Filed under — EU AI Act · Article 14 · Human Oversight · High-Risk AI · Compliance · AI Governance
Prefer this source on Google →EU AI Act Article 14: Human Oversight Requirements Explained
Article 14 mandates human oversight for high-risk AI systems. Learn what oversight measures you must implement and how to document them before December 2027.
Free check · no account
Check your own system against what this article describes.
EU AI Act Article 14: Human Oversight Requirements Explained
Updated 2 October 2026 — correction. Earlier versions cited deployers' obligations as Article 29 (they are Article 26) and the right to explanation as Article 72 (it is Article 86), presented human-in-the-loop, on-the-loop and in-command as types Article 14 names (they come from the High-Level Expert Group's 2019 Ethics Guidelines), and said systems already deployed must comply by 2 December 2027, which Article 111(2) does not say.
Updated 4 August 2026 — deadline changed. The Digital Omnibus (adopted by the European Parliament on 16 June 2026 and the Council on 29 June 2026) deferred the Annex III high-risk obligations from 2 August 2026 to 2 December 2027. The Article 50 transparency obligations were not deferred and have applied since 2 August 2026. This article has been corrected accordingly.
If your AI system is classified as high-risk under the EU AI Act, Article 14 requires you to design it so that humans can effectively oversee its operation. This isn't a checkbox exercise — it's a fundamental architectural requirement that affects how you build, deploy, and monitor your system.
Article 14 mandates that high-risk AI systems must be designed to enable human oversight through appropriate measures. These measures must allow humans to understand system outputs, interpret results, and intervene when necessary. These obligations apply from 2 December 2027 for Annex III systems — with a statutory maximum of €15 million or 3% of global annual turnover under Art. 99(4), and for SMEs whichever of the two is lower.
This guide explains what Article 14 requires, what oversight measures satisfy the regulation, and how to implement human oversight that works in practice.
What Article 14 Requires
Article 14 applies to providers of high-risk AI systems (those classified as high-risk under Article 6, through Annex I or Annex III). It requires that systems be designed and developed in such a way that they can be effectively overseen by natural persons during their use (Regulation (EU) 2024/1689, Article 14).
Core Human Oversight Obligations
Human oversight must aim to prevent or minimize risks to health, safety, or fundamental rights that may emerge when a high-risk AI system is used in accordance with its intended purpose, or under conditions of reasonably foreseeable misuse.
Under Article 14(4), the system must be provided so that the people assigned to oversee it are enabled, as appropriate and proportionate, to:
- Properly understand the relevant capacities and limitations of the high-risk AI system and duly monitor its operation, including to detect and address anomalies, dysfunctions and unexpected performance
- Remain aware of the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system (automation bias)
- Correctly interpret the system's output, taking into account, for example, the interpretation tools and methods available
- Decide not to use the system or otherwise disregard, override, or reverse the output in any particular situation
- Intervene in the operation of the system or interrupt it through a "stop" button or similar procedure that brings it to a halt in a safe state
Additionally, oversight measures must be commensurate with the risks, level of autonomy and context of use, and must be either identified and built into the system by the provider before it's placed on the market, where technically feasible, or identified as appropriate for implementation by the deployer — or both (Article 14(3)).
Three Ways to Organise Oversight
Article 14 does not name oversight types. A common way to organise the measures it requires borrows three patterns from the High-Level Expert Group on AI's 2019 Ethics Guidelines for Trustworthy AI: human-in-the-loop, human-on-the-loop and human-in-command. They are one reading of Article 14, not categories the Act defines.
1. Human-in-the-Loop (HITL)
The AI system provides a recommendation, but a human makes the final decision before any action is taken.
Example: An AI system recommends rejecting a loan application, but a human loan officer must review the recommendation and approve the rejection before the applicant is notified.
When it fits: High-stakes decisions affecting individuals (hiring, credit, benefits eligibility). Article 14 itself prescribes a specific human check in one case only: for remote biometric identification systems (Annex III, point 1(a)), no action or decision may be taken on an identification unless at least two competent natural persons have separately verified and confirmed it, subject to exceptions in law enforcement, migration, border control and asylum (Article 14(5)).
2. Human-on-the-Loop (HOTL)
The AI system operates autonomously, but a human monitors its operation in real-time and can intervene if necessary.
Example: An autonomous vehicle drives itself, but a safety operator monitors the system and can take control at any time.
When it fits: Real-time systems where human-in-the-loop would introduce unacceptable latency, but human intervention must remain possible.
3. Human-in-Command (HIC)
A human oversees the overall operation of the AI system, including the ability to deactivate or shut it down.
Example: A hospital administrator can disable an AI-powered diagnostic tool if it begins producing unreliable results.
When it fits: As a baseline for every high-risk system. Article 14(4)(e) requires that the overseers be able to intervene in the system or interrupt it through a "stop" button or a similar procedure, as appropriate and proportionate.
Most high-risk AI systems need multiple oversight layers — for example, human-in-the-loop for individual decisions plus human-in-command for system-level control.
Article 14 Compliance Checklist
Here's what you must implement and document:
| Requirement | What You Must Implement | Evidence Needed |
|---|---|---|
| Understanding capacities and limitations | Training materials, system documentation, performance disclosures | User manual, training completion records, instructions for use (Article 13) |
| Awareness of automation bias | Warnings, training on over-reliance risks, decision-forcing functions | UI warnings, training materials, decision audit logs |
| Interpretation tools | Explainability features, confidence scores, feature importance | Explainability reports, UI screenshots, interpretation guide |
| Ability to override or disregard | Override button, manual review workflow, rejection mechanism | UI design docs, override logs, workflow diagrams |
| Ability to intervene or stop | Emergency stop button, system shutdown procedure, escalation path | Technical architecture, stop button design, incident response plan |
| Oversight role assignment | Who oversees the system, qualifications required, escalation hierarchy | Role definitions, RACI matrix, training requirements |
Practical Example: AI-Powered Hiring Tool
Suppose you provide an AI system that screens CVs and recommends candidates for interviews — a high-risk system under Annex III, point 4(a).
Step 1: Identify Required Oversight Type
Your system shapes decisions that significantly affect individuals' access to employment. Article 14 does not prescribe a pattern for it; it requires measures commensurate with the risks, the level of autonomy and the context of use. For a CV screener, human-in-the-loop oversight is the pattern that fits best: a human reviews and approves every hiring decision before candidates are notified.
Step 2: Design Interpretation Tools
You implement explainability features so hiring managers can understand why the system recommended or rejected a candidate:
- Feature importance scores: "This candidate was ranked highly due to: relevant experience (35%), education match (28%), skills alignment (22%), other factors (15%)"
- Confidence score: "Confidence: 78% (medium confidence — manual review recommended)"
- Comparison view: Side-by-side comparison of top candidates with key differentiators highlighted
Step 3: Implement Override Mechanism
You build a workflow where hiring managers can:
- Accept the AI recommendation (candidate moves to interview stage)
- Reject the AI recommendation (candidate is manually reviewed by senior recruiter)
- Flag for review (case escalated to hiring committee)
Every override is logged with a reason code (e.g., "AI missed relevant experience," "candidate has unique background," "bias concern").
Step 4: Mitigate Automation Bias
You add UI warnings to prevent over-reliance:
- Decision-forcing prompt: "Before accepting this recommendation, have you reviewed the candidate's full CV?"
- Randomized manual review: 10% of AI recommendations are flagged for mandatory manual review, even if the hiring manager agrees with the AI
- Training requirement: All hiring managers must complete a 30-minute training on automation bias before using the system
Step 5: Provide System-Level Control
You implement human-in-command oversight:
- System administrator (Head of HR) can disable the AI system at any time
- Performance dashboard shows accuracy, bias metrics, and override rates in real-time
- Automatic shutdown triggers: System disables itself if accuracy drops below 80% or if bias metrics exceed predefined thresholds
Step 6: Document Everything
You create an Oversight Design Document that includes:
- Role definitions (who oversees what)
- Oversight workflows (diagrams showing decision paths)
- Interpretation tools (screenshots, user guide)
- Override mechanisms (technical design, logs)
- Training requirements (curriculum, completion tracking)
- System-level controls (shutdown procedures, escalation paths)
This document becomes part of your Article 11 technical documentation and informs your Article 13 instructions for use.
Common Gaps and How to Fix Them
Gap 1: No Explainability Features
Problem: Your system produces recommendations, but users can't understand why.
Fix: Implement interpretation tools:
- Confidence scores (how certain is the system?)
- Feature importance (what factors drove this decision?)
- Counterfactual explanations (what would need to change for a different outcome?)
- Comparison views (how does this case compare to similar cases?)
Gap 2: Override Mechanism Exists But Isn't Used
Problem: Users can override the system, but in practice they almost never do (automation bias).
Fix: Implement decision-forcing functions:
- Require users to actively confirm decisions (not just click "accept all")
- Randomize mandatory manual reviews
- Track override rates and investigate if they're too low
- Train users on when and how to override
Gap 3: No System-Level Shutdown Capability
Problem: Individual users can reject recommendations, but no one can stop the entire system if it starts malfunctioning.
Fix: Implement human-in-command controls:
- Designate a system owner with shutdown authority
- Build an emergency stop mechanism (e.g., admin dashboard with "disable system" button)
- Define automatic shutdown triggers (accuracy thresholds, bias thresholds, incident reports)
- Document escalation procedures (who gets notified, how quickly, what happens next)
Gap 4: Oversight Roles Are Undefined
Problem: It's unclear who is responsible for overseeing the system, what qualifications they need, or what they're supposed to do.
Fix: Define oversight roles and responsibilities:
- Who reviews individual decisions? (e.g., hiring manager, loan officer)
- Who monitors system-level performance? (e.g., compliance lead, ML engineer)
- Who has authority to shut down the system? (e.g., CTO, Head of Compliance)
- What qualifications are required? (e.g., training completion, domain expertise)
- How are oversight activities logged and audited?
How Article 14 Connects to Other Articles
Article 14 oversight requirements intersect with several other obligations:
- Article 9 (Risk Management): Risks identified in your Article 9 risk assessment inform what oversight measures are needed under Article 14.
- Article 13 (Transparency): The oversight measures you implement under Article 14 must be described in your Article 13 instructions for use.
- Article 26 (Obligations of Deployers): Deployers must assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support (Article 26(2)) — which requires that you (the provider) have designed the system to support effective oversight.
- Article 86 (Right to Explanation of Individual Decision-Making): A person affected by a decision that a deployer takes on the basis of the output of an Annex III high-risk system (other than those in point 2), and that produces legal effects or similarly significantly affects them, has the right to obtain from the deployer clear and meaningful explanations of the role of the AI system and the main elements of the decision. The deployer can only give that explanation if your system's output can be interpreted — another reason to build interpretation tools in.
What Regulators Will Look For
When a market surveillance authority examines your high-risk AI system, expect questions such as:
- Show me how humans oversee this system. (What workflows, tools, and controls exist?)
- How do the overseers understand what the system is doing? (Are explainability features built in?)
- Can the overseers override or reject system outputs? (Is there a documented override mechanism?)
- How do you prevent automation bias? (What training, warnings, or decision-forcing functions exist?)
- Who can shut down the system if it malfunctions? (Is there a designated owner with shutdown authority?)
- How do you know oversight is working? (Are override rates, review times, and incident reports tracked?)
If you can't demonstrate effective oversight with documentation and logs, the system appears not to meet Article 14.
Timeline and Enforcement
| Date | Milestone |
|---|---|
| 2 August 2025 | GPAI obligations and the penalty regime apply |
| 2 August 2026 | Article 50 transparency obligations apply (not deferred) |
| 2 December 2027 | Article 14 obligations apply for Annex III high-risk AI systems |
| 2 August 2028 | Annex I high-risk obligations apply (AI embedded in regulated products) |
If your high-risk AI system was already placed on the market or put into service before its date of application, the high-risk requirements reach it only if its design changes significantly from that date on (Article 111(2), as replaced by the Digital Omnibus). Systems intended to be used by public authorities are the exception: their providers and deployers must comply by a separate deadline set in Article 111(2), whether or not the design changes. If you're building a new system, Article 14 applies from the design phase.
This article is for informational purposes only and does not constitute legal advice. Consult a qualified legal professional for guidance on your specific situation.
Related dispatches