Trust & Accountability¶
What people should be able to demand from AI—and how those commitments could be checked.
When people rely on AI, they need to know what it is allowed to do, what supports its claims, who answers for failures, and how to challenge a result. Assurance means examining evidence for a specific claim about a system, within a stated scope and with clear limits. A promise or published policy alone does not show that a control works.
Our AI Charter connects three kinds of work. They have different purposes and should not be treated as interchangeable.
| Part | What it answers | Where the work stands |
|---|---|---|
| Charter commitments | What should people be able to demand from AI systems and those responsible for them? | Public draft, open for discussion. |
| Evaluating AI answers | Do factual answers match their cited sources, handle uncertainty and support correction and challenge? | Draft method. No pilot evaluation has been run or is scheduled. |
| Future certification | How could independent assessment, oversight and withdrawal of claims work? | Proposed arrangement. No operating certification scheme, Trust Mark or Charter-certified systems. |
Start with the question you need answered¶
- Using or affected by AI? Read what people should be able to demand, followed by the five public obligations.
- Evaluating a factual AI system? Start with the method summary. The full draft protocol explains sampling, measurement, limitations and the proposed report. A narrow evaluation cannot establish general safety, legal compliance or Charter alignment.
- Responsible for organisational AI use? The institutional crosswalk proposes how evaluation findings could inform procurement, policy and oversight.
- Examining independence and accountability? The future certification proposal discusses separate roles, conflicts of interest, access to evidence and consequences for unsupported claims.
How the documents fit together¶
The Charter's five obligations express the public commitments. Its eight operational duties describe responsibilities that could be assessed. Modules examine particular subjects, such as factual grounding; assurance depth distinguishes documentation from observed evidence, checked implementation and tested effectiveness. An unassessed material area limits the claim that can be made.
The structure map explains these relationships; worked assurance cases show how obligations connect to duties and evidence. Both support discussion of the draft rather than establish an operating assessment scheme.
Research and supporting evidence¶
The research menu separates rights and risk foundations from proposed controls, record custody and correction mechanisms. These pages retain their own source dates, assumptions and open questions.
Evidence-Gated Agents is the current development focus: research and prototype work on permission, evidence and controlled release of consequential answers and actions. Its proposed user workflow describes decisions during use; those workflow steps are distinct from the Charter's public obligations. EGA does not establish whole-Charter conformity. See the current project status and what the research needs to demonstrate.
To contribute, raise a specific question or proposed correction through the contribution guide.