Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

Human Oversight Under AI Act Article 14: Design Requirements That Survive Audit

Human oversight is the AI Act requirement most likely to fail an audit, because assessors test behaviour rather than paperwork. What Article 14 demands of providers and deployers, and the evidence to keep.

Renomira Team · 11 min read
A professional reviewing a folder alongside her laptop at a desk
Photo: Shixart1985 / CC BY 2.0 via Wikimedia Commons

Human oversight looks like the softest requirement in the AI Act and fails hardest in assessment. Every mid-market company running a high-risk system can say a person is “in the loop”; almost none can show that person had the information, authority, time and interface needed to disagree. Here is what Article 14 requires, how the duty splits, and what assessors test.

What Article 14 actually requires

Three working paragraphs precede the list everyone quotes:

  • Article 14(1) — oversight must be possible. High-risk systems “shall be designed and developed in such a way, including with appropriate human-machine interface tools, that they can be effectively overseen by natural persons during the period in which they are in use.”
  • Article 14(2) — oversight has a purpose. It must prevent or minimise risks to health, safety and fundamental rights from intended use or reasonably foreseeable misuse — supervision that reduces none of them does not discharge the duty.
  • Article 14(3) — oversight is proportionate. Measures must be “commensurate with the risks, level of autonomy and context of use” and come from (a) measures built in by the provider, (b) measures the provider identifies for the deployer, or both.

For Annex III systems — hiring, credit, essential services — the duty applies from 2 December 2027 (Digital Omnibus timetable). Two duties already bite: Article 4 AI literacy since 2 February 2025, Article 86 explanation right since 2 August 2026.

Provider design duties versus deployer duties

The provider designs the affordances. Under 14(3)(a) it must, where technically feasible, build oversight into the product: confidence displays, reason codes, borderline-case routing to a person, override and halt controls, intervention logs. It must also identify what it expects the deployer to run under 14(3)(b), setting both out in the Article 13 instructions.

The deployer makes oversight real. Article 26: use it as instructed (26(1)); assign human oversight to natural persons with the necessary competence, training and authority, and the necessary support (26(2)); resource the provider’s measures (26(3)); keep input data relevant (26(4)); monitor and suspend on risk (26(5)); retain logs six months (26(6)); inform workers’ representatives before workplace use (26(7)).

Three consequences for a buyer:

  • You cannot contract the duty away. A “solely responsible for oversight” clause does not transfer design obligations, and a shipped dashboard does not erase your 26(2) duty.
  • The split must be written down. Ask which 14(4) measures are built in and which your team runs; if neither contract nor instructions say, the gap surfaces at assessment.
  • Get change notification. Commit vendors to telling you when oversight-relevant instructions change; features disappear in releases.

The five things an overseer must be enabled to do

Article 14(4) requires overseers, as appropriate and proportionate, to:

  • (a) understand and monitor the system’s capacities, limitations and operation, including anomalies and unexpected performance;
  • (b) resist automation bias — awareness of the tendency to rely or over-rely on output, named explicitly in the article;
  • (c) interpret the output correctly, taking into account the interpretation tools available — the article 14 ai act interpretability requirement, and the mechanism making the other four possible;
  • (d) override — decide not to use the system, or disregard, override or reverse its output;
  • (e) stop — intervene through a “stop” button or similar procedure halting the system safely.

Preventing automation bias in the workflow

Automation bias is accepting the machine’s answer because it is there. It is invisible in policy and obvious in data, which is why it is the most common finding against a human in the loop ai act compliance arrangement. Three patterns give it away:

  • Overrides near zero. If “reject” has never been overridden in six months, auditors assume nobody is reading it.
  • Review times near zero. Time-to-decision from render shows engagement or scrolling.
  • Justifications that copy the model. A rejection note echoing the score verbatim means the human added nothing.

The fix lives in the interface: show reasons before the accept button; require free-text justification when following a borderline recommendation; make disagreement one click; and hide rankings until operators have formed a view. Prevent automation bias AI Act-style in the product, not the handbook.

Override and stop in three mid-market settings

CV screening. Override: reject or reorder the ranking, reason recorded; stop: ranking suspended, applicants revert to an unranked queue. Credit decisions. Override: underwriter declines the AI recommendation for manual assessment; stop: automated declines suspended, adverse decisions fall to manual review. Support ticket routing. Override: re-route a misclassified ticket or reopen an auto-closed conversation; stop: auto-routing and auto-close suspended, first-in-first-out resumes.

Two rules keep this survivable. The override must be one click with no approval queue — one needing a manager’s sign-off will not be used on a Friday afternoon. The stop control must be tested, not documented: Article 26(5) obliges a deployer to suspend use when it has reason to think the system presents a risk, so the halt needs a dated drill and defined safe state.

Competence, authority, support: Article 26(2) and Article 4

Article 26(2) has four legs, and all must exist together. Competence: the person can read the output and knows the domain. Authority: documented power to override, escalate and halt without asking. Training: refreshed when the model or rule pack changes. Support: protected time to review, access to the record, escalation, cover. A named owner with no time is oversight on an organigram only.

Article 4 AI literacy is the floor beneath it. Since 2 February 2025 providers and deployers must support AI literacy among staff operating their AI systems, accounting for knowledge, experience, training, context of use and the people affected — with no obligation to guarantee a level for any individual. Keep the evidence of that effort.

AI Act human oversight training should cover capabilities and limits; known failure modes and their data conditions; reason codes; when overriding is expected rather than discouraged; escalation, logging and the incident clock. Scope it by role, version it against the model release, keep completion records — a slide deck without attendance proves nothing.

The oversight control matrix

Oversight control Who implements Evidence to keep
Interface showing capacities and limits (Art. 14(4)(a)) Provider (Art. 14(3)(a)) Instructions of use, screenshots
Automation-bias prompts, reason display (Art. 14(4)(b)–(c)) Provider builds; deployer keeps enabled Configuration record, change log
Per-decision override with reason (Art. 14(4)(d)) Deployer Override log: actor, timestamp, rationale
Stop control with defined safe state (Art. 14(4)(e)) Provider supplies; deployer operates Procedure, drill record
Named overseer: competence, training, authority, support (Art. 26(2)) Deployer Appointment record, cover policy
Literacy and oversight training (Art. 4) Deployer Curriculum, completion records
Monitoring and suspension on risk (Art. 26(5)) Deployer Monitoring reports, suspension records
Explanation of individual decisions (Art. 86) Deployer Explanation template, response log

How auditors test oversight in practice

Assessors start with transactions: they cannot be rehearsed.

  1. Sample reconstruction. The last twenty decisions: what each overseer saw, changed and when; divergence from documented process is a finding.
  2. Distributions, not averages. Override, acceptance and time-to-decision per operator; a flat zero reads as control failure, not performance.
  3. Interview the overseers, not the owner. What do you do when a score looks wrong? When did you last override one? Vague answers read badly.
  4. Stop drill. Demonstrate the halt in a test tenant and confirm the safe state: does the queue revert to unranked?
  5. Explanation test. File a live Article 86 request: the system’s role and the main elements of the decision. Annex III point 2 biometrics are excluded; “GDPR covers this” must be a documented position.
  6. Instructions-of-use reconciliation. Tick the measures assigned under 14(3)(b) against live configuration; divergence is a finding.
  7. Worker information check. Article 26(7): date, audience, content, and whether oversight was described.

How to check your own position

Run this against every high-risk system before the December 2027 duties land:

  1. List your systems. Recruitment, creditworthiness, essential services and worker monitoring are Annex III high risk unless documented.
  2. Split the measures. Mark each 14(4) measure provider-built or deployer-operated; confirm both appear in the instructions.
  3. Name the overseers. One accountable person per system, written authority to override and halt, named cover.
  4. Pull the numbers. Override rate, median review time and acceptance rate over six months — investigate rubber-stamping first.
  5. Test the stop control. Dated drill, defined safe state, access list reviewed.
  6. Check the records. Training completion per role, six-month log retention, override reasons captured, Article 26(7) notice dated.
  7. Draft one explanation. Take a real declined candidate and draft the Article 86 response — system role plus main elements, or it is not ready.

Items 4, 5 and 7 cannot be fixed in a week.

Frequently asked questions

Is a human in the loop enough to satisfy Article 14?

No. A person clicking “confirm” is presence, not oversight. Article 14(4) requires them to understand capacities and limitations, monitor for anomalies, stay aware of automation bias, interpret output, override or reverse it, and interrupt it. Without the information, tools, authority and time, the arrangement fails — an override rate of zero and review times in seconds is how assessors prove it.

Who is responsible for human oversight — us or the vendor?

Both, on different terms. The provider owes Article 14(3)(a): build oversight into the system where feasible and document it. The deployer owes Article 26(2): assign oversight to persons with competence, training, authority and support, and implement the measures identified under 14(3)(b). A contract moves neither duty.

Does Article 14 apply if the AI only recommends rather than decides?

Yes, Article 14(4)(b) names that case: the duty applies “in particular for high-risk AI systems used to provide information or recommendations for decisions to be taken by natural persons”. Recommendation-only screening, scoring and routing tools are exactly the pattern targeted; where the decision has legal or similarly significant effects, the Article 86 explanation right and GDPR Article 22 may also apply.

How much AI Act human oversight training is enough?

Article 26(2) requires “the necessary competence, training and authority” for oversight persons; Article 4 requires measures to support AI literacy, without guaranteeing a level for anyone. In practice: role-based training on capabilities, limitations, failure modes, reason codes, override expectations and escalation, refreshed on each material model change, with completion records kept.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

ai act article 14 human oversighthuman oversight requirements high risk aihuman in the loop ai act complianceoversight measures ai act deployerprevent automation bias ai actai act human oversight trainingarticle 14 ai act interpretability

Keep reading