Skip to main content

Article 4 executed: the literacy program you must actually stand up, with evidence

The short answer

Article 4 is the EU AI Act's first live obligation, and it is about people, not paperwork

It has applied since 2 February 2025, before the high-risk rules, and requires providers and deployers to take measures to support the development of AI literacy among their staff and anyone operating AI on their behalf. If you are in scope, it is overdue, not upcoming.

What you will be able to do

  • State what Article 4 of the EU AI Act actually requires, in plain words: that providers and deployers take measures to support the development of AI literacy among their staff and other persons operating AI on their behalf, the duty as rewritten by Regulation (EU) 2026/1744 (in force 27 July 2026), and that this has applied since 2 February 2025 to AI systems of every risk level, not only high-risk ones.
  • Define AI literacy the way the Act does (Article 3(56)): the skills, knowledge, and understanding that let people make informed use of AI and stay aware of its opportunities, its risks, and the harm it can cause.
  • Explain why literacy, and not a policy or a tool, is the obligation that prevents the class of harm the bromism case shows: a person acting on a fluent, confident, wrong AI answer they did not know to distrust.
  • Scope an Article 4 program to your own organization by naming every role that touches AI on your behalf, drawn from your AI systems inventory and your workforce, so no operating hand is left out.
  • Design role-based literacy content that teaches each group the real failure modes of the actual systems it uses, rather than a single generic course, and set an honest sufficient-level bar for each.
  • Build the evidence record that turns a program into a defensible one: who was trained, on what, when, tied to which systems and roles, and how it refreshes when systems change.
  • Defend the program against the challenge it will actually face: a regulator or a claimant asking you to prove that the person who used your AI to harm someone had been taught the risk.

The ideas, one by one

Literacy is the control that sits in the human at the moment of use

The bromism case shows the whole reason the obligation exists: a person acted on a fluent, confident, wrong AI answer they did not know to distrust. Literacy is the skill of knowing an AI can be confidently wrong and knowing to verify, refuse, or escalate. No policy or tool reaches that moment; only a trained human does.

Scope covers everyone who touches AI on your behalf, including non-staff

The text names "other persons dealing with the operation and use of AI systems on your behalf." Contractors and outsourced operators are the population organizations forget and the one most able to harm a customer with your AI. Scope from your inventory and workforce, and leave no operating hand out.

It applies regardless of risk level, and requires no test

Article 4 is not limited to high-risk systems, and the Commission's guidance confirms there is no obligation to measure staff knowledge or issue certificates. That is a relief and a trap: no test is not no proof, and no risk threshold means you cannot wait for a classification to begin.

A generic company-wide course is the classic failure

A module that never names your actual systems, their failure modes, or the harm your use could cause does not teach the awareness Article 3(56) demands; it yields completion records and no literacy. Tailor by role, lead with real failures, and end each audience with a reflex it can use under pressure.

"Sufficient" is proportionate and relative

Set the bar per role by one question: what must this person understand so they will not unknowingly cause the harm this system can do? Sharp and short for frontline and leaders; deep for human-oversight reviewers and developers. Write the bar down, tied to the harm.

Human oversight is only real if the reviewer understands the failures

A person assigned to review AI outputs who does not know how the system fails, and does not know their own instinct will be to trust it (automation bias), provides a rubber stamp, not oversight, and that stamp launders the machine's error into a human decision. The sufficient-level bar for oversight roles is therefore deeper than a frontline reflex, because a false assurance is worse than no oversight at all.

The evidence is what makes the program count

Enforcement by national market surveillance authorities begins 2 August 2026, and after any AI harm a regulator or claimant will ask whether the person understood the risk. A dated record tying named roles to the specific system risks they were taught answers that with a document; a completion tick or a memory does not. Keep the record beside your inventory and refresh it when systems change.

Literacy is a strong control, not the only one

It reduces the chance a human acts on a bad output; it does not fix the model or replace rollout discipline, monitoring, human oversight, and conformity evidence (see Topic 5.6). Treat it as necessary and high-leverage, not sufficient on its own, and the program stays honest.

Scope by the AI people actually use, and teach before they use it

The tidy program scoped only to approved systems and delivered after an incident misses the two exposures that bite most: shadow AI (unsanctioned tools, vendor-switched-on features) that never reached your inventory, and operators running a live system with no literacy yet. Teach a general reflex alongside the named systems, feed any admitted unlisted tool back into the inventory, and trigger literacy at onboarding and at each system rollout, so nobody operates on your behalf ahead of the literacy their work requires.

The proportionality is a scale, not an escape

The calibrated duty sizes the program to your means and your systems' risk; it does not excuse having none. A small organization owes a smaller program, not a missing one: identify who uses your AI, teach them the real failures, keep the record, and be able to show a genuine effort proportionate to your size when someone asks.

This program is the design; Module 9 is the delivery

The Article 4 literacy program evidence you build here is the artifact Topic 9.4 rolls out across a real, resistant, forgetful workforce (see Topic 9.4), and the roles come from the workforce map in Topic 9.1 (see Topic 9.1). Build it specific and evidenced here, or the rollout has nothing real to carry and the record has nothing to show.

Listen to it

The same lesson, talked through by two hosts as episode 35 of the podcast.

Real cases

These examples show the literacy gap, and the obligation that closes it, in real settings. The deep anchor is the bromism case; the others sharpen a specific point and are treated in depth by their owner topics.

Example 1 (the anchor): the ChatGPT bromism poisoning. A 60-year-old man with no psychiatric history developed bromism, an old and now-rare poisoning, after replacing table salt (sodium chloride) with sodium bromide for about three months, a substitution he made after consulting an AI system about removing chloride from his diet. He was hospitalized with paranoia and hallucinations and recovered over roughly three weeks of treatment. Because his original chat logs were unavailable, the treating physicians queried a version of the same AI system and found it would suggest bromide as a chloride replacement in plain language, without a health warning and without asking the purpose of the substitution (Eichenberger and colleagues, Annals of Internal Medicine: Clinical Cases, August 2025). Read through Article 4, this is a pure literacy failure: the user could operate the tool but lacked the one piece of literacy that would have saved him, the knowledge that a fluent, confident AI answer in a health context is exactly the kind you must verify against a real authority before acting. The case is a consumer story, not an organizational one, which is precisely why it is the sharpest possible illustration of the obligation: it shows the harm with no organizational layer present. Article 4 requires you to put that missing layer, literacy, around every person who uses AI on your behalf, so the same fluent-and-wrong failure does not reach your customers through your staff.

Example 2 (the literacy gap made public inside an organization): the Vanderbilt condolence email. In 2023 an office at Vanderbilt University used ChatGPT to draft a mass email responding to a mass shooting and left the AI's own attribution line in the sent message, making the use plainly visible and, to many recipients, offensive (CNN, 2023). The deep treatment of the org-wide literacy rollout belongs to Topic 9.4 (see Topic 9.4); here it makes one point about scope. The people who sent that email were staff using AI on the organization's behalf, exactly the Article 4 population, and the failure was a literacy failure (not understanding what the tool was appropriate for, and not reviewing its output before it went out under the institution's name). A tailored literacy reflex for that role, review AI output before it represents the organization, is cheap; the reputational cost of its absence was not.

Example 3 (why "other persons on your behalf" is in the text): outsourced operators behind an "AI" product. Several enforcement cases have turned on work that was marketed as AI but performed substantially by offshore human operators, and the reverse pattern is just as relevant to Article 4: organizations whose AI is operated or supervised by contractors and outsourced teams rather than employees. The deep treatment of AI-washing and vendor truth belongs to Module 4 (see Topic 5.1) and its own anchors; the Article 4 point is narrow and important. The obligation explicitly covers "other persons dealing with the operation and use of AI systems on your behalf" (Article 4), so if your AI is run or fed by a contractor, their literacy is your obligation. A program that trains only badge-holding employees and assumes the outsourcing partner handled the rest has left the exact population the text names outside its scope.

Example 4 (the high-stakes operator who must be genuinely literate): a clinician reading an AI summary. Consider any clinical or advice setting where staff act on AI-generated summaries or suggestions. External validation work on deployed clinical AI has repeatedly shown these systems can miss cases or fire false alerts at rates their marketing did not suggest; the deep case belongs to its owner topic (see Topic 5.3). For Article 4, the lesson is about the sufficient-level bar for human-oversight roles: a clinician who is assigned to review an AI output but who does not understand how that system fails cannot provide real oversight, and their sign-off becomes a rubber stamp that launders the machine's error into a human decision. For high-stakes operators, "sufficient" literacy is deep literacy, because the harm they can pass along is severe. There is a specific failure worth naming here, called automation bias: the well-documented human tendency to over-trust a machine's output and under-use one's own judgment, which is strongest exactly where the operator is busy, the tool is usually right, and the sign-off is routine. A human-oversight role that does not understand this bias, and does not know the system's real failure modes, will drift into approving whatever the machine produced, which is the rubber stamp that turns oversight into a laundering step. So the sufficient-level bar for these roles is not only "know how the system fails" but "know that your own instinct will be to trust it, and build the habit of checking against that instinct." That is deeper literacy than a frontline reflex, and it is owed precisely because the oversight role exists to catch what the machine got wrong.

Example 5 (the generic tick-box that proves the gap, illustrative method). Picture an organization that assigns a single twenty-minute "Intro to AI" video to all fifteen thousand employees, records the completions, and files that as its Article 4 compliance. The course never names the organization's actual AI systems, never shows a domain-specific failure, and never mentions the harm a wrong output could cause a customer. When an incident later happens, the completion record does not help; it documents that everyone watched a generic video and learned nothing role-specific, which is closer to evidence of a box-ticking culture than of literacy. This is not a named incident; it is the single most common real-world shape of Article 4 non-compliance, and it is worth picturing because it feels like compliance while being its opposite: a paper trail of the wrong thing.

Example 6 (literacy as the cheap control, illustrative). A fintech deploys an internal AI assistant that drafts customer messages. Before rollout, it runs a ninety-minute session for the messaging team built entirely around real, wrong outputs the assistant produced in testing: a fabricated fee, a confident but outdated policy, a misread of a customer's balance. The team leaves with one reflex, "the draft is a suggestion, verify any figure or policy against the real record before you send," and a dated record of the session tied to that system. Months later, when the assistant confidently drafts a wrong overdraft figure, an agent catches it because that is exactly the failure they were shown. The cost was ninety minutes; the avoided harm was a wrong financial statement sent under the company's name. This is the shape of a program that satisfies Article 3(56): tailored, failure-first, and evidenced.

Example 7 (the missing verification reflex, in a real organization): the Chicago Sun-Times AI reading list. In May 2025 the Chicago Sun-Times, and the Philadelphia Inquirer, ran a syndicated summer section ("Heat Index: Your Guide to the Best of Summer") whose reading list paired real, famous authors with book titles that do not exist, such as a fabricated "Tidewater Dreams" credited to Isabel Allende. The list had been produced with an AI tool by a freelance writer working for a third-party content supplier (King Features, a Hearst unit), and it ran as licensed editorial content that the newsroom had not created or reviewed before publication. Chicago Public Media's chief executive said plainly that the list "was created through the use of an AI tool and recommended books that do not exist" (NBC News, 2025; NPR, 2025). Read through Article 4, this is two of the topic's points at once. First, the "other persons acting on your behalf" scope: the content reached the public under the papers' names, produced by a non-staff writer via a supplier, exactly the population the text names and the population easiest to leave outside a literacy program built around employees. Second, the missing reflex: nobody in the chain applied the one habit failure-first literacy installs, which is to verify a fluent, confident AI output (here, plausible-sounding book titles) against a real source before it goes out under your name. A trained reflex at either the writer or the review step, "the AI's list is a draft, not a fact; confirm each title exists before publishing," was cheap; the public correction was not. The deep treatment of AI-generated content and disclosure belongs to its owner topics; the Article 4 lesson is narrow: scope reaches your suppliers, and the verification reflex is the literacy that would have caught it.

Where people go wrong

  • "Article 4 is a future obligation; we will get to it with the high-risk rules." Wrong on timing. Article 4 has applied since 2 February 2025, the earliest operative date in the entire EU AI Act, before the high-risk obligations. If you are in scope, this duty is already live and overdue, not pending. The high-risk timeline is a separate, later matter (see Topic 5.3).
  • "It only applies to high-risk AI." No. Article 4 applies to providers and deployers of AI systems regardless of risk classification. A low-risk internal assistant still has operators who need literacy. Waiting for a high-risk determination to start literacy misreads the scope.
  • "A single company-wide AI awareness course covers it." A generic module that never names your actual systems, their failure modes, or the harm your use could cause does not teach the awareness Article 3(56) requires, and the Commission's guidance warns that relying on instructions-for-use or generic material may be insufficient. It produces completion records and no literacy: cost without protection, and a paper trail that documents the gap.
  • "We only have to train employees." The text explicitly covers "other persons dealing with the operation and use of AI systems on your behalf." Contractors, outsourced operators, temporary staff, and service providers who run or feed your AI are inside the obligation. A program that stops at the payroll leaves the named population uncovered.
  • "There is no test required, so there is nothing to prove." True that Article 4 does not require testing or certificates; false that there is nothing to prove. Enforcement by national market surveillance authorities begins 2 August 2026, and a documented, dated record of who was trained on what is your demonstration of compliance and your shield in any adjacent claim. No certification mandate is not the same as no proof burden.
  • "Literacy means teaching people how to use the AI tools." Usability is not literacy. The bromism victim could use the tool perfectly. Literacy in the Article 3(56) sense is understanding the opportunities, the risks, and the possible harm, which centers on knowing how the tool fails and when to distrust it, not on prompt technique.
  • "Our developers are technical, so they are already literate." Technical skill is not the same as Article 4 literacy. A strong engineer may never have considered the harm a system does to an affected person, which is squarely inside the definition. Builders need literacy tailored to bias, evaluation, drift, security, and downstream harm, not an assumption that competence covers it.
  • "Sufficient means maximal, so we need a huge curriculum." The duty is deliberately proportionate: measures calibrated to people's knowledge, the context, and who the AI is used on. A frontline user needs a sharp, short, memorable reflex, not a machine-learning syllabus. Over-engineering the program is a common way to blow the budget or never finish and therefore train no one.
  • "We wrote an AI usage policy, so literacy is handled." A policy filed in a drawer does not reach the human at the moment of use, which is the only place this class of harm is caught. Literacy is a capability in people; a policy is a document. You need both, and the policy is not a substitute for the trained reflex.
  • "Literacy makes our AI safe." Literacy is a control, not a cure. It reduces the chance a human acts on a bad output; it does not fix the model, and it works alongside rollout discipline, monitoring, human oversight, and conformity evidence, not instead of them. Overselling literacy as safety is its own error.
  • "One training and we are done." Systems change, and the literacy tied to them goes stale. When the assistant is updated, retuned, or replaced, the failure modes shift and the literacy (and its evidence record) must refresh. A one-time session with no refresh is a snapshot of literacy the organization no longer has.
  • "The completion record is the evidence." A completion record proves attendance, not literacy, and if the content was generic it proves the wrong thing. The evidence that matters ties a named role to the specific system risks it was taught, dated and refreshable, so it can answer "did the person who caused this harm understand the risk," which a bare completion tick cannot.
  • "We only need to cover the AI tools we officially approved." Scope by the AI people actually use, not by the list you wish they used. Shadow AI (unsanctioned chatbots, tools switched on by a vendor update, AI features buried inside familiar products) still produces the confident-wrong risk on your behalf, so a program scoped only to approved systems trains people for the wrong estate. The fix is to teach the general reflex alongside the named systems and to treat any admitted unlisted tool as a gap to add to your inventory, not a use to ignore.
  • "We built the AI, so our builders don't need the deployer training." The provider and deployer hats are different, and an organization that fine-tunes or wraps a model wears both. Builders need provider-side literacy (the failure modes they design against: bias, evaluation, drift, security, downstream harm), which the generic deployer awareness session does not deliver. The fix is role-appropriate depth: deployer-side reflex for the frontline, provider-side literacy for the people who could build the harm in rather than merely pass it along.
  • "We ran the training after the incident, so we are covered now." Literacy owed only after a harm reads as reactive and leaves every person who operated the system beforehand unprepared at the exact moment the obligation targets. The fix is to wire literacy to two triggers that come before the hands touch the system: a person joining a role that touches AI (onboarding), and a system being rolled out to a role. Train ahead of use, not after the apology.
  • "An obligation of effort means we can say we tried and move on." The duty to take measures scales the size of the program to your means and the risk; it does not excuse having no program. A small organization is not expected to build an academy, but it is expected to have identified who uses its AI, taught them the real failures, and kept a record. The fix is to make a genuine, resourced effort proportionate to your size, and to be able to show it, rather than treating "we took measures" as a synonym for "little."

Questions people ask

What is EU AI Act (Artificial Intelligence Act)?
Regulation (EU) 2024/1689, the European Union's horizontal law on artificial intelligence, in force since 1 August 2024, whose obligations phase in over several years. Article 4 (AI literacy) and the Article 5 prohibitions were the first to apply, from 2 February 2025.
What is article 4 (AI literacy obligation)?
The provision requiring providers and deployers to take measures to support the development of AI literacy among their staff and other persons operating AI on their behalf, taking into account those persons' knowledge, the context of use, and who the AI is used on. Rewritten by Regulation (EU) 2026/1744 (in force 27 July 2026) from the original duty to ensure a sufficient level; the amended duty expressly guarantees no individual's level. Applies regardless of risk classification.
What is AI literacy (Article 3(56) definition)?
The skills, knowledge, and understanding that allow providers, deployers, and affected persons to make an informed use of AI and to gain awareness of its opportunities, its risks, and the possible harm it can cause. Centered on understanding how AI fails and when to distrust it, not on tool usability.
What is provider?
An organization 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. Providers owe Article 4 for the people who operate their systems on their behalf. More on Provider
What is deployer?
An organization that uses an AI system under its own authority in the course of its activity. Most organizations are deployers of AI they bought; deployers owe Article 4 for their staff and operators. More on Deployer

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →