A rights-based AI treaty asks who can challenge the system

The Council of Europe framework shifts attention from model capability to dignity, remedies and democratic accountability.

Most AI demonstrations begin with what a system can do. A rights-based assessment begins with what its use could do to a person. Can they understand an important decision? Can they correct the information behind it? Does the system treat groups differently, or interfere with participation in public life? These questions remain relevant even when a model performs well on a technical benchmark.

The Council of Europe’s [framework convention overview](https://www.coe.int/en/web/artificial-intelligence/the-framework-convention-on-artificial-intelligence) describes a treaty opened for signature on 5 September 2024, focused on human rights, democracy and the rule of law across the AI lifecycle. The [Treaty Office record](https://www.coe.int/en-GB/web/conventions/full-list2?module=treaty-detail&treatynum=225) is the appropriate place to check treaty details. The [explanatory report](https://rm.coe.int/1680afae67) provides context but expressly is not an authoritative interpretation. This historical analysis does not assert that all countries have ratified the convention or implemented identical duties.

The Council of Europe is a separate institution from the European Union. Its framework and the EU’s AI rules therefore should not be treated as interchangeable. A treaty signature, ratification and domestic implementation are also different steps. Organisations seeking a legal determination must check their jurisdiction and the relevant current instruments with qualified advisers. This article is educational analysis, not legal advice.

An original illustrative public agency uses AI to identify applications that need additional review. It records the purpose, the data categories and the office responsible for the decision. A rights-focused assessment then asks what happens when a person’s name, address history or language produces a misleading flag. The harm may be delay, repeated requests for documents or a burden that falls unevenly on particular communities.

Testing average accuracy will miss some of that experience. The agency could compare delays and correction rates across relevant groups, while handling any demographic information with appropriate safeguards. It could ask service users whether the explanation is understandable and whether the challenge route is accessible. This proposed exercise examines the pathway through which an error becomes a public consequence.

A meaningful remedy requires operational capacity. If a reviewer can only repeat the model’s output, the review adds little. The person conducting it needs the relevant evidence and authority to correct the record or reconsider the decision. In the illustrative agency, a challenge about inaccurate data follows a different path from a disagreement about policy. The interface should help the person explain which problem they experienced.

Democratic accountability adds questions beyond an individual service. A government may use AI to summarise public submissions or prioritise messages from residents. The process should not erase dissent or treat polished, repeated text as more representative than short accounts of lived experience. Analysts need recoverable source material and a documented method for checking whose views were omitted.

The framework’s scope has boundaries, including its treatment of national defence and research activities. Those details matter when comparing policy instruments. A broad statement that the treaty regulates every AI application would conceal the distinctions in the official text. The safe editorial approach is to name the relevant instrument and explain the specific context, rather than extrapolating a universal rule from its headline.

The potential benefit is a governance conversation that remains connected to people when technical performance improves. The failure mode is a formal assessment that describes risks without changing the service. An organisation can publish a careful document and still leave users unable to correct a harmful result. Evidence of recourse needs an actual process, tested with realistic cases, alongside the written policy.

For a team designing an AI service, the practical questions are direct: who is affected, what evidence supports the decision, who can intervene and how will changes be tracked? The 2024 convention provides an international rights-focused reference for that work. It should be read with its scope and national context intact. Policy ambition becomes useful when it is translated into decisions and remedies that people can reach.

A useful management review can ask for one completed challenge case, with private details removed. Trace the initial concern, the reviewer’s evidence and the correction. That exercise can reveal where a promised remedy depends on information or staff capacity that the service does not actually have.