Direct answer
Can we test our AI system before it goes to market?
This falls under Article 60: testing in real world conditions outside a sandbox. That obligation applies today. There is one exception you have to assess yourself.
This could go the other way
- Article 60(1), third subparagraph, leaves Union and national law on real-world testing of high-risk systems related to products under the Annex I harmonisation legislation unaffected. Article 60(4), point (i), contains a law enforcement carve-out: where seeking informed consent would prevent the system from being tested, testing may proceed without consent, provided the testing and its outcome have no negative effect on the subjects and their personal data are deleted after the test. For the systems referred to in Annex III, points 1, 6 and 7, in law enforcement, migration, asylum and border control, registration runs through the secure non-public section of the EU database under Article 49(4), point (d). For the systems referred to in Annex III, point 2, Article 49(5) requires registration at national level.
First step: Submit the testing plan, obtain approval and register the test.
You describe: You want to develop, train and validate an AI system before placing it on the market, in a supervised sandbox or under real world conditions outside the laboratory. Likely role: provider (you are preparing market placement).
This applies now
- Article 60: testing in real world conditions outside a sandboxApplicable
- Article 57: AI regulatory sandboxesApplicable
- Article 4: AI literacyApplicable
Coming up
- Annex III: high-risk AIfrom 2 December 2027
- Article 14: human oversightfrom 2 December 2027
There are two routes and they are not the same. The AI regulatory sandbox of Article 57 is a controlled environment supervised by the competent authority, and every member state must provide at least one. Testing in real world conditions under Article 60 happens outside that environment, with an approved testing plan, informed consent from participants and a limit in time. Both are voluntary routes: they replace no obligation, but they do give you supervision and a written record of what you tested.
Your first actions
- Submit the testing plan, obtain approval and register the test. Draw up a real-world testing plan, submit it to the market surveillance authority, obtain approval, register the test with a Union-wide unique single identification number, and record the division of roles with your deployer.
- Apply to a sandbox and agree the sandbox plan. Apply to the competent authority, agree a specific sandbox plan, and record which uncertainty about the Regulation you want resolved inside the sandbox.
- Take role- and context-specific AI literacy measures. Determine for each role, system and context which combination of instruction, guidance, practice or training is appropriate.
Record this
- Dated and documented informed consent of test subjects
- Written proof of participation and the exit report
- AI literacy measures record
recruitment and selection
Candidate recommendation that automatically becomes a decision
An employer uses a system that ranks applicants and recommends a candidate to hire. In one setup a recruiter weighs that recommendation in their own assessment; in the other the outcome is applied automatically and a candidate is rejected without anyone looking at it.
Provenance: The Commission draft guidelines of 19 May 2026 address this case when determining whether an application falls under Annex III. The document is a consultation version: non-binding and not yet final.
Assess a recruitment system on its intended purpose rather than on whether a recruiter reviews the output, because adding or removing human involvement does not change its high-risk classification.
Draft guidelines on high-risk AI classification, 19 May 2026, annex on Annex III
biometrics and identification
Facial recognition at access control: the guard behind the camera counts too
An organisation secures the entrances to its buildings with facial recognition and uses that biometric access control to register visitors as well. When the system returns no match, a security officer reviews the camera images and decides personally whether someone may enter. The question is whose measures have to reach that officer: those of the supplier of the model, those of the department that deploys the system, or both.
Provenance: Article 4(1) provides that providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf. In doing so they must take into account technical knowledge, experience, education and training and the context the AI systems are to be used in, and consider the persons or groups of persons on whom the AI systems are to be used. The same provision states that this obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.
We read the phrase about the persons on whom the system is used as the centre of gravity for biometrics: whoever stands in front of the camera is subject to the outcome and has little to set against it. That argues for equipping the officer who decides for himself when no match comes back more substantively than the colleague who merely switches the system on and off. The article itself names no sufficient level and expressly states that you need not guarantee one, so where the floor lies for each role stays open. In our assessment a record kept per role, stating the choice made and the reason for it, is easier to defend than one organisation-wide session backed only by an attendance list.
Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.
Article 4(1)
law enforcement
Police using AI in investigations: context sets how deep the training goes
A police force uses AI to search large volumes of investigation files and surface connections a detective would otherwise miss. The outputs feed into the choice of which suspect is pursued further and end up in documents that enter the criminal process. The question is whether one and the same basic instruction is enough for the analyst operating the model and for the detective who acts on its output.
Provenance: Article 4(1) requires providers and deployers of AI systems to take measures supporting the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf. The provision prescribes that they take into account technical knowledge, experience, education and training and the context the AI systems are to be used in, as well as the persons or groups of persons on whom the AI systems are to be used. It also states that this obligation does not require any specific level of AI literacy of any individual to be guaranteed.
Article 4 requires you to weigh the context of use and the people the system is applied to, and in law enforcement both factors run high on our reading. Whether a general introduction to what AI can do is then enough for someone carrying an output into a file that affects a person's position as a suspect, we doubt, but the provision expressly names no level you must guarantee, so that floor is yours to justify. We would record for each role what someone must be able to recognise, for instance that a discovered connection is not yet evidence, and revisit that choice periodically.
Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.
Article 4(1)
media and content
Newsroom with generative AI: do freelancers count within your measures?
A newsroom uses generative AI to prepare summaries, headlines and imagery, after which an editor finishes the piece and the desk decides to publish. Part of that work sits with freelancers, and an outside agency produces marketing content with the same tools. The question is whether your AI literacy measures must reach those freelancers and that agency, or only the people on the payroll.
Provenance: Article 4(1) is addressed to providers and deployers of AI systems and requires them to take measures supporting the development of AI literacy of their staff and of other persons dealing with the operation and use of AI systems on their behalf. The provision requires them to take into account technical knowledge, experience, education and training and the context the AI systems are to be used in, as well as the persons or groups of persons on whom the AI systems are to be used. It does not require any specific level of AI literacy of any individual to be guaranteed.
Alongside staff, the text expressly names other persons dealing with the operation and use of AI systems on your behalf, and we read that as a functional boundary rather than a contractual one. On that reading a freelance editor using your tool inside your workflow and on your instruction sits within your measures, employment contract or not. The outside agency is a harder case: if it works in your environment and on your instruction, the argument that it acts on your behalf holds up, but if it runs its own tools on its own account it is a deployer in its own right, and Article 4 does not say your measures must cover that work. In practice, in our assessment, that means recording in your agreements who works in which role and what instruction you give, rather than trusting the other side to arrange it.
Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.
Article 4(1)
No mandatory course format, no certificate, no exam and no AI officer
The Commission Q&A on AI literacy states that there is no one size fits all when it comes to AI literacy and that no strict requirements or mandatory trainings are imposed. On certification, the Q&A states literally that there is no need for a certificate and that organisations can keep an internal record of trainings or other guiding initiatives. On assessment, it states that Article 4 of the AI Act does not entail an obligation to measure the AI knowledge of employees. On governance, it states that no specific governance structure is mandated to comply with Article 4, so that unlike the data protection officer under the GDPR, no AI officer needs to be appointed. On the level, the Q&A states that following the Digital Omnibus amendment AI literacy remains an obligation for providers and deployers of AI systems, but that no specific or sufficient level is mandated and that the Regulation does not require guaranteeing any specific level of AI literacy of any individual. Against that, the Q&A states that simply relying on the AI systems' instructions for use or asking staff to read them might be ineffective, and that organisations should take into account general AI understanding within the organisation, whether they are a provider or a deployer, the risks associated with the systems deployed, staff knowledge gaps considering technical knowledge, experience, education and training, and contextual factors such as sector, purpose and affected populations. The Q&A further states that organisations may implement different levels of training or learning approaches depending on knowledge, experience, education and role, and that staff with a degree or experience in AI development are normally considered AI literate, while the organisation must still verify that those persons understand the specific AI systems of the organisation, know how to deal with them and are aware of all risks.
Commission Q&A on AI literacy, sections on required level, training formats, certificates, assessment of knowledge and governance structures (consulted 9 August 2026)
Article 4 reaches beyond your own staff, and the national supervisor enforces it
The Commission Q&A on AI literacy states that Article 4 applies to providers and deployers of AI systems and in addition to other persons dealing with the operation and use of AI systems on their behalf, covering persons broadly within the organisational remit, with a contractor, a service provider and a client given as examples. On clients, the Q&A states that they may need AI literacy depending on the specific risk, reasoning that affected persons should understand how decisions taken with the assistance of AI will have an impact on them. On geographic scope, the Q&A states that the AI Act's legal framework applies to both public and private actors inside and outside the EU as long as the AI system is placed on the Union market, used in the Union, or its use has an impact on people located in the EU. On supervision, the Q&A states that the supervision and enforcement of Article 4 is not with the AI Office but under the remit of national market surveillance authorities, and that supervision and enforcement began on 2 August 2026, while Article 4 itself entered into application on 2 February 2025. On sanctions, the Q&A states that national market surveillance authorities could impose penalties and other enforcement measures for infringements of Article 4, that this will be based on national laws that Member States were due to adopt by 2 August 2025, that any sanction must be proportionate and based on the individual case taking into account factors such as the nature and gravity of the infringement and its intentional or negligent character, and that sanctions are more likely if there is proof of an incident due to a lack of appropriate training and guidance. Article 4 is not listed in the enumeration in Article 99(4) of the AI Act, which covers only Articles 16, 22, 23, 24, 26, 31, 33(1), (3) and (4), 34 and 50, so the level of any penalty for Article 4 follows from national law rather than from the Regulation's own ceilings. The Q&A further states that Article 4 reinforces the transparency provisions of Article 13 and the human oversight provisions of Article 14 and indirectly contributes to the protection of affected persons, and that for deployers of high-risk systems the Article 26 obligation to ensure staff are trained to ensure human oversight is a distinct requirement; that requirement becomes applicable on 2 December 2027 for standalone Annex III systems and on 2 August 2028 for Annex I systems.
Commission Q&A on AI literacy, sections on target groups, geographic scope, supervision and enforcement, and sanctions (consulted 9 August 2026)
Article 6 has two separate routes to high-risk
The European Commission's draft guidelines on the classification of high-risk AI of 19 May 2026, which are expressly non-binding, state in paragraph (7) that an AI system is high-risk in two scenarios: first, where it is intended to be used as a safety component of a product, or is itself a product, covered by the Union harmonisation legislation listed in Annex I and required to undergo third-party conformity assessment; and second, where it falls within one of the use cases in the areas listed in Annex III. Paragraph (448) of those same draft guidelines notes that the Article 113 application dates have been postponed by the AI Omnibus to 2 December 2027 for the Article 6(2) route and 2 August 2028 for the Article 6(1) route.
Draft guidelines on high-risk AI classification (19 May 2026), General principles chapter, section II, paragraph (7); section V, paragraph (448)
Broadly positioned and general purpose AI systems: a disclaimer is not enough
According to the non-binding draft guidelines of 19 May 2026 on the classification of high-risk AI, paragraph (12) provides that where the instructions for use, contractual arrangements, terms of service, usage policy, promotional and sales materials or technical documentation present the AI system as broadly applicable across a generality of contexts and functions, and do not consistently limit its application or exclude high-risk uses, the system's intended purpose will be deemed to also encompass high-risk use cases and therefore qualify as high-risk. Under these draft guidelines this applies in particular where such uses are feasible and reasonably foreseeable given the system's functionalities and capabilities. The same paragraph states that merely asserting, for example in the terms of service, that high-risk uses are excluded is insufficient where the provider's overall presentation, examples or product positioning effectively provides for or promotes such uses, and that any limitations of use must be described clearly, concretely and coherently across all materials.
Draft guidelines on high-risk AI classification (19 May 2026), General principles chapter, section II.2, paragraph (12)
prEN 18229-3: AI trustworthiness framework part 3, transparency and human oversight
prEN 18229-3 (AI trustworthiness framework, Part 3: Transparency and human oversight) is the JTC 21 deliverable under M/613 addressing Articles 13 and 14 of the AI Act: transparency and provision of information to deployers, and the design for effective human oversight of high-risk AI systems. As at June 2026 the deliverable was at the drafting stage. It has not yet been published as an EN and is not cited in the Official Journal. This deliverable concerns Article 14 (high-risk) and not the Article 50 transparency obligations, which apply since 2 August 2026.
General interpretation, not legal advice. Checked against Regulation (EU) 2024/1689 and the Digital Omnibus (EU) 2026/1744; the official source remains authoritative.
Full map for your situation