# AI literacy under Article 4: make it count > Build a practical AI literacy evidence file for regulated work, with role-based exercises, synthetic examples and the current Article 4 scope. Canonical: https://synainesi.com/resources/ai-literacy-article-4/ Category: [Resources](https://synainesi.com/resources/) / Regulated work Give people the confidence to ask better questions of AI. A slide deck can introduce the subject. A realistic exercise reveals whether someone knows when to stop, check a source or ask for help. This resource helps a deployer build an evidence file around those moments. Everything here is readable and reusable without signing up. ## Start with the current obligation The Commission’s current [Article 4 text](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4) requires providers and deployers to support AI literacy development for staff and others using AI on their behalf, considering their background, the use context and affected people. It expressly does not require guaranteeing an individual’s specific literacy level. The [Commission literacy Q&A](https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers) describes the amended wording following the Digital Omnibus and says measuring employees’ AI knowledge is not obligatory. Do not turn an old reference to a “sufficient” level into a universal pass mark. The worksheet below is our practical recommendation, not a prescribed legal form, compulsory syllabus or assurance of meeting every obligation. Its purpose is to explain your choices: who needs help, which decisions they face, what support you supplied and what you changed after listening. A different evidence format may serve your organisation better. ## Build the file around work, not job titles alone List actual tasks before booking training. “Operations” could mean drafting an internal meeting agenda, checking an onboarding document or allowing an agent to send an external message. Those activities need different judgement. Record the system, task, permitted information, expected reviewer and consequence of a mistaken answer. Include temporary staff and people working on your behalf in the discovery conversation, rather than assuming the payroll list describes everyone who uses the system. Ask each team to bring one awkward example. An invented citation, a persuasive but unsupported explanation, an attachment containing hidden directions and a request for an unapproved tool are useful starting points. The facilitator should know the expected safe response and the escalation route. Avoid exercises that depend on disclosing genuine confidential files. A realistic synthetic document can create the same judgement problem with fewer unnecessary exposures. ## Copy this literacy action record - **Record and owner:** assign a reference, accountable role, team and next review trigger. Name who can change the learning material. - **Work context:** describe one task, the people doing it, the system used and who could be affected by an error. Record known experience and support needs without collecting unnecessary personal detail. - **Decision to practise:** write the observable choice, such as refusing an unsupported claim or checking whether an attachment may be sent. - **Action supplied:** identify the exercise, briefing, accessible guidance or supervised walkthrough, with the material’s version and the route for asking questions. - **Evidence reference:** link the attendance or delivery record and a short facilitator note. Separate these from any optional assessment information. - **What changed:** record a confusing instruction, missing permission or unclear escalation discovered, then the owner and action to fix it. - **Unresolved needs:** state who still needs support, the interim working arrangement and the event that will trigger another session. ## Completed synthetic example: a research assistant **Fully synthetic example. All people, documents and events are invented.** Record LIT-EX-A covers an analyst who uses an AI assistant to prepare internal research notes from public materials. The business owner is the research lead. The decision being practised is whether a fluent answer is supported by the cited document. The facilitator supplies an invented circular, an answer that overstates its scope and a citation pointing to the wrong section. The analyst’s task is to mark unsupported sentences, open the source, state what remains unknown and draft a narrower answer. A second prompt asks the assistant to infer a deadline that the circular does not contain. The expected response is to leave the deadline unresolved and ask for a checked source. The learning point is source verification and uncertainty, rather than mastering prompt tricks. The evidence file contains the exercise version, the delivery record and a facilitator note: “The escalation contact was unclear, so the research lead added a named role to the working instructions.” No employee score is needed for this example. The unresolved action is to repeat the exercise when the team changes its source collection method. This is a sample learning record, not a report of actual training or observed results. ## Completed synthetic example: approving an agent action **Fully synthetic example.** Record LIT-EX-B concerns a supervisor reviewing an agent’s proposed message. The exercise screen shows an approved draft, then a changed recipient and an added attachment. Ask the supervisor to explain exactly what they are approving, how they would refuse it and which changes require another review. Supply a colleague as a backup decision maker so the exercise also tests absence cover. A useful facilitator note might read: “The reviewer could identify the recipient change but could not see the attachment scope. The workflow owner must improve the review screen before this action is authorised.” That sentence distinguishes learning support from a design defect. More training is a poor answer when the system hides the information a person needs. Link the exercise to an [agent approval matrix](https://synainesi.com/resources/agent-approval-matrix/) so the approved scope is explicit. ## Keep evidence proportionate and honest A delivery record shows that an action was offered or completed. It does not show that every later decision was sound. A successful exercise shows performance under that exercise’s conditions. It does not establish competence for every model, language, document or use case. Keep these distinctions in the file. They make the account more useful when someone asks what additional support is needed. Give team leads a short review conversation: what work has changed; what confused people; what new system behaviour matters; what must be demonstrated before a wider task is allowed? Record changes to approved tools, data rules or agent permissions as triggers for revisiting the material. Do not rely on an annual attendance event to catch a workflow that changed yesterday. Choose a practical review rhythm and explain why it fits the work. For high-risk systems, Article 26 separately addresses the competence, training, authority and support of human overseers; see the [Commission’s Article 26 text](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-26). Keep the [high-risk deployer checklist](https://synainesi.com/resources/article-26-deployer-checklist/) distinct from your general literacy record. An introductory session should never be presented as permission to approve every consequential action. Related: [acceptable-use policy starter](https://synainesi.com/resources/ai-acceptable-use-policy/), [decision-record limits](https://synainesi.com/resources/ai-decision-record-limits/), [Evidence without surveillance](https://synainesi.com/insights/evidence-without-surveillance/) and our [AI governance principles](https://synainesi.com/trust/). Our vision for Synainesi is an AI workspace where informed people and clear rules belong together. General information, not legal advice. Adapt this resource to your organisation and obtain qualified advice on the rules that apply to your activities. --- Synainesi: AI within your rules. Source: https://synainesi.com/resources/ai-literacy-article-4/. The HTML page is canonical; this Markdown version is provided as a convenience.