Article 4 of the EU AI Act requires providers and deployers of AI systems to take measures that support the development of AI literacy among relevant staff and other people dealing with the operation or use of AI on their behalf.
The obligation has applied since February 2, 2025. It reaches much further than high-risk AI. A business whose employees use ChatGPT, Microsoft Copilot, Gemini or another AI assistant can be a deployer and therefore needs an appropriate literacy approach, even when the tool is used only for writing, translation, research or routine office work.
The 2026 Digital Omnibus changed the wording and removed the idea that organizations must guarantee a particular or “sufficient” level for every individual. It did not delete Article 4 or turn it into a voluntary aspiration. The current rule is deliberately contextual: measures should reflect people’s technical knowledge, experience, education and training, as well as the context in which the AI systems are used.
For the full regulatory structure, start with our complete EU AI Act guide. To determine whether your organization is a provider, deployer or another operator, use the AI Act roles guide.
Article 4 at a glance
| Question | Practical answer |
| Who has the duty? | Providers and deployers of AI systems |
| Who should receive support? | Relevant employees and other people operating or using AI on the organization’s behalf |
| Does it apply only to high-risk AI? | No. It can apply to ordinary workplace AI and generative assistants |
| Is a formal course mandatory? | No single format is prescribed; training, guidance, workshops and embedded controls can be combined |
| Is certification required? | No |
| Must every employee pass an exam? | No general knowledge-testing requirement is imposed by Article 4 |
| Must a company appoint an AI officer? | No specific governance structure is mandated for Article 4 |
| Is reading the vendor manual always enough? | No. The measures must fit role, context and risk; high-risk human-oversight duties can require more |
| How should compliance be evidenced? | Keep records of the systems, target groups, measures, attendance or delivery, materials, decisions and reviews |
| Who enforces Article 4? | National market-surveillance authorities, under the AI Act’s enforcement framework |
What Article 4 now says
The amended provision has three connected parts.
Providers and deployers must act
Providers and deployers must take measures to support the development of AI literacy among their staff and other persons dealing with the operation and use of AI systems on their behalf.
The organization must consider:
- the technical knowledge of the people involved;
- their experience;
- their education and previous training;
- the context in which the systems are used;
- and the people or groups on whom the AI systems may be used.
The provision also states that a provider or deployer is not required to guarantee a specific level of AI literacy for any individual. That distinction matters. The duty is to design and take reasonable, contextual measures—not to promise that every employee will reach the same measurable proficiency.
Public institutions must support implementation
The Commission and Member States are required to support and facilitate providers’ and deployers’ efforts, particularly for smaller organizations. The Commission is expected to publish practical examples on the AI Act Single Information Platform.
Those examples can inform a program, but copying another company’s initiative does not automatically create a presumption of compliance. Your measures still need to fit your own systems, roles and risks.
The AI Board will develop common objectives
The European AI Board is to adopt recommendations supporting AI literacy, taking European competence frameworks into account. These recommendations may make programs more comparable over time, but they do not replace the organization-specific analysis required now.
The Commission’s current AI literacy Q&A is the most useful operational starting point.
What changed in 2026?
Before the Digital Omnibus, Article 4 was commonly summarized as requiring providers and deployers to ensure a “sufficient level” of AI literacy. That formulation caused uncertainty because the Act did not define a universal threshold, examination or certificate.
The amendment made four practical points clearer:
- The obligation remains. Providers and deployers still need measures supporting AI literacy.
- No fixed level is guaranteed. The law does not prescribe a single competence threshold for every person.
- The approach is risk- and context-based. A generic annual awareness video may be inadequate for staff controlling consequential AI decisions.
- Public support increases. The Commission, Member States and AI Board have stronger roles in examples, recommendations and common objectives.
Organizations should update policy language that still says they must “ensure a sufficient level” without qualification. More importantly, they should not use the amendment as a reason to stop training or governance work already started.
Who is in scope?
Article 4 uses two legal roles: provider and deployer.
A provider is generally the party that develops an AI system or has one developed and places it on the market or puts it into service under its own name or trademark. A deployer uses an AI system under its authority, except for personal non-professional activity.
Many organizations are deployers simply because they use third-party AI tools in business processes. Others are providers because they build, commission, brand or materially change systems offered to customers or used internally under the relevant conditions.
The target group is not limited to employees. The Commission explains that “other persons” can include people under the organization’s operational remit, such as:
- contractors;
- external service providers;
- temporary workers;
- consultants;
- and, in some contexts, clients using or operating a system on the organization’s behalf.
The correct scope is functional. Ask who actually configures, prompts, monitors, validates, overrides, maintains or relies on the AI system—not only who appears on the payroll.
Does every company using ChatGPT need AI literacy measures?
The Commission’s answer is effectively yes where employees use an AI system for professional activity and the organization is its deployer. It specifically gives the example of staff using ChatGPT for advertising text or translation and says they should be informed about relevant risks such as hallucinations.
That does not mean every small company needs a large academy. A proportionate program for low-impact use might consist of:
- a short approved-tools and data-handling policy;
- practical instruction on verification and hallucinations;
- examples of acceptable and prohibited prompts;
- copyright and confidentiality guidance;
- a clear escalation route;
- and periodic updates when tools or uses change.
A bank using AI for creditworthiness, a hospital deploying diagnostic support or an employer using automated candidate ranking needs substantially deeper controls. Article 4 is broad, but the intensity of measures is not uniform.
What should an AI literacy program cover?
The content should follow the systems and decisions in your inventory. A useful baseline has six layers.
1. What AI is—and what it is not
People should understand enough about the systems they use to avoid false assumptions. Depending on the audience, this can include:
- the difference between models and applications;
- probabilistic output rather than guaranteed truth;
- training, inference and retrieval;
- limitations of generative AI;
- model updates and version drift;
- and the difference between automation and autonomous decision-making.
The objective is not to turn every employee into an engineer. It is to give them the mental model needed to use the tool safely.
2. The organization’s actual AI systems
General theory should be connected to the approved inventory:
- Which tools are authorized?
- What business purposes are approved?
- Which features are disabled or restricted?
- What data may be entered?
- Which outputs require review?
- Who owns the system and contract?
- How are changes communicated?
Training that never names the systems employees actually use is less likely to change behavior.
3. Error, bias and decision risk
Users should recognize relevant failure modes, including:
- fabricated facts, sources or calculations;
- stale or incomplete information;
- automation bias and over-reliance;
- disparate performance across groups;
- unsafe recommendations;
- prompt injection or manipulated inputs;
- and confident output outside the system’s intended purpose.
The examples should match the role. A marketing employee needs source verification and brand safeguards. An HR user needs discrimination, profiling and employment-risk awareness. A developer needs evaluation, access control and secure integration practices.
4. Data, confidentiality and intellectual property
People need practical rules on:
- personal data;
- special-category data;
- confidential company information;
- customer and supplier data;
- credentials and security secrets;
- copyrighted input and output;
- contractual restrictions;
- retention and vendor training settings;
- and cross-border processing.
Article 4 does not replace GDPR, trade-secret protection, cybersecurity or copyright law. Literacy should help staff recognize when those regimes are engaged.
5. Human oversight and accountability
Training should define what “human review” means in the specific workflow. A reviewer needs:
- enough time and information;
- authority to reject the output;
- competence to recognize failure;
- access to relevant source material;
- a way to record overrides;
- and protection from incentives that make review purely ceremonial.
For high-risk systems, separate duties concerning trained staff and effective human oversight remain. Read our high-risk AI systems guide before treating Article 4 as the only training requirement.
6. Escalation, incidents and complaints
Users should know what to do when something goes wrong. A program should explain:
- when to stop using the system;
- how to report a suspected data leak;
- how to challenge an incorrect or discriminatory output;
- when security teams must be involved;
- how affected-person complaints are routed;
- and who decides whether regulatory reporting is required.
A literacy program is stronger when it connects directly to operational controls rather than ending with a quiz.
Build training by role, not by job title alone
A role matrix avoids both overtraining and dangerous gaps.
| Audience | Minimum useful focus | Additional focus where relevant |
| All approved AI users | Basic limitations, allowed tools, data rules, verification, escalation | Sector and workflow examples |
| Managers and process owners | Accountability, procurement, risk classification, oversight design, metrics | Fundamental-rights and impact assessments |
| Marketing and communications | Factual verification, copyright, synthetic-media disclosure, brand controls | Article 50 deepfake and public-interest text rules |
| HR and people teams | Bias, employment use cases, sensitive data, prohibited emotion recognition | Annex III high-risk classification and worker information |
| Customer-service teams | Chatbot disclosure, escalation to humans, vulnerable users, complaint handling | Essential-service and consumer-law implications |
| Developers and data scientists | Dataset governance, evaluation, logging, security, model limitations, documentation | GPAI and high-risk provider obligations |
| Procurement and vendor management | Role allocation, documentation rights, change notices, audit and incident clauses | High-risk supply-chain evidence |
| Human-oversight operators | Instructions for use, output interpretation, override authority, incident response | Article 14 and Article 26 high-risk duties |
| Executives and board | Risk appetite, accountability, reporting, resources and enforcement exposure | Portfolio-level compliance roadmap |
One person may appear in several rows. Training records should reflect the actual function performed.
Examples of proportionate measures
Low-impact generative writing assistant
A communications team uses an approved enterprise assistant for first drafts. Appropriate measures might include:
- a 45-minute practical onboarding module;
- a prohibited-data list;
- source and quotation verification rules;
- copyright and disclosure guidance;
- an editor sign-off requirement;
- a prompt library showing safe patterns;
- and quarterly updates on product changes.
Customer-service chatbot
A chatbot answers routine questions and escalates complex cases. The relevant staff may need:
- knowledge of Article 50 interaction disclosure;
- escalation thresholds;
- vulnerable-user safeguards;
- monitoring for systematically wrong answers;
- handling of personal data;
- and complaint and incident procedures.
Recruitment-ranking system
A recruitment system scores or ranks candidates. Article 4 measures should sit inside a broader high-risk program and can include:
- Annex III classification awareness;
- limits on intended purpose;
- bias and performance interpretation;
- human-review requirements;
- record keeping;
- worker and candidate rights;
- and authority to disregard the score.
A short general AI course would not be enough for this role.
Software team integrating a third-party model
Developers integrating a model API may need:
- secure prompt and output handling;
- model evaluation;
- prompt-injection defense;
- version and change management;
- upstream documentation limits;
- logging and observability;
- and understanding of when modifications could change the organization’s legal role.
Is formal training mandatory?
Article 4 does not impose one mandatory format. Organizations can use a combination of:
- instructor-led training;
- e-learning;
- workshops;
- role-specific playbooks;
- approved-use guidance;
- system prompts and in-product notices;
- simulations and tabletop exercises;
- office hours;
- communities of practice;
- and targeted refreshers.
However, flexibility is not permission to do nothing. The Commission cautions that simply asking staff to read instructions for use may be ineffective. For high-risk systems, the deployer’s duty to ensure appropriately trained staff for human oversight can require more concrete measures.
Choose formats based on behavior. A one-page policy may work for a narrow, low-risk tool. A live exercise may be necessary where users must identify unsafe output or practice an override.
Does Article 4 require a certificate, exam or AI officer?
No general certificate is required. The Commission says organizations may keep internal records of training and other guidance initiatives.
Article 4 also does not require:
- a particular external course;
- a minimum number of training hours;
- a universal examination;
- an AI officer modeled on the GDPR’s data protection officer;
- or a mandatory AI governance board.
An organization may still choose those structures because its scale or risk profile justifies them. The legal point is that the governance design is not prescribed by Article 4 itself.
How to document compliance
Article 4 is easier to defend when evidence shows a reasoned program rather than a collection of course-completion screenshots.
Keep an AI literacy file containing:
Scope and rationale
- the AI inventory used to identify target groups;
- provider/deployer role analysis;
- risk and context assessment;
- affected persons and business processes;
- the chosen literacy objectives;
- and why the measures are proportionate.
Delivery evidence
- training materials and versions;
- policy acknowledgements;
- attendance or completion records where used;
- workshop agendas;
- guidance distributed to contractors;
- in-product notices or controls;
- and dates of refreshers.
Effectiveness and follow-up
- common errors or questions;
- incidents linked to knowledge gaps;
- spot checks or scenario exercises;
- changes made after feedback;
- metrics such as unapproved-tool use or verification failures;
- and management review.
Article 4 does not generally require employee testing, but organizations may use proportionate assessment to determine whether measures work. Avoid collecting more personal data than necessary.
A 30/60/90-day implementation plan
Days 1–30: establish scope
- Name an accountable program owner.
- Build or update the AI-system inventory.
- Identify providers, deployers and target groups.
- Stop clearly unsafe or unapproved uses.
- Publish interim rules for tools, data and verification.
- Preserve existing training and policy evidence.
Days 31–60: design and deliver
- Create a baseline module for all relevant users.
- Add role-specific modules for higher-impact workflows.
- Train managers, procurement and support functions.
- Include contractors and service providers where in scope.
- Connect learning to incident, privacy and security procedures.
- Record why the program is proportionate.
Days 61–90: test and govern
- Run practical scenarios rather than relying only on attendance.
- Review incidents and questions for knowledge gaps.
- Add training triggers to procurement and deployment gates.
- Define refresh cycles and change events.
- Report coverage and unresolved risk to management.
- Integrate the evidence into the wider AI Act compliance checklist.
When should AI literacy be refreshed?
Use event-based triggers as well as a calendar. Revisit measures when:
- a new AI system is approved;
- a model or major feature changes;
- the intended use expands;
- the organization becomes a provider;
- a high-risk use is introduced;
- Article 50 disclosures change;
- an incident or complaint reveals misunderstanding;
- new official guidance appears;
- or staff move into oversight roles.
Annual training alone can lag behind systems that change monthly.
Enforcement and penalties
Article 4 is supervised by national market-surveillance authorities, not primarily by the AI Office. National authorities can use penalties and other enforcement measures provided under the AI Act and national law.
The practical enforcement question is likely to be evidence: Did the organization identify relevant people, consider their knowledge and the context, take measures, and update those measures when risks changed?
The Act’s maximum penalty tiers should not be treated as automatic invoices. Authorities consider factors such as gravity, duration, affected people, cooperation, mitigation and the size of the organization. The compliance and fines guide explains the enforcement framework.
Common Article 4 mistakes
Treating the 2026 amendment as a repeal
The duty remains; only the formulation and support architecture changed.
Sending everyone the same generic module
Uniform awareness can be a baseline, but it does not address the different risks faced by developers, HR users and human-oversight staff.
Training employees but forgetting contractors
The target group includes other people operating or using systems on the organization’s behalf.
Recording attendance but not the rationale
A completion list does not show why the content, audience and format were appropriate.
Separating literacy from controls
People need approved tools, escalation routes and real authority—not only information.
Assuming technical staff need no additional guidance
Engineers may understand models while still needing organization-specific legal, security, copyright or incident rules.
Frequently asked questions
When did Article 4 start applying?
The obligation began applying on February 2, 2025. The 2026 amendment changed its wording but did not restart the clock.
Does Article 4 apply to small businesses?
Yes, where they are providers or deployers. Measures can be proportionate to size, systems, users and risk.
Is ChatGPT use covered?
Professional use can make the organization a deployer. Staff should receive measures appropriate to the use, including guidance on hallucinations, data and verification.
Does every employee need training?
Article 4 targets staff and other people dealing with AI operation and use on the organization’s behalf. The organization should identify relevant groups rather than assume every person needs identical content.
Must employees pass a test?
No general test is prescribed. Assessment may still be useful where risk or human-oversight responsibilities justify it.
Is an external certificate proof of compliance?
It can be evidence of one measure, but it does not prove that the program fits the organization’s systems, context and risks.
Can internal training satisfy Article 4?
Yes. No external provider is mandatory. Internal measures should be accurate, role-specific, maintained and documented.
Is an AI officer required?
No specific officer or governance board is mandated by Article 4. Organizations can assign accountability through structures that fit their scale and risk.
Does human-in-the-loop use remove the requirement?
No. Both ordinary users and the people performing human oversight need appropriate knowledge and skills.
Does Article 4 replace high-risk training duties?
No. Deployers of high-risk systems have additional obligations concerning trained staff and effective human oversight.
Bottom line
Article 4 is not a certificate requirement and not a demand that every employee become an AI expert. It is an operational duty to identify who deals with AI, understand the systems and risks they face, and support them with measures that fit their role and context.
The strongest program starts with the AI inventory, differentiates audiences, connects learning to real controls, records the reasoning and refreshes whenever the technology or use changes.
byDavid Porter
Popular news
Claude rejects far fewer harmless biology questions
EU AI Act: Complete Guide to Rules, Deadlines and Compliance
Apple links Alibaba’s Qwen with Siri on Macs in China
Nvidia plans up to $3 billion investment in the company behind the Stargate data center
Nvidia DGX and AI Factories: Complete Guide to AI Infrastructure
North Korean hackers build local AI systems for cyberattacks
Latest comments
Loading