Article 14 of the EU AI Act wants a human who can stop the system. Can your repo show one? A developer published a reference guide mapping Article 14 of the EU AI Act's human-oversight requirements onto repository-level technical measures, covering the two kinds of oversight measures providers must build in or specify for deployers. The guide details Article 14(4)'s overseer capabilities, including the ability to override or interrupt a system via a stop procedure, and notes compliance deadlines of 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems under Regulation (EU) 2026/1744, with penalties up to EUR 15 million or 3% of worldwide turnover. Article 14 of the EU AI Act requires providers of high-risk AI systems to design them so that people can effectively oversee them. It applies to providers that place such a system on the EU market or put it into service, wherever the provider is based. Systems that are not high-risk are not bound by it. Annex III systems the Act's list of high-risk uses, such as credit scoring and hiring tools must comply from 2 December 2027. Annex I systems AI inside regulated products must comply from 2 August 2028, with exceptions below. Regulation EU 2026/1744 set both dates Article 113 . Breaking Article 14 can cost up to EUR 15 000 000 or 3% of worldwide annual turnover, whichever is higher Article 99 4 . SMEs, including start-ups, and small mid-cap companies a size class the Omnibus added pay whichever amount is lower. First, find the place in your product where a person can reject or stop a model output. Article 2 2 , Article 2 13 and Article 111 2 each limit Article 14, and all three were replaced or added by Regulation EU 2026/1744. Annex I Section B is a list of Union product laws. For systems tied to those products, Article 2 2 says: "only Article 6 1 , Article 60a and Articles 102 to 112 shall apply". Article 14 is not on that list. Annex I Section A is a separate list of Union product laws. For those systems, Article 2 13 says the application of Articles 9 to 15 "may be limited". That holds only "where and to the extent that" a product law gives "an equivalent or higher level of protection". The limit must not reduce the overall level of protection. The Commission is to specify the details in delegated acts by 2 August 2027. Article 111 2 covers a system placed on the market before its date of application. The Act then applies "only if, as from that date, those systems are subject to significant changes in their designs". Systems intended for use by public authorities must comply "by 2 August 2030". Systems named in Article 111 1 are excluded from that rule. Article 14 asks for design, so a repository can show the technical measures but not that someone used them well. Article 14 1 says the system must be designed so that natural persons can effectively oversee it while it is in use. Article 14 3 says the measures must fit the risks, the level of autonomy and the context of use. They are ensured through "either one or both" of two kinds of measure. The provider builds one kind in before market placement or putting into service, where technically feasible. The provider identifies the other kind for the deployer to implement. A deployer is the organisation that uses the system. Article 14 4 lists what the overseer must be enabled to do, as appropriate and proportionate. Understand the system's capacities and limits and monitor its operation. Stay aware of automation bias, the tendency to over-rely on the output. Interpret the output correctly. Decide not to use the system, or disregard, override or reverse its output. Intervene or interrupt it through a "stop" button or similar procedure, so that it halts in a safe state. Article 14 5 adds a rule for remote biometric identification systems, which identify people without their active involvement, typically at a distance. For those systems under Annex III point 1 a , no action may follow from an identification unless two natural persons have separately verified it. That rule does not apply where Union or national law finds it disproportionate for law enforcement, migration, border control or asylum. The Act prescribes no tool or file format. The Act's own text is on EUR-Lex: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 Article 14 is in Chapter III, Section 2 . Our reference page for the article: https://scanara.io/en/learn/eu-ai-act/article-14-human-oversight/ https://scanara.io/en/learn/eu-ai-act/article-14-human-oversight/ These four traces are my engineering judgement, not a list from the Act. Each one ties to a paragraph of Article 14 4 . 1. An override path that records who decided and why. Article 14 4 d asks that a person can disregard, override or reverse an output. A function that writes one record per human decision shows the path exists and was used. Requiring a reason even for a confirmation makes a blind click-through visible. python def review decision decision, reviewer, action, reason : action is "confirm", "override" or "reverse" illustrative if not reason: raise ValueError "A review needs a recorded reason" audit log.append { "decision id": decision.id, "model version": decision.model version, "model output": decision.output, "reviewer": reviewer, "action": action, "reason": reason, } 2. A stop path with a test. Article 14 4 e asks for a stop button or similar procedure. A flag that every inference path checks, plus a test that fails when a new output appears after the stop, is a trace a reviewer can run. python def infer request : if stop switch.is set : raise SystemStopped "Stopped by operator" return model.predict request def test stop blocks new outputs : stop switch.set with pytest.raises SystemStopped : infer sample request 3. An output contract for the overseer. Articles 14 4 a to c ask that the overseer can monitor, interpret and not over-rely on the output. A schema test can require every response to carry the model version, a confidence value and a low-confidence flag. The test proves the fields exist. It does not prove anyone reads them. 4. A release gate that fails closed. The gate blocks the release when the override test or the stop test is missing or failing. As in the logging and risk posts, a warning leaves nothing that anyone must act on. REQUIRED = "test override is recorded", "test stop blocks new outputs" def gate passed tests : missing = t for t in REQUIRED if t not in passed tests if missing: raise SystemExit f"Oversight tests missing or failing: {missing}" None of this proves that a person oversaw the system well. It proves that the paths exist, were exercised and are checked at release. For a system that takes actions, the Act's phrase "safe state" is not defined further, and a stop test on new outputs says nothing about actions already in flight. Scanara's code scan checks Article 14 narrowly. Take a file that calls a model, such as .predict or .generate . Three separate rules apply to it, and each is satisfied only by its own function name: flag low confidence output , apply with override and install abort handler . The scan merges the rules still open for that file into one Article 14 finding, so the finding clears only when all three names are present. Each rule is a presence check on a function name. It cannot tell whether the call reflects a real oversight process. Only the stop rule has a guard: it does not count a call whose reason is empty, "none" or "n/a". A clear result is a prompt for a person to look, not proof of oversight. A second set of keyword rules looks for oversight vocabulary in identifiers, such as human-in-the-loop, manual approval, stop function and override. It also has vocabulary for Article 14 4 a , b and d . Who oversees the system, with what training and authority, is not something a code scan can see. Passing a scan does not mean an organisation is compliant. Scan your repository on the free plan: https://scanara.io/en/ https://scanara.io/en/ Is a stop test on new outputs enough? Does "safe state" in Article 14 4 e mean more for an agent that has already sent an email or changed a record? Requiring a reason on every confirmation may produce one-word reasons typed to clear the form. Is the record then evidence of oversight, or evidence of friction? A check that looks for a named helper may teach teams to add a stub call to clear it. Is that signal worth having at all? I have no data on this and would like to hear from people who run similar gates.