How often should AI literacy training be refreshed?
The law sets no interval. Refresh when something material changes: a new tool or feature enters use, people change role, or the rules move, as they did when the Digital Omnibus rewrote Article 4 in July 2026. In practice that means a short review of the inventory every six to twelve months and a follow-up session when it turns up a gap. A refresh is a new dated entry in the same record, which is what keeps the record credible.
Why is there no fixed interval in the law?
Article 4 describes a duty to take measures that support AI literacy, calibrated to the people, the systems and the context of use. Context changes, so the measure has to be revisited, but the article does not say when. The Commission's Q&A on AI literacy does not set a frequency either. It does say that organisations can keep an internal record of trainings and other guiding initiatives, and a record whose only entry is two years old shows nothing done since. The AI Board is due to adopt recommendations under Article 4(3), including common objectives, and those may say more about frequency. Nothing has been published yet.
What does learning research say about follow-up?
It supports a short follow-up over one long session a year. In 2015 Murre and Dros replicated Ebbinghaus's forgetting curve and got results similar to the original: what was learned fell away fastest soon after learning. The study had one subject memorising lists of nonsense syllables, so it gives no figure for how fast people forget a workplace session, and we would not quote one.
What counts as a material change?
- A new tool or feature. Vendors switch AI features on by default in updates. If the CRM gained a summariser or the HR platform gained a ranking feature since the last session, the people who use them have a new system in scope.
- A change of role. A user who becomes a configurer, or a team lead who now signs off AI-assisted decisions, has moved into a group that needs more depth.
- New joiners. Anyone who arrived after the last session has had no measure at all. A catch-up for new hires is the most common reason for a follow-up.
- A change in the rules. Regulation (EU) 2026/1744 rewrote Article 4 with effect from 27 July 2026, moved most Annex III high-risk obligations to 2 December 2027, and left the Article 50 transparency duties on 2 August 2026. Guidance from the Commission and from Irish regulators will keep arriving.
- An incident. A confidential document pasted into a public tool, or a ranking acted on without review, is a reason to refresh the affected team now, not at the next scheduled review.
What a sensible cycle looks like
| When | What | Record entry |
|---|---|---|
| Month 0 | Baseline session for everyone, depth for configurers and overseers | Full record: inventory, sessions, attendance, calibration note |
| Months 3 to 6 | Ninety-minute follow-up: what changed, new hires, a clinic on real situations | Dated entry with attendance |
| Every 6 to 12 months | Inventory review by the named owner: new tools, changed roles, new guidance | Updated inventory, note of any gap |
| On a material change | Targeted session for the affected group | Dated entry |
Who owns the cycle?
One named person. After the calibration note, the most useful line in the record is the name of the person who will look at it again and the date they will do it. Our two-day tier ends with that name and date agreed, and the follow-up session exists to make the second entry happen.
Sources
- Regulation (EU) 2024/1689, the AI Act, Article 4, as amended by Regulation (EU) 2026/1744, for the July 2026 rewrite and the revised application dates.
- European Commission, AI literacy: questions and answers, for the record passage.
- Murre and Dros, Replication and Analysis of Ebbinghaus' Forgetting Curve, PLOS ONE, 2015.
This is general information about the EU AI Act, not legal advice. Enforcement is proportionate and case by case, and your sectoral regulator may issue guidance that applies to you.
