outcomes should not lie with the system itself | Each register row carries an assigned operator (a named user of the organisation) and an approval record (approved_by, approved_at); every event records human_operator_id, so each action is attributable to an agent and an accountable person; remediations from assessments carry a named owner and a due date, and the owner is notified. Four-tier role-based access control (viewer / member / admin / owner) is enforced server-side on every route. |
| 1.2.2 | Identify and mitigate significant risks | Guideline 1.2.2.3 suggests internal risk assessments and 1.2.2.7 continual performance monitoring. Each register row carries a risk classification (minimal / limited / high / prohibited, with a justification required for high and prohibited) and can hold scored risk and impact assessments with tracked remediations; 7 anomaly-detection rule families run against the live event stream with severity-classified alerts; the pack reports the LLM providers and models each agent called. Data accuracy, model tuning and the research basis of the system remain the customer's. |
| 1.2.7 | Operators should understand the AI systems they use | The register documents each agent's intended use, description, asset type, version and lineage beside its risk classification and its declared tools and permissions; every row passes through Approval before Deployment in the lifecycle Discovered → Intake → Development → Validation → Approval → Deployment → Retirement, with Intake, Development and Validation available for the organisation to require before it; Preflight compares the recorded tape against the declared set. The judgement that a system suits its use case is the customer's; Wytness records what was declared and what the agent did. |
| 1.3.1 | Traceability for significant decisions | Guideline 1.3.1.3 asks for the decision journey of an outcome to be logged. Every tool call, resource access and inter-agent message is logged with its inputs, outputs, tool parameters and the agent's reasoning_summary where the runtime supplies one, hash-chained (tamper-evident) and signed per event with the customer's Ed25519 key, and grouped by session so the steps behind an outcome read in order; register changes are logged with actor and fields changed. Events are retained for seven years by a storage lifecycle policy on Wytness-hosted storage, or under the customer's own policy on bring-your-own-storage. Training-data provenance, model design and code-change history (1.3.1.2) are outside Wytness's view. |
| 1.4.2 | Means to request explanations for significant decisions | When a person asks about a decision, the customer's owners and admins can find the events that concern them by subject search (email, operator id or free text) and read the recorded inputs, outputs, tool parameters and reasoning summary behind it. The channel through which people make the request, and the explanation given to them, are the customer's. |
| Security 1 | AI systems will be safe, secure and controllable by humans | Wytness evidences the security of the record, not of the model: the hash chain is verified link by link per agent, per-event Ed25519 signatures are checked against the customer's registered public keys, and a three-way ingest / query-store / archive reconciliation runs on a schedule. The register's Retirement stage records when a system is withdrawn. Overriding or reversing an agent's decisions belongs to the customer's runtime; the kill action records the request and does not stop the agent. |
| Inclusiveness 4 | We will respect people's privacy | Where the customer enables PII keys, personal data in payloads is replaced with HMAC pseudonyms in the customer's runtime before egress and the token map is encrypted to the customer's X25519 public key, so Wytness stores pseudonyms it cannot reverse. Every event records inputs_source and inputs_classification (the provenance and sensitivity class of the data acted on), the register tags each agent with the data classes it accesses, and an event classified restricted raises a restricted_data anomaly. Surveillance limits and data minimisation in the agent itself are the customer's. |
the boundary
What Wytness does not cover
| Item | Guideline or principle | Why the pack does not evidence it |
|---|---|---|
| 1.1.1 | Data ingested should be representative of the affected population | Training and evaluation data are outside Wytness's view. |
| 1.1.2 | Decision-making processes should not introduce bias | Wytness does not test models or decision processes for bias; the customer's own assessment is recorded under 1.1.3. |
| 1.1.4 | AI systems accessible and usable fairly across user groups | Accessibility of the customer's product is outside the record of agent activity. |
| 1.1.5 | Effect of diversity on development and deployment | Team composition and consultation with AI subjects are outside the product. |
| 1.2.3 | External audit of AI systems informing critical decisions | Suspended in the guidelines themselves until an external auditing mechanism is established; an Evidence Pack can be handed to an auditor, but no Dubai audit scheme exists to evidence against. |
| 1.2.4 | AI subjects can challenge significant automated decisions and opt out | The challenge, review and opt-out procedure is the customer's; Wytness has no channel for AI subjects. |
| 1.2.5 | No value judgements on people's behalf without consent | A design property of the customer's system; the record cannot show it. |
| 1.2.6 | Significant-decision systems developed by diverse teams | Team composition and expert consultation are outside the product. |
| 1.3.2 | People informed of their interaction with AI systems | Disclosure to people happens in the customer's product, not in the governance record. |
| 1.4.1 | High-level explanation of how the AI system works | Explanations to AI subjects are the customer's; the register's intended-use description is an internal record, not a public explanation. |
| 1.4.3 | Explanations accessible, free of charge and user-friendly | The explanation service is the customer's. |
| Security 2 | AI systems should not be able to autonomously hurt, destroy or deceive humans | A design and policy commitment, including on autonomous weapons; the record of agent activity cannot evidence it. |
| Humanity 1 | We will give AI systems human values and make them beneficial to society | A societal commitment, outside what a governance record can show. |
| Humanity 2 | We will plan for a future in which AI systems become increasingly intelligent | Long-term and general-intelligence risk planning is a policy matter outside the product. |
| Inclusiveness 1 | We will govern AI as a global effort | A government commitment to international cooperation and standards. |
| Inclusiveness 2 | We will share the benefits of AI throughout society | Employment, training and education aims for society, outside the product. |
| Inclusiveness 3 | We will promote human values, freedom and dignity | A societal commitment; human sign-off on individual agents is recorded under guideline 1.2.1, but the principle itself is not evidenced. |
- —Neither the guidelines nor Wytness certify anything. An Evidence Pack is a mapping with evidence; your self-assessment score is yours to calculate in Dubai's own tool.
- —Fairness of data and models (guidelines 1.1.1 and 1.1.2) is outside Wytness's view. We record the assessments you make; we do not test the model.
- —Challenge, opt-out and explanation channels for the people your AI affects (1.2.4, 1.4.1, 1.4.3) live in your product. Wytness gives you the record to answer from.