- Article 4 was amended by the Omnibus, not repealed. AI literacy remains an obligation on providers and deployers; no specific or sufficient level is mandated, and the Commission and the Member States take a stronger promotion role.
- It is one of the few AI Act duties enforceable today. It applied from 2 February 2025 and supervision commenced on 2 August 2026 with the national market surveillance authorities. It was not deferred to 2027 or 2028.
- Removing the mandated level removes a ceiling, not a floor. The operator's own definition of adequate is now the artefact under assessment, which makes literacy a governance question rather than a training question.
- Completion certificates for generic AI awareness training score close to nothing. Role-mapped competence tied to a named system and a named intervention authority scores well.
- In this framework, literacy evidence is scored inside the Autonomy Envelope and Governance dimensions rather than as a category of its own, because it is a property of the oversight arrangement.
Between 27 July and the middle of August 2026, two incompatible statements about Article 4 circulated widely enough that both reached this desk from operators seeking assessment. The first was that the Omnibus had left Article 4 alone. The second was that it had repealed it. Neither is correct, and the correct position is more consequential than either.
Regulation (EU) 2026/1744 amended Article 4 of Regulation (EU) 2024/1689. AI literacy remains an obligation on providers and deployers. What the amendment removed was the mandating of a specific or sufficient level, with the European Commission and the Member States taking a stronger role in promoting literacy. For a deployer of a high-risk system, the duty to train staff so that human oversight is real remains in place. Supervision and enforcement of the Act commenced on 2 August 2026, with the national market surveillance authorities.
This article is about what that means for evidence. It is written for operators preparing an assessment and for the procurement, insurance and audit readers who will be handed the result.
Why removing the level made the obligation harder, not easier
A mandated threshold is a comfortable thing to be regulated by. It converts a judgement into a test. Where a rule names the level, an organisation can meet it, evidence it, and stop; the boundary of the obligation is set externally and disputes are about whether the boundary was reached.
Remove the named level and the obligation does not shrink. It relocates. The organisation must now decide what adequate means for its own systems and its own people, and must be able to defend that decision. An assessor arriving in that situation does not ask whether an external threshold was cleared. The assessor asks how the operator arrived at its answer, whether the reasoning is coherent, and whether the reality in the building matches it.
In our experience with the first assessments run after the amendment, this is a real change in difficulty for two specific groups. Organisations that had procured a generic AI awareness course specifically to point at Article 4 now have an artefact that answers a question nobody is asking. Organisations that had genuinely mapped competence to systems are largely unaffected, because they were already producing the evidence that survives.
Where literacy sits in the seven dimensions
Agent Certified does not score AI literacy as a dimension of its own, and the amendment is a good occasion to explain why. Literacy is not an independent property of an organisation. It is a property of a specific oversight arrangement around a specific system, and it only means anything relative to what that system is permitted to do without a human.
So the evidence lands in two places. The Autonomy Envelope dimension, weighted at fourteen points, asks what the agent may do unsupervised and who may stop it. Literacy evidence is scored there as part of whether the stated intervention authority is capable of being exercised. The Governance dimension, weighted at sixteen, asks how the organisation decides, records and reviews such things, and literacy evidence is scored there as part of whether competence is managed rather than assumed.
A consequence worth stating plainly, because it surprises operators: an organisation can hold extensive training records and still score poorly on both. Training volume with no mapping to intervention authority is an input with no demonstrated output. The framework does not reward inputs.
The reverse also holds and is more common than expected. A small operator with three people, no training platform, and a precise written description of who may halt which agent and what each of them must understand to do so, will score well. The evidence is thin in volume and complete in substance.
What an assessment asks for
Five questions, in this order. They are deliberately answerable by an operator who has never used the word literacy internally.
| The question | What a strong answer looks like | What a weak answer looks like |
|---|---|---|
| Who is expected to intervene in this agent? | Named individuals or a named role with a defined roster, tied to a specific system, with an escalation path when they are unavailable. | A team name. A function. Anybody with access to the admin console. |
| What must they understand to intervene correctly? | A written description of the system's failure modes and the signals that precede them, expressed in the operator's own language rather than the vendor's. | A copy of the vendor's product documentation, or a general statement about understanding the limitations of AI. |
| Can they demonstrate that understanding? | An exercise, an incident retrospective, or a structured conversation in which the person describes their authority and its limits without prompting. | A completion certificate. A signed policy acknowledgement. |
| How did the organisation decide this was enough? | A recorded judgement referencing the consequence of the agent failing, with a named owner and a review date. | Reference to an external standard as though it set the level. Since the Omnibus, no instrument does. |
| What happens when the system changes? | A trigger that reopens the literacy question when the model, the scope of autonomous action, or the personnel change. | An annual training cycle unconnected to system changes. |
The fourth question is the one the amendment created. Before it, an operator could point outward. Now the reasoning has to be the operator's own, and the assessment scores its coherence.
The one place literacy evidence gets read by someone other than a regulator
Underwriting. A specialist underwriter assessing an AI liability submission is answering a question adjacent to the one an assessor answers: whether the scope of autonomous action described in the submission is actually bounded in practice. Both questions turn on whether the people named as the boundary can act as one.
We are careful about how far to take that observation. No carrier or managing general agent writing AI liability publishes a rate card, and no published underwriting guideline names AI literacy as a rating factor. What can be said, and is checkable, is that specialist underwriters ask for a defined scope of autonomous action, an approval threshold, and performance monitoring evidence, and that the human side of those three is the literacy evidence under a different name. An operator who cannot produce it is asking to be priced for an unknown. The mechanics of that submission are set out on agentinsured.eu.
What this does not change
Three boundaries, since the amendment has been over-read in both directions.
It does not change Article 14. Where a high-risk system is in scope, the design obligation that it be capable of effective human oversight is unchanged, and it applies on the deferred high-risk timetable rather than on the Article 4 timetable. Our reading of Article 14 evidence is unchanged and set out in the human oversight evidence guide.
It does not change the tier floors. Certified requires a raw score of four on every dimension, Advanced six, Elite eight. An operator with weak literacy evidence will find the Autonomy Envelope and Governance scores capped, and a capped dimension caps the tier regardless of weighted total. The framework does not reward lopsided agents and did not become more forgiving in July.
It does not change the timetable an operator should be working to. The obligations that moved are the Annex III set, to 2 December 2027, and the Annex I set, to 2 August 2028. Article 4 did not move, Article 5 did not move, Article 50 did not move, and the general-purpose AI obligations did not move. The date cluster that does arrive this year is set out on agentliability.eu.
Where to go next
Operators preparing for assessment should start from the full methodology and the certification levels, then read the preparation guide, which sets out the evidence pack in the order an assessor reads it. Operators whose interest in Article 4 is driven by a procurement or insurance requirement rather than by supervision will find the crosswalk in how the seven dimensions map to EU AI Act obligations the faster route.
Questions
Did the AI Omnibus repeal the Article 4 AI literacy obligation?
No. Article 4 of Regulation (EU) 2024/1689 was amended, not repealed. AI literacy remains an obligation on providers and deployers. What the amendment removed was the mandating of a specific or sufficient level of literacy, with the European Commission and the Member States taking a stronger role in promoting it instead. For a deployer of a high-risk system, the duty to train staff so that human oversight is real remains in place. Two readings circulating since July 2026 are both wrong: that Article 4 was untouched by the Omnibus, and that it was repealed.
Is the Article 4 AI literacy duty enforceable now, or has it moved to 2027?
It is enforceable now. Article 4 applied from 2 February 2025 and was not among the obligations deferred by the Omnibus. Supervision and enforcement commenced on 2 August 2026, with the national market surveillance authorities. The obligations that moved to 2 December 2027 and 2 August 2028 are the Annex III and Annex I high-risk obligation sets. AI literacy is not one of them, which makes it one of the few AI Act duties an operator can be supervised on today.
How does a certification assessment score AI literacy when no level is mandated?
By assessing competence against the specific role a person holds in relation to a specific system, rather than against an external threshold. An assessment asks who is expected to intervene in this agent, what they must understand to intervene correctly, whether they can demonstrate that understanding, and whether the organisation can show how it reached that judgement. The absence of a mandated level removes a ceiling, not a floor: an operator still has to defend its own definition of adequate, and a defensible definition is a scoreable one.
Does a completion certificate from an AI training course satisfy Article 4?
On its own, no, and in an assessment it scores close to nothing. A completion record evidences attendance rather than competence, and generic AI awareness training is rarely mapped to the systems an organisation actually runs. The evidence that scores is role-mapped: a named person, a named system, a description of what they are authorised to do when the system behaves unexpectedly, and a demonstration that they can do it. Training records support that picture. They do not constitute it.
What is the relationship between Article 4 literacy and Article 14 human oversight?
Article 14 requires that high-risk systems be designed so they can be effectively overseen by natural persons. Article 4 is the condition that makes that oversight real rather than nominal. A well-designed intervention control operated by someone who does not understand when to use it satisfies neither article in substance. In the Agent Certified framework this is why literacy evidence is scored inside the Autonomy Envelope and Governance dimensions rather than as a standalone category: it is a property of the oversight arrangement, not a training statistic.