On 10 November 2026, UNESCO published AI Answers. You Decide., a guide to teaching media and information literacy in AI-mediated information environments. UNESCO introduced the guide during Global Media and Information Literacy Week, 24 to 31 October 2026. It is written mainly for classrooms. Its lesson reaches employers that count trained staff and call the count literacy.
The stakes sit with the person who signs. When someone with sign-off authority approves an AI-assisted decision without the ability to catch a flawed output, the error stops being the machine’s and becomes the organization’s accepted decision. A training count does not tell anyone whether that person can catch it.
A guide built for teaching judgment
The guide is a companion to UNESCO’s 2021 Media and Information Literacy curriculum, which UNESCO says was piloted and adapted in more than 40 countries. UNESCO says eight international experts developed the guide and ten additional specialists peer-reviewed it. It is designed primarily for educators, teacher-training institutions, and curriculum developers.
The guide covers five areas. One of them, AI, the human mind and knowledge, addresses critical thinking, informed trust, and the ability to evaluate AI-generated information. UNESCO says the guide examines how AI influences information, knowledge, human behaviour, and society, and does not center on specific technologies or technical skills. Mariya Gabriel, UNESCO’s Assistant-Director General for Communication and Information, said people must “retain the ability to think critically, question what they see and make informed choices.” The guide combines conceptual explanations with discussion questions and practical activities.
That is a sound design for teaching. Discussion questions and activities build a habit. They do not, by themselves, show that a particular person can catch a flawed AI output on a particular day.
What the EU text asks of employers
The AI literacy article of the EU AI Act, Article 4, appears in Regulation (EU) 2026/1744, the Digital Omnibus on AI, in this form: providers and deployers of AI systems must take measures to support the development of AI literacy among their staff and other people who operate or use AI systems on their behalf. The measures account for technical knowledge, experience, education, training, and the context of use. The same article says the obligation “does not require providers or deployers to guarantee any specific level of AI literacy of any individual.” It also tasks the Commission with publishing practical examples of compliance and the European Artificial Intelligence Board with adopting recommendations.
The legal duty in that text is to take measures, and the text stops short of requiring proof of an outcome for each person. That is a workable rule for a law that covers every provider and deployer in the Union. The check described below goes beyond what the article asks. It is a governance recommendation for the people whose signatures carry the most weight, and each organization decides what evidence of ability it keeps for them.
Counting trained staff is Responsible AI
Some of the work around a training program is machine checking machine. A learning platform’s automated quiz scores a learner’s answers, a dashboard compiles completion rates, and a monitoring tool reports usage. Those automated checks are Responsible AI. Their output is a number of people trained, and the number is useful for showing that measures exist.
AI Governance asks a narrower question about a specific person. When that person signs to accept, modify, or reject an AI-assisted decision, can they direct the work, challenge the output, verify the facts, and own the result? Assessing that ability, the governance competence at the center of AI literacy, is the work of an instrument the author calls the Human Enhancement Quotient. A completion rate cannot answer the question. A whole team can finish a course while the one employee holding sign-off authority approves a fabricated citation.
The gap sits between two counts. One counts staff who finished training. The other counts signers whose ability to reject a bad output was confirmed by a task they performed.
Four steps for the people who sign
- List the sign-off roles. Write down every role that signs off on AI-assisted decisions, such as claims, credit, hiring, clinical notes, filings, or published content. Name the current holder of each.
- Run a competence check before granting authority. Give the person a sample task with a seeded flaw in the AI output, such as a wrong figure, an invented source, or a conclusion the evidence does not support. The person must catch it, reject or modify the output, and explain why. A reviewer other than the trainer scores the result.
- Keep a record per signer. Record the check, the date, the result, and who scored it. The head of the function that grants the authority signs the record.
- Re-check on change. Repeat the check when the AI system changes or when the person moves into a different role.
The method pairs each fact with a tactic and a measured result, a discipline the author calls Factics. The fact is that a completion count shows participation and does not show whether a signer can catch a flawed output. The tactic is the seeded-flaw check. The measures are three:
- Share of sign-off holders with a recorded competence check. The target is every holder, checked before authority is granted.
- Share of sampled AI-assisted decisions where the signer modified or rejected the output. Review the sample monthly against an independent review of the same items, so a low rate can be read as sound outputs or as missed errors.
- Time from an AI system change to the signer’s re-check, in days.
What stays open
A seeded-flaw task tests a person on a planted error. Whether that performance holds on live decisions is the open question, and the independent sample review in the second measure is the way to find out. The Commission’s practical examples and the Board’s recommendations under Article 4 are the next texts to watch for how far the EU expects measures to reach the people who sign.
This column is general information and is not legal advice. Employers should ask qualified counsel what Article 4 and their national rules require of them.
Sources
- European Parliament and Council of the European Union. (2026). Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI). Official Journal of the European Union, L series. EUR-Lex. https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=OJ%3AL_202601744
- UNESCO. (2026, November 10). AI Answers. You Decide: UNESCO launches a new Media and Information Literacy curriculum. https://www.unesco.org/en/articles/ai-answers-you-decide-unesco-launches-new-media-and-information-literacy-curriculum
FAQ
What did UNESCO publish on 10 November 2026?
UNESCO published “AI Answers. You Decide.”, a companion guide to its Media and Information Literacy curriculum. It is designed primarily for educators, teacher-training institutions, and curriculum developers. It covers five areas, including critical thinking and the ability to evaluate AI-generated information.
Does Article 4 of the AI Act require employers to guarantee staff AI literacy?
No. In the text set out in Regulation (EU) 2026/1744, Article 4 requires providers and deployers to take measures to support the development of AI literacy. The text says the obligation does not require them to guarantee any specific level of AI literacy of any individual.
Why does course completion not show that a person can reject AI output?
A completion record shows that a person finished a course. It does not show that the person caught a flawed output when it mattered. A sample task with a seeded flaw tests that ability directly, because the person must find the flaw and reject or modify the output.
What is a competence check for a sign-off holder?
It is a short task in which an AI output carries a planted error, such as a wrong figure or an invented source. The person must catch the error, reject or modify the output, and explain the reason. A reviewer other than the trainer scores it, and the result is recorded.
What should an organization measure?
Three measures fit this problem: the share of sign-off holders with a recorded competence check, the share of sampled AI-assisted decisions where the signer modified or rejected the output, checked against an independent review of the same sample, and the days between an AI system change and the signer’s re-check.
Is this legal advice?
No. This column is general information based on the published texts. Ask qualified counsel what Article 4 and national rules require of your organization.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply