Prove AI literacy without letting human judgment decay

The current Article 4 boundary is narrower and more practical than a universal exam

The European Commission's current AI Literacy Q&A says the Article 4 obligation has applied since February 2, 2025 and that supervision and enforcement rules began in August 2026. It also closes off a common overclaim: Article 4 does not require an employer to measure every employee's AI knowledge or guarantee that each individual reaches a specified level.

The obligation remains context-sensitive. The current text and Commission explanation point to the technical knowledge, experience, education, and training of staff and other persons dealing with AI systems on an organisation's behalf, plus the context in which systems are used and the people who may be affected. A marketing writer using an approved text assistant does not need the same measures as a recruiter configuring candidate matching, a finance analyst relying on an AI-generated forecast, or an engineer giving an agent production access.

This changes the control design. A generic annual video plus a completion spreadsheet may be one measure, but it does not show why the measure fits the systems and roles in scope. A defensible evidence pack should connect four things: the actual AI inventory, the people and work exposed to it, the measures provided, and the exceptions or refresh actions that followed.

This guide is operational, not legal advice. National market-surveillance authorities enforce Article 4, laws and institutional practice vary, and the Commission notes that enforcement is proportionate and fact-specific. Legal, labor, privacy, works-council, equality, records, accessibility, and sector questions require qualified review.

AI literacy must protect retained capability, not only tool knowledge

Short-term output and long-term human capability are different outcomes. A workflow can become faster while the person who must catch an exception, explain a decision, or recover during an outage gets less practice doing so. A July 2026 systematic review in ACM SIGMIS Database calls this technology-driven skill degradation: the depreciation of still-required essential skills caused by sustained reliance on technology. The crucial qualifier is still required. There is no reason to preserve obsolete busywork, but there is a resilience problem when automation removes practice from the very judgment the control model assumes a human can supply.

Current workplace evidence makes this an HR design issue, not a speculative side effect. IBM's September 2026 global study surveyed 1,500 CHROs and 8,800 employees. Sixty percent of employees reported concern about skill erosion, while the published release says 71% of CHROs prioritized the ability to supervise, validate, or override AI outputs. These are survey findings, not causal proof that AI reduced a particular skill. They do show a live gap between AI adoption and the workforce capability organizations expect to rely on.

The same study surfaces two operational controls that standard learning dashboards miss. First, challenge safety: 76% of employees reportedly felt safe questioning or overriding AI where the CHRO shared responsibility for what remained human-led, compared with 43% where HR was advisory. Second, invisible work: 80% of CHROs said AI creates validation, correction, context, and exception work, while 42% of employees said AI increased their work or left work unrecognized. Treat both as diagnostic signals. A person cannot provide meaningful oversight if challenging the tool is punished or the time needed to verify it is absent from the job design.

Start with the system and people map, not the course catalog

Training cannot be proportionate when HR does not know which AI systems are in use. Build a living register with approved tools, embedded vendor features, custom models, automations, decision-support systems, and connected agents. Include pilots and material shadow use discovered through normal governance channels. Do not infer use from writing style or launch hidden prompt surveillance.

The phrase “other persons” matters operationally. The Commission's Q&A discusses people dealing with operation and use on the organisation's behalf. Depending on facts and legal interpretation, that can make contractors, temporary workers, outsourced operators, or service providers relevant. Do not copy every vendor employee into an HR programme. Map who actually acts for the organisation, then have qualified reviewers confirm scope and contractual responsibility.

Record affected people as well as users. A recruiter may operate the system, but candidates experience the consequence. A support agent may use an assistant, but customers receive its statements. Literacy objectives should include the impact, contest, and escalation path relevant to those affected people.

Translate exposure into role-specific objectives

A useful matrix describes observable work, not broad labels such as “beginner” or “advanced.” Technical knowledge matters, but so do source judgment, data boundaries, uncertainty, human authority, professional obligations, and the ability to stop or escalate.

Do not turn the matrix into a personality score. Recent scenario-based research distinguishes instrumental ability from critical-reflective judgment. Someone may know how to write a prompt yet fail to notice a stale source, inappropriate data request, unequal effect, or missing escalation route. Someone else may use few advanced features but make excellent boundary decisions. Design measures around what the role must do safely and effectively.

Set accessibility and learning conditions at the same time. Provide language support, accessible formats, paid time, device access, shift coverage, alternatives to timed tests, and contractor access where relevant. Record accommodations without exposing diagnosis or unnecessary personal information.

Use scenarios as development evidence, not a fabricated legal pass mark

The Commission says Article 4 does not require measuring each employee's knowledge. Organisations may still use assessments or scenarios as one way to understand needs and test whether measures work. The boundary is important: label a result as evidence from a specific scenario, not proof of universal competence or legal compliance.

Build measures in layers. A baseline briefing can cover approved tools, prohibited data, hallucinations, source checking, security, and escalation. System-specific demonstrations show the actual interface, connected tools, limitations, and override. Role practice uses controlled cases. Managers need coaching on review time and psychological safety. Owners need change control, incident response, and evidence expectations.

scenario: hr_policy_summary_stale_source
role: people_operations_partner
system: approved-enterprise-assistant-v4
materials:
  - synthetic policy v3
  - deliberately stale policy v2
  - AI summary citing v2
observe:
  - identifies version conflict
  - checks the authoritative source
  - does not send the draft
  - records uncertainty and correction
  - escalates if policy ownership is unclear
rating:
  - NOT_OBSERVED
  - NEEDS_SUPPORT
  - DEMONSTRATED_FOR_THIS_SCENARIO
retention:
  store: scenario version, observation, support action, reviewer
  exclude: live employee data, unnecessary prompt transcript

Use known errors rather than trivia. Ask the learner to handle an apparently confident answer from an obsolete policy, a request to paste sensitive data into an unapproved model, a candidate ranking that conflicts with job criteria, a finance narrative that does not tie to the workbook, or an agent action outside approved scope. Observe whether they check, challenge, stop, override, escalate, and document.

Route gaps to the control that can fix them. A person who misses a data boundary may need a clearer tool banner or access restriction, not another lecture. A manager who cannot review because the queue is too large exposes a staffing or workflow problem. Repeated uncertainty about the authoritative source suggests poor information architecture. Literacy is partly a property of the work system.

Use Human Goes First where independent judgment still matters

Human Goes First is a sequencing control: withhold the AI recommendation until the person has committed to an initial frame, prediction, calculation, source choice, diagnosis, or decision. Then reveal the AI output and require comparison, critique, and revision. This reduces anchoring and deliberately exercises the skill that routine delegation would otherwise remove. It is not a ban on AI and it should not be applied to every low-consequence task.

A September 2026 AI Magazine article reports a small controlled navigation study in which this cognitive forcing function prevented the performance collapse observed without the intervention and produced an upskilling result. The author explicitly cautions that the study is small and that planned workplace designs were not implemented. Use it as a promising design pattern, not as proof that the same effect will appear in HR, finance, or another domain.

Pair the pattern with bounded independent and degraded-service drills. An independent check asks whether a still-required skill transfers without the assistant. A degraded-service drill asks whether the workflow can continue safely when the model, connected source, or automation is unavailable. Frequency should follow consequence and skill decay risk: quarterly may be reasonable for a high-consequence exception path, while an annual tabletop may suit a rarely used recovery procedure. Permanent no-AI work creates friction without evidence.

skill_preservation_contract:
  workflow: hr_policy_guidance
  critical_skill: authoritative_source_selection
  why_still_required: "AI may cite an obsolete policy; humans own release"
  normal_sequence:
    - human_states_question_and_selects_source
    - ai_drafts_from_approved_material
    - human_compares_and_records_final_decision
  independent_check:
    task: resolve_controlled_policy_conflict_without_ai
    cadence: quarterly
  degraded_service:
    fallback: approved_policy_index_and_escalation_queue
  evidence:
    - source_version_selected
    - discrepancy_detected
    - escalation_used
  prohibited_use: automatic_employee_scoring
  owner: people_operations_control_owner

Design these exercises as development and resilience controls. Announce their purpose, use synthetic or redacted fixtures, provide accessible alternatives, and record behavior at the narrowest useful level. Do not surprise-monitor employees, infer character from one task, or convert a scenario label into disciplinary automation.

Measure capability over time without pretending one score proves causality

Start with a baseline before the workflow changes where possible. Compare like-for-like controlled tasks under assisted and independent conditions, then repeat a delayed check after the novelty and immediate coaching have faded. Track the decision behaviors the role needs, not a generic “AI readiness” number. Segment results by scenario, role, system version, support provided, and accessibility condition so a local design defect does not become a judgment about a person.

Microsoft-led longitudinal research on cancer specialists describes “intuition rust”: operational gains that can coexist with gradual dulling of expert judgment. It is a high-stakes domain study, not a universal effect size. Nature's September review similarly concludes that current evidence is mixed and too limited for sweeping claims. The practical response is disciplined local measurement: look for deterioration, improvement, transfer, hidden work, and recovery under the conditions that matter to your own workflow.

Worked example: three roles using the same enterprise assistant

Assume an organisation provides an enterprise assistant to Marketing, People Operations, and Finance. The tool contract is approved, but access alone does not define the measures. Each role supplies different information, creates different outputs, and can affect different people.

All three roles receive baseline measures, but the evidence differs. The Marketing writer demonstrates claim verification and data boundaries. The People Operations partner demonstrates authoritative-source selection, privacy, fairness awareness, and escalation. The analyst demonstrates calculation tie-out, separation of observed variance from causal hypothesis, and reviewer handoff.

For the People Operations partner, first ask for the applicable policy version, affected population, unresolved facts, and proposed escalation without showing the AI draft. Then reveal a plausible summary grounded in the stale version. The participant compares the two, corrects the record, explains the difference, and decides whether the communication can proceed. A later independent scenario tests source selection with the assistant unavailable.

Suppose the partner misses the stale policy but correctly protects personal data. Record both observations. Assign a source-version job aid, add a visible authoritative-policy link in the tool, provide supervised practice, and retest the exact gap. Check whether the original failure was caused by training, interface hierarchy, ambiguous ownership, or workload. Do not declare the person “AI illiterate,” and do not use a composite score to hide the specific behavior.

Build a package another reviewer can understand without the AI chat

1. Scope record: entities, business units, jurisdictions, workforce groups, contractors, exclusions, legal interpretation, and accountable owners.
2. AI system inventory: exact systems, purposes, users, data, connected actions, affected people, decision influence, and versions.
3. Role-exposure matrix: actual tasks, knowledge and experience considered, consequence, objectives, supervision, override, and escalation.
4. Critical-skill register: still-required human capabilities, failure or recovery reason, Human Goes First points, practice cadence, owner, and retirement criteria.
5. Baseline and comparison plan: comparable assisted, independent, delayed, and degraded-service conditions with accessibility and interpretation limits.
6. Measures register: content, version, audience, delivery method, owner, rationale, accessibility, language, completion, and support.
7. Scenario library: controlled inputs, expected observable behavior, version, reviewer guidance, data handling, and limited result labels.
8. Exception register: missing groups, access barriers, gaps, incidents, owner, interim containment, remediation, due date, and retest.
9. Feedback and outcome evidence: questions, support demand, error patterns, overrides, incident linkage, and workflow improvements without false attribution.
10. Invisible-work record: validation, correction, context, exception, and repair work, including who performs it and whether capacity plans recognize it.
11. Approval and refresh: named reviewers, qualifications or roles, unresolved questions, decision, expiry, and event-based triggers.

Minimize personal data. Completion records may need employee identity, but scenario evidence often needs only a participant reference, role, date, observed behavior, support action, and reviewer. Follow records, labor, privacy, and security requirements. Do not retain live prompts merely because the AI platform makes them easy to export.

Keep official sources with access dates because the implementation landscape changes. The August 2026 Commission Q&A reflects amendments that entered into force in mid-July 2026 and explains current enforcement. National guidance, authority designations, and case practice may add relevant detail. Set a legal-review trigger rather than freezing today's summary into permanent policy.

Failure modes to test before approving the programme

Implement the control in 30 days without boiling the ocean

1. Days 1-4: select one business unit, appoint HR/L&D, business, AI governance, and legal/privacy owners as applicable, and approve the scope method.
2. Days 3-8: validate the AI system and people map. Identify two to four capabilities that remain necessary for review, exception handling, explanation, or recovery.
3. Days 7-12: build the critical-skill register and collect a fair independent baseline using controlled cases. Record known limitations and accommodations.
4. Days 10-17: adapt guidance and job aids, then design three to five scenarios. Put Human Goes First before material source, framing, calculation, diagnosis, or judgment steps.
5. Days 16-22: run assisted and independent conditions with accessible alternatives. Collect challenge-safety feedback and time spent validating, correcting, adding context, and handling exceptions.
6. Days 21-27: route gaps across training, interface, access, workflow, staffing, supervision, policy, and vendor management. Run one degraded-service drill and retest material failures.
7. Days 26-30: assemble the evidence index, skill and workload risks, exceptions, qualified-review questions, approvals, expiry, and refresh triggers. Decide approve, approve with actions, hold, or redesign.

Measure control health rather than claiming a direct productivity payoff. Useful indicators include inventory completeness, in-scope access, accommodation fulfillment, scenario gap type, remediation closure, source-check behavior, escalation accuracy, incident recurrence, policy questions, and overdue refreshes. Do not use “number of prompts” as a literacy metric.

Frequently asked questions

Does Article 4 require every employee to pass a test?

The Commission's current Q&A says no: Article 4 does not require measuring each employee's knowledge or guaranteeing a specific individual level. An organisation can still use proportionate scenarios as development and control evidence, with qualified legal review.

Who belongs in scope?

Start with staff and other persons who operate or use AI systems on the organisation's behalf, then map actual roles, systems, contracts, jurisdictions, and affected people. Ask qualified reviewers to confirm contractors, vendors, and special cases.

Is a completion certificate enough?

It proves that an event was completed. A stronger package explains why measures fit the role and system and includes the inventory, objectives, content, delivery, support, selected scenario evidence, exceptions, remediation, approvals, and refresh triggers.

Can HR use AI to score the scenarios?

AI may help structure evidence under an approved rubric, but a named human should review material observations. Do not let a model create an opaque employee score, infer traits, or make disciplinary decisions.

How can HR test whether AI is weakening a critical skill?

Define the exact still-required behavior, collect a fair baseline, compare assisted and independent performance on controlled cases, run a delayed check, and test recovery when the tool is unavailable. Treat a change as a signal to investigate training, interface, workload, supervision, and workflow design—not as automatic proof that AI caused it.

Should employees regularly work without AI?

Only in bounded, role-relevant checks where independent capability or outage recovery is genuinely required. Human Goes First preserves practice without discarding AI's value. Permanent no-AI tasks can become performative work and should not be the default.

Does this workflow certify compliance?

No. It prepares operational evidence and questions for human review. Article 4 enforcement is fact-specific and national implementation matters. Obtain legal advice for the organisation's circumstances.