# What is auditable AI?

> Auditable AI is an approach to building AI systems whose conclusions can be inspected, challenged and reproduced after the fact. It requires traceability from each decision back through the reasoning, the rules applied and the interpretation made, to the authoritative source that justifies them. In regulated work it complements explainability: not only why a model answered as it did, but which provision, which rule and which evidence the answer rests on.

Canonical: https://sireto.com/auditable-ai

## Source-to-decision traceability

NormLogic is designed around a chain of evidence with five links: source, interpretation, rule, reasoning, decision. Each link records what it was derived from. A decision can be followed back to the reasoning step, the formal rule, the interpretation a person approved, and the provision in the regulation. The chain is recorded using provenance standards such as W3C PROV so it can be exchanged and verified.

## Trustworthy AI in practice

European expectations of trustworthy AI translate into concrete properties a system either has or lacks.

- Transparency: the structure, rules and sources are visible, not hidden in weights and prompts.
- Traceability: every output links to its inputs and justification.
- Human oversight: interpretations are reviewed and approved by people, and that approval is recorded.
- Robustness: deterministic rules give the same answer for the same inputs, and are tested with cases.
- Documentation: the model of the regulation is itself the documentation of what the system does.
- AI assurance: auditors can reproduce a conclusion from the recorded chain.

## AI governance with structured constraints

Autonomous agents increasingly act in real systems. Giving an agent documents does not make it act within the rules. Auditable AI, as NormLogic approaches it, gives agents explicit regulatory constraints linked to authoritative sources, so an agent can ask what it is allowed and required to do, which conditions apply, and which provision created the obligation, and so its actions can be reviewed against the same constraints afterwards.

## Frequently asked questions

### Is auditable AI the same as explainable AI?

They overlap. Explainable AI focuses on why a model produced an output. Auditable AI focuses on whether the whole chain from source to decision can be inspected and reproduced, including the human interpretations and formal rules around the model.

### Does auditable AI require avoiding large language models?

No. It requires placing them where probabilistic interpretation is appropriate, recording what they proposed, having people approve interpretations, and executing decisions with deterministic rules. The model assists; the structure and rules are what gets audited.

### Does NormLogic make an AI system compliant with the EU AI Act?

No. NormLogic supports the traceability, documentation and human oversight that trustworthy AI calls for, and can model the Act’s requirements for compliance checks. Whether a system complies is a determination for the organisation and its advisers.

## Related

- [What is automated regulatory compliance?](https://sireto.com/automated-regulatory-compliance)
- [What is machine-executable regulation?](https://sireto.com/machine-executable-regulation)
- [What is Regulatory Engineering?](https://sireto.com/regulatory-engineering)
- [NormLogic](https://sireto.com/normlogic)