First legally binding international AI treaty. EU effective September 1, 2026. Treaty obligations mapped to SWT3 witness procedures.
Audience: International compliance officers, legal counsel advising on treaty obligations, EU AI governance teams, organizations operating across CoE member states, and public sector AI program managers.
Status: EU ratified May 15, 2026 (EU Parliament approved March 11). Effective for EU: September 1, 2026 (19 days). US and UK signed September 2024 but have not ratified. 46 Council of Europe member states. Applies primarily to public sector AI; private sector coverage determined by each party.
The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy, and the Rule of Law (CETS No. 225) is the first legally binding international AI treaty. It establishes obligations focused on protecting human rights, democratic processes, and rule of law in the context of AI systems.
Critical distinction: This is a treaty, not a regulation. It creates obligations for states to implement, not direct obligations on companies. But those state implementations will create compliance requirements for organizations operating within each ratifying party's jurisdiction. Organizations should treat the Convention's principles as the minimum standard that national implementing legislation will enforce.
| Convention Article | Obligation | SWT3 Procedure | Evidence Produced |
|---|---|---|---|
| Art. 4 -- Transparency | Ensure transparency about AI system use | AI-TRANS.1 |
Disclosure type, recipient, timestamp |
| Art. 5 -- Accountability | Establish accountability for AI outcomes | AI-AUDIT.1 + AI-INF.1 |
Audit integrity hash + inference provenance chain |
| Art. 6 -- Non-discrimination | Prevent discriminatory AI outcomes | AI-FAIR.1 |
Bias disparity ratio, group count, max disparity |
| Art. 7 -- Privacy | Protect personal data in AI systems | AI-CONSENT.1 |
Consent type, withdrawal status |
| Art. 8 -- Safety / Risk | Assess and mitigate AI risks | AI-SAFE.1 + AI-DRIFT.1 |
Safe state attestation + drift detection threshold |
| Art. 9 -- Human oversight | Ensure meaningful human control | AI-HITL.1 |
Override decision, rationale, authority level |
| Art. 10 -- Reliability | Ensure AI system reliability | AI-PERF.1 + AI-ROBUST.1 |
Performance metrics + adversarial robustness score |
| Art. 14 -- Remedies | Provide effective remedies for AI harms | AI-INCIDENT.1 + AI-VIO.1 |
Incident severity classification + violation reporting |
Article 4 of the Convention requires transparency about when AI systems are used and how they affect people. This is the treaty's most broadly applicable obligation -- every ratifying party must ensure that individuals interacting with AI systems can understand the nature and consequences of that interaction.
SWT3 witnesses transparency disclosures via AI-TRANS.1 and explanation quality via AI-EXPL.1. Each anchor records the disclosure type, the recipient category, and the timestamp of the event, creating an immutable evidence trail that national implementing authorities can audit.
AI-TRANS.1 anchors prove disclosure events occurred at specific points in time. AI-EXPL.1 anchors document explanation method and quality metrics. Together they satisfy both the "notice" and "understanding" dimensions of Art. 4 transparency.
Article 6 requires parties to adopt measures to prevent discriminatory AI outcomes. This maps directly to fairness monitoring -- organizations must be able to demonstrate that their AI systems do not produce biased results across protected characteristics.
AI-FAIR.1 measures bias disparity ratios across protected groups, recording group count, max disparity value, and the evaluation methodology used. This creates quantitative evidence that fairness obligations are being monitored and enforced.
Look for AI-FAIR.1 anchors with max_disparity values. Ratios exceeding defined thresholds should correlate with documented remediation actions. Absence of AI-FAIR.1 anchors for high-risk AI systems is itself a compliance gap under Art. 6.
Article 5 accountability requires that AI decisions be attributable -- someone must be responsible, and the decision path must be traceable. This is the treaty's anchor principle: without accountability, none of the other protections are enforceable.
AI-AUDIT.1 proves audit log integrity through cryptographic hashing. AI-INF.1 provides per-inference provenance, linking each AI output to its model, inputs, and processing context.
The combination of AI-AUDIT.1 and AI-INF.1 proves both that the evidence chain is intact (no tampering) and that individual decisions can be traced back to their origin. This dual-anchor pattern is the strongest Art. 5 evidence package.
Article 9 requires meaningful human control over AI systems, particularly for decisions with significant impact on individuals. Article 8 requires risk assessment and mitigation. These two obligations are tightly coupled -- human oversight is itself a safety mechanism.
AI-HITL.1 records human-in-the-loop override decisions including the rationale and the authority level of the person intervening. AI-SAFE.1 anchors prove that risk mitigation capabilities exist and are operational.
Verify that AI-HITL.1 anchors exist for high-impact decisions -- their absence indicates that human oversight mechanisms are either missing or not being exercised. AI-SAFE.1 anchors prove the system can be brought to a safe state when required by Art. 8 risk assessments.
The United States signed the Convention in September 2024 but has not ratified it. Signing creates a political signal and a good-faith obligation not to undermine the treaty's purpose, but it does not create binding legal requirements.
| Examiner Question | Where to Find Evidence |
|---|---|
| How do you prove AI transparency disclosures occurred? | AI-TRANS.1 anchors in the sovereign witness ledger. Each anchor records disclosure type, recipient, and timestamp. |
| Can you demonstrate non-discriminatory AI outcomes? | AI-FAIR.1 anchors with quantitative disparity ratios. Historical trend via the AI Witness dashboard. |
| How is accountability maintained for AI decisions? | AI-AUDIT.1 (log integrity) + AI-INF.1 (inference provenance). Full chain from input to output. |
| What human oversight mechanisms exist? | AI-HITL.1 anchors documenting override decisions, rationale, and authority level of the intervening person. |
| How do you assess and mitigate AI risks? | AI-SAFE.1 (safe state capability) + AI-DRIFT.1 (performance drift detection). Risk assessment cadence in scan history. |
| What remediation process exists for AI harms? | AI-INCIDENT.1 (severity classification) + AI-VIO.1 (violation reporting). Incident response timeline in ledger. |
pip install swt3-aifrom swt3_ai import SWT3Witnesswitness = SWT3Witness(api_key="YOUR_KEY", endpoint="https://sovereign.tenova.io")anchor = witness.witness_inference(model_id="gpt-4o", prompt="...", response="...")print(anchor.fingerprint)
npm install @tenova/swt3-aiimport { SWT3Witness } from "@tenova/swt3-ai";const witness = new SWT3Witness({ apiKey: "YOUR_KEY", endpoint: "https://sovereign.tenova.io" });const anchor = await witness.witnessInference({ modelId: "gpt-4o", prompt: "...", response: "..." });console.log(anchor.fingerprint);
Full SDK documentation: sovereign.tenova.io/docs. SDKs available in Python, TypeScript, Rust, C#, Ruby, Swift, Kotlin, and MCP.