Regulatory Intelligence

EU AI Act Article 12 Explained: What Logging Actually Requires

Following the Council of the EU's final adoption of the Digital Omnibus on 29 June 2026, Article 12 of the EU AI Act becomes enforceable for Annex III high-risk AI systems on 2 December 2027. Article 50 transparency obligations remain live from 2 August 2026. Article 12 requires that every high-risk system technically allow for the automatic recording of events over the system's lifetime. The minimum retention period is six months. Fines for non-compliance with high-risk AI system obligations reach €15 million or 3% of global annual turnover (EU AI Act Article 99(4)).

Most organisations believe they are covered. Most are not. The gap between what teams call an audit trail and what Article 12 actually requires is where regulatory exposure lives.

What Article 12 Actually Says

The full text of Article 12(1) requires that high-risk AI systems be designed and developed with capabilities enabling the automatic recording of events relevant to identifying risks and ensuring human oversight. These logs must be:

• Automatically generated, not manually assembled after the fact
• Retained for a minimum of six months after generation
• Sufficient to enable post-hoc reconstruction of the system's operation
• Available to national competent authorities on request

Article 12(2) goes further, specifying that for systems used in real-time biometric identification, logging must capture start and end time of each use, the reference database used, and the individuals involved.

Why Standard Logs Fail the Article 12 Standard

The most common response to Article 12 is to point to existing application logs. This is insufficient, and auditors are beginning to understand why.

The distinction matters because in a regulatory proceeding, a log that can be modified is not evidence. It is a statement. Statements can be challenged. Evidence stands independently.

The Stabilisation Problem

Producing tamper-evident, cryptographically verifiable audit records at institutional throughput is harder than it appears. The naive implementation, hashing each record individually, does not produce a chain. Any record can be replaced wholesale without breaking adjacent records. Hash chaining solves this, but doing it at the speed and scale that financial services, healthcare, and legal institutions require has proven difficult.

Most implementations that reach proof-of-concept stage fail under institutional load. The hashing overhead compounds. The chain structure creates bottlenecks. Systems that work in testing collapse in production.

What RFC3161 Timestamping Adds

Article 12 requires that logs capture when events occurred. A timestamp written by the application is self-reported. An RFC3161 timestamp is issued by a trusted third-party time authority and is cryptographically bound to the record at the moment of capture.

The difference is significant in a legal context. A self-reported timestamp can be altered. An RFC3161 timestamp cannot, altering the record changes its hash, which breaks both the chain and the timestamp binding simultaneously.

Independent Verification: The Court-Admissibility Test

The final criterion for Article 12 compliance, and the most frequently overlooked, is that evidence must be independently verifiable. This means an auditor, regulator, or court must be able to confirm the integrity of the record without access to the vendor's systems.

If verification requires calling an API, logging into a dashboard, or trusting the vendor's attestation, the evidence has a single point of failure. Vendors can be acquired, shut down, or compelled. The evidence layer must survive the vendor entirely.

KairoNull produces audit packages verifiable with standard OpenSSL tools. The verification command is public. No account required. No API call. The mathematics speak for themselves.

Which AI Systems Are In Scope

Article 12 applies to high-risk AI systems as defined in Annex III of the EU AI Act. This includes:

• Financial services: Credit scoring, fraud detection, AML/KYC automated review
• Employment: CV screening, candidate scoring, promotion recommendation systems
• Healthcare: Clinical decision support, diagnostic imaging analysis, triage systems
• Education: Automated assessment, student evaluation systems
• Law enforcement: Risk assessment tools used in criminal proceedings
• Critical infrastructure: AI systems managing essential services
• Biometric identification: Real-time and post-hoc recognition systems

What Compliant Implementation Looks Like

Article 12 compliance requires four capabilities working together:

1. Capture at generation: The AI input, output, model identifier, and decision context must be recorded at the exact moment the decision is made, not written to a file afterward.
2. Policy evaluation: Each decision must be tested against the organisation's governance policies and the result recorded alongside the decision.
3. Hash-chained ledger: Records must be written to a SHA-256 hash-chained structure where tampering with any record breaks every subsequent hash, making modification immediately detectable.
4. On-demand export: Cryptographically verifiable audit packages must be producible on request, without manual assembly, suitable for regulators, auditors, or courts.

This is precisely what the KairoNull Umbra Trust Protocol delivers. No architectural changes to existing AI systems. No model lock-in. The evidence layer wraps existing AI calls and runs underneath everything already in place.