Article 14 Human Oversight: engineering checklist for high-risk AI systems
As of 18 November 2026 · checked against Regulation (EU) 2024/1689 and Regulation (EU) 2026/1744
Information only, not legal advice. This page is an engineering reference, not a compliance opinion. Confirm any compliance decision with qualified counsel.
Article 14 of the EU AI Act applies to every high-risk AI system under Article 6, whether it is high-risk through an Annex I product or an Annex III use case. It does not list specific UI controls or thresholds: it lists outcomes the human-oversight design must achieve. This page maps those outcomes to concrete engineering work: where in the system you intervene, what you log, and what you hand to the deployer.
What Article 14 actually requires
- Provider must design and develop the system so it can be effectively overseen by natural persons during the period it is in use.
- Oversight measures must be commensurate with the risks, level of autonomy, and context of use.
- Persons assigned oversight must be able to: understand capacities and limitations; remain aware of automation bias; correctly interpret output; decide not to use the output or override it; intervene or interrupt operation by a stop button or comparable procedure.
- For systems under Annex III(1)(a) (remote biometric identification), no action or decision may be taken by the deployer on the basis of an identification unless it has been separately verified and confirmed by at least two natural persons with the necessary competence, training, and authority (Article 14(5)). This does not apply for law enforcement, migration, border control or asylum where Union or national law considers it disproportionate.
The provider builds the oversight surface; the deployer staffs and uses it. Article 26(2) puts the deployer obligation on top.
Implementation options by intervention point
Three points in the inference path are typically wired for oversight:
| Intervention point | Mechanism | Latency budget |
|---|---|---|
| Before invocation | Human-in-the-loop gate; system serves a recommendation only. | Seconds to minutes: synchronous |
| During invocation | Confidence-based routing, soft holds, dual-control verification (Annex III(1)(a)). | Sub-second: synchronous |
| After invocation | Override, recall, rollback; statistical sampling for review. | Minutes to days: asynchronous |
Engineering checklist
- Stop control. A reachable, documented stop or graceful-degrade path. Must work even if the model service is unhealthy.
- Override path. Authorised persons can replace the system output with a human decision; the override is the binding decision of record.
- Confidence and uncertainty surfaced. The output exposes a meaningful confidence signal and the operator-facing UI does not hide it.
- Limitations brief. A short, current document describing scope, known failure modes, drift signals, and out-of-distribution behaviour. Versioned with the model.
- Automation-bias countermeasure. At least one design feature actively works against rubber-stamping: e.g. randomised second-opinion prompts, periodic blind reviews, or required justification on high-impact decisions.
- Dual-control for biometric identification. Two distinct authenticated identities required to commit a downstream action. Recital 73 notes that it can be enough for the separate verifications to be recorded automatically in the system logs.
- Training and competence record. The system records which oversight role each operator holds and the date of last training acknowledgement.
- Deployer hand-off pack. Documented under Article 13: instructions for use that explain how to operate the oversight surface in practice.
Logging requirements
Article 12 (automatic logging) and Article 14 reinforce each other. The oversight log is part of the Article 12 record. The Regulation does not prescribe a schema for it; Article 12(3) sets minimum fields only for remote biometric identification systems. An example of what such a record could hold:
{
"event_id": "ovr_01HZY3K2T9P4MQR7N5B1A8",
"occurred_at": "2026-04-27T14:22:18.412Z",
"system_id": "hrs-credit-scoring-v3.2.1",
"session_id": "sess_8f2c...",
"operator_id": "u_4419",
"operator_role": "reviewer",
"model_version": "credit-risk-3.2.1",
"input_hash": "sha256:9f8c...",
"system_output": { "decision": "deny", "score": 0.83 },
"confidence": 0.83,
"operator_action": "override",
"operator_decision": { "decision": "approve" },
"justification": "Applicant submitted updated income documentation post-scoring.",
"dual_control_required": false,
"stop_invoked": false
} Retention follows Article 19 for providers and Article 26(6) for deployers: at least six months unless Union or national law requires otherwise. The Regulation does not require a storage technique. Append-only storage with a cryptographic chain is our suggestion for making later tampering visible.
Common engineering anti-patterns
- "Override" buttons that route back through the same model. That is not an override, it is a re-score.
- Confidence shown only in admin dashboards, not in the operator UI where decisions happen.
- Stop control gated behind a separate role nobody on shift holds.
- Dual-control implemented as two clicks by the same logged-in user.
- Limitations brief written once at certification and never updated: drift makes it stale within a release or two.
- Oversight events logged to the same database the model writes its own outputs to, with no separation of duties.
What to track on the regulatory side
Article 14 itself is stable, but its operational expectations are shaped by harmonised standards (CEN/CENELEC JTC 21), AI Office guidance, and the date from which the high-risk rules apply. The Digital Omnibus moved that date to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products, see the deadline timeline. AI Act Radar emits events on the human-oversight topic for:
- Harmonised standards cited in the Official Journal that touch oversight, transparency, or logging. As of this page none is cited for Article 14; the only cited AI Act standard is EN 18286:2026 on quality management.
- AI Office Q&A entries clarifying Article 14 expectations.
- National supervisory authority guidance on operator competence and training.
- Notified body common positions on conformity assessment for oversight design.
Sources
- Regulation (EU) 2024/1689, Articles 12, 13, 14, 19, 26 and recital 73: EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus), Article 113 dates: EUR-Lex
- Commission Implementing Decision (EU) 2026/2274 (EN 18286:2026): EUR-Lex
→ See live updates: https://aiactradar.com
→ Subscribe to webhook: https://docs.aiactradar.com/quickstart