Three years ago, ChatGPT launched and changed how we work.
Today, 30% of European workers use AI daily. Only 15% have received any training.
We have been experimenting in the dark. It is time to turn the lights on.
The Experiment We Are All Running
Maria is a marketing manager at a mid-sized company in Bucharest. She is sharp, capable, and, like millions of professionals across Europe, she discovered ChatGPT in early 2023.
At first she used it tentatively. A subject line here, a quick rewrite there. Then she grew bolder. Draft emails. Blog outlines. Competitive analysis summaries.
Nobody taught her how to do this. There was no training session, no company policy, no course. She figured it out herself, through trial and error, through late nights, and through a considerable pile of deleted outputs that did not quite work.
Maria is not unusual. She is typical.
Across the European Union, 30% of workers now use AI tools at work. That figure comes from a 2024 survey by the Joint Research Centre covering 70,316 workers across all 27 member states. Nearly one in three of us, interacting daily with systems that can write, analyse, summarise, and create.
A separate survey by Cedefop, the European Centre for the Development of Vocational Training, found that only 15% of workers in eleven EU countries have participated in any AI-related training in the past year. Forrester’s research paints a similar picture: just 39% of European employees have received any formal AI training, compared to 52% in the United States. We are not merely undertrained. We are undertrained relative to our competitors.
Think about what this means.
We have built and deployed some of the most powerful cognitive tools in human history. We have given them to hundreds of millions of workers. And then we said: “Figure it out yourself.”
This is the experiment we are all running, whether we realise it or not.
And we are running it in the shadows. A survey by Salesforce found that 55% of employees use AI tools their companies have not approved. Forty percent use tools their companies have actively banned. Two thirds have used AI output without properly checking whether it was accurate.
We have not just adopted AI without training. We have adopted it without guardrails, without visibility, without any real understanding of what is happening inside our own organisations.
What Happens When We Do Not Know What We Are Doing
There is a study that keeps me up at night.
In 2023, researchers from Harvard Business School and Boston Consulting Group ran an experiment with 758 consultants. Not students. Not novices. Skilled knowledge workers at one of the world’s top consulting firms. The researchers wanted to understand what happens when professionals use AI for real work.
The results were stunning, in both directions.
When AI was suited to the task, the gains were remarkable. Consultants using AI completed 12% more tasks, finished 25% faster, and produced results more than 40% higher in quality. A separate study by Harvard and the Federal Reserve found knowledge workers were 33% more productive in each hour when using generative AI properly. Lower-performing consultants saw the biggest improvements. Their output quality rose by 43%. AI was not just helping. It was equalising. Lifting everyone up.
But here is the catch. When Cedefop asked AI users directly whether the technology had made them better at their jobs, 54% said no. Not because the tools are weak. Because nobody taught them how to use the tools well. The potential is there. The training is not.
And then there was the other finding. The one that should concern every professional and every organisation.
When AI was poorly suited to the task, it made smart people worse.
The researchers designed one task that fell outside AI’s current capabilities. A business problem that looked similar to the others but required different reasoning. For this task, consultants using AI performed 19 percentage points worse than those working without it.
I want to make sure you understand what happened. These were elite consultants. They had AI assistance. And they performed significantly worse than colleagues working alone with just their own minds.
The AI did not simply fail to help. It actively degraded their judgment.
The researchers gave this phenomenon a name: the “jagged technological frontier.” AI capabilities, they found, do not form a smooth line where more complex tasks are harder and simpler tasks are easier. Instead, the frontier is jagged. AI excels brilliantly at some surprisingly difficult tasks and fails completely at some surprisingly simple ones.
And there is no obvious way to tell which is which.
Maria, our marketing manager in Bucharest, does not know where the frontier is. Neither do her colleagues. Neither, frankly, do most of the executives making decisions about AI deployment. We are all navigating a landscape with invisible cliffs. We are doing it without a map.
The Confident Wrong Answer
Here is an example you might recognise.
Imagine you ask an AI to summarise a quarterly report for a board presentation. The report is 47 pages of financial data, market analysis, and strategic recommendations. You upload it, ask for a one-page executive summary, and within seconds you have exactly that. A crisp, well-written summary with clear conclusions.
You read it. It sounds authoritative. The language is polished. The structure is logical. You are running late for the meeting, so you paste it into your slide deck and move on.
Here is what you did not notice.
The AI combined two separate revenue figures from different sections, creating a number that appears nowhere in the original report. It attributed a strategic recommendation to the CFO when it actually came from an external consultant. And it confidently stated that Q3 showed “continued growth” when the report actually described a 3% decline that was offset by cost reductions.
None of these errors are obvious. The summary reads beautifully. It is wrong in ways that are almost impossible to detect without reading the original 47 pages yourself. Which defeats the purpose of the summary.
This is not a hypothetical. Variations of this happen thousands of times every day, across every industry, in every country.
The AI is not lying to you. It has no concept of truth or falsehood. It is doing exactly what it was designed to do: produce fluent, confident text that pattern-matches against what summaries typically look like. The fact that this particular summary contains fabricated details is not a bug from the AI’s perspective. It is simply what happened.
And without training, without a systematic method for catching these errors, we have no reliable way to know when we are holding gold and when we are holding fool’s gold.
The European Moment
On February 2, 2025, Article 4 of the EU AI Act took effect. This single article, just a few paragraphs of legal text, creates an obligation that affects every organisation operating in the European Union.
Here is what it says, in essence. If you deploy AI systems, you must ensure that your staff have “sufficient AI literacy.”
Not recommended. Required.
The European Commission has been explicit. In their official guidance, they stated: “Relying on instructions of use is NOT sufficient.”
Read that again. You cannot simply hand employees an AI tool and point them to the help documentation. You cannot assume they will figure it out. You need actual evidence that your people understand how to use AI safely, effectively, and responsibly.
What does “sufficient” mean? The Act deliberately leaves this flexible, but it specifies that literacy must be appropriate to the technical knowledge and experience of the staff, the context in which the AI systems are used, and the people or groups affected by those systems.
A customer service representative using an AI chatbot needs different training than a data scientist building recommendation algorithms. A healthcare administrator using AI for scheduling needs different training than a radiologist using AI for diagnostic assistance.
One-size-fits-all will not work. Generic “intro to AI” videos will not satisfy the requirement. You need a framework that adapts to roles, contexts, and risk levels.
Enforcement begins August 2, 2026. Eighteen months from when Article 4 took effect. Eighteen months to train entire workforces on skills that most organisations have not even defined yet.
The penalties are sobering. AI literacy violations under Article 4 are not directly fined at the highest levels, but they serve as an aggravating factor for any other violation. If your employee misuses a high-risk AI system because they were not properly trained, the consequences for that misuse multiply. For serious violations involving prohibited practices, fines can reach €35 million or 7% of global annual turnover, whichever is higher.
This is not a distant regulatory concern. This is a countdown.
A View from Romania
I live and work in Romania. It is worth sharing where I am writing from, because it illustrates the challenge perfectly.
According to Eurostat’s December 2025 data, 20% of EU enterprises now use AI technologies, up from 13.5% in 2024. That is genuine progress across the continent.
The distribution, however, is wildly uneven.
Denmark leads the EU at 42% enterprise AI adoption. Finland follows at 37.8%. Sweden at 35%.
Romania stands at 5.21%. We rank 27th out of 27 member states. Dead last.
When I first saw this number, I felt a familiar mixture of frustration and recognition. Romania has 226,000 IT workers, one of the highest concentrations of tech talent in Europe. We have engineers building AI systems for companies around the world. And yet our own enterprises barely use the technology.
There is a paradox here, though, and it gives me hope. According to a report on CEE small and medium enterprises, 61% of Romanian employees actively seek ways to use AI at work. That is one of the highest rates in Central and Eastern Europe.
The curiosity is there. The hunger to learn is there. What is missing is structure.
When researchers asked Romanian companies about their barriers to AI adoption, 57% cited lack of know-how as their primary obstacle. That is the highest rate in Central and Eastern Europe after Bulgaria. One quarter of Romanian firms have taken no steps whatsoever to train employees in AI. Only 8% believe they would be ready for a compliance audit.
The digital foundation is shaky too. Romania scores 30.6 on the EU’s Digital Economy and Society Index, against an EU average of 52.3. Only 28% of Romanians have basic digital skills, compared to 54% across the EU. We are not starting from zero, but we are starting from behind.
People want to use AI. They are already experimenting, often without guidance. They know something powerful is available. What they do not have is a framework. A systematic way to know when to use it, how to use it well, and how to verify that it is actually helping.
This gap between individual curiosity and organisational capability is where the risk lives. Employees experimenting. Companies not guiding. Regulators preparing to enforce.
Romania is not alone in this. We have simply made the problem more visible by ranking last. The same gap exists across Europe, in different proportions, with different statistics, but with the same fundamental challenge.
For context: Romania has not yet designated a national AI authority despite the EU deadline of August 2, 2025. Our draft AI legislation was rejected by both the Senate and Chamber of Deputies. A new draft is under review, but there is no timeline.
We are not ready. But we are also not unique. Most member states are behind.
What Already Exists
Before I describe what I have built, I want to acknowledge what is already out there. The AI literacy landscape is not empty. Some excellent resources exist, and I recommend them.
Elements of AI from the University of Helsinki has educated over 1.8 million people globally. It is free, accessible, available in dozens of languages, and does a remarkable job explaining what AI is and how it works at a conceptual level. If you have not taken it, you should.
Anthropic, Google, Microsoft, IBM, and OpenAI all offer training resources. Anthropic’s courses teach prompting techniques. Google’s AI Essentials covers fundamentals. Microsoft’s LinkedIn Learning paths are comprehensive. These are valuable, especially for understanding specific tools.
The KI-Campus “EU AI Act Essentials” course from TU Munich focuses specifically on the regulatory requirements. It is the best free resource I have found for understanding what the law actually requires.
I recommend all of these. They are well-made and useful.
They also share a common limitation. They exist in silos.
One course teaches you what AI is. Another teaches you how to prompt. Another teaches you what the law requires. None of them give you an integrated framework for the actual workflow of AI-assisted work. None guide you from deciding whether to use AI for a task, through setting it up correctly, to verifying the output and measuring whether it helped.
No existing training I have found gives you a systematic framework for deciding when NOT to use AI. They all assume you have already decided to use it. They just teach you how. But the Harvard/BCG study showed that the most important skill might be knowing when to step back.
Then there is verification. Most training mentions “check the output” but does not teach you how to do it systematically. Where is the methodology for catching errors before they cause harm?
And measurement. How do you know if AI is actually helping you, versus just producing output that feels productive? If you cannot measure the value, you cannot improve the process. And you cannot prove compliance.
The skills and the compliance requirements exist in separate courses, taught by different people, with no connection between them. But in real work, they are inseparable. Every AI task has both a skill dimension and a responsibility dimension.
We need something that ties it all together.
What I Have Built
Over the past year, I have developed a framework called SPARKS-SIX™. It is not a tutorial for a specific AI tool. Those become outdated the moment a new model launches. It is an operating system for AI-assisted work that applies regardless of which tool you are using.
The name comes from the structure. Six pillars, each represented by a letter that spells SPARKS.
SCOPE: Should we use AI for this task?
This is where most training fails to start. Before you write a single prompt, you need to decide whether AI is the right choice for this particular task.
SCOPE walks you through four questions. Is there a capability match, based on what you know about this AI? Is there genuine value, or will you spend more time prompting and fixing than doing it yourself? What is the real-world damage if the AI gets this wrong? And what kind of partnership do you need: automation where AI does it for you, augmentation where you work together, or agency where AI works independently?
This is how you navigate the jagged frontier. You pause before each task and ask whether it is likely to be inside or outside AI’s capabilities.
PROMPT: How do we ask clearly?
Once you have decided to use AI, the quality of your instruction determines the quality of the output.
PROMPT teaches you to structure your requests in four parts. Give the AI a role: “Act as an experienced financial analyst…” State your goal: “Identify the three largest expense categories…” Show what good looks like: “For example, here is how I would describe a category…” And set boundaries: “Limit to 200 words. Use bullet points. No jargon.”
This is not about clever tricks or “hacks.” It is about clear communication. The same skills that make you effective working with humans, applied to working with AI.
ARCHITECT: How do we set AI up for success?
Most AI failures happen before you even start prompting. The AI simply does not have the information it needs.
ARCHITECT teaches you to ask a simple question: Does this task require the AI’s general knowledge, a specific document, or current information? If you are asking about your company’s Q3 report, the AI needs to see that report. It cannot answer from general knowledge. If you are asking about news from this morning, the AI needs web access. Its training data is months or years old.
This pillar also includes the Agent Canvas, a one-page template for designing complex AI workflows with clear goals, instructions, knowledge, and tools.
REVIEW: Is this answer actually correct?
Remember Maria and her board presentation summary? This pillar is about not becoming Maria.
REVIEW teaches the Four C’s, a quality check you can run in sixty seconds. Is it Correct? Do the facts, names, and numbers actually check out? Is it Coherent? Does it make logical sense, or does it just sound good? Is it Complete? Did it answer everything you asked? And is it in Context? Is the tone and style appropriate for your audience?
There is also what I call the Golden Question: “Show me your work.” Ask the AI to cite sources, quote specific passages, or explain its reasoning. Hallucinations often collapse when challenged.
When output is bad, the Fail-Tree Analysis helps you diagnose why. Was the information wrong? That points to an Architect problem. Was the instruction unclear? That points to a Prompt problem. Or was this simply a task the AI cannot do reliably? That points back to Scope.
KEEP-SAFE: Is this safe, fair, and compliant?
This is where AI skills meet AI responsibility. And where EU AI Act compliance lives in daily practice.
KEEP-SAFE starts with what I call the “Public Computer” Rule. Treat any AI tool without explicit company approval like a shared computer at the library. Would you paste confidential client data into a public computer? Then do not paste it into ChatGPT.
The pillar includes a five-point responsibility checklist covering privacy, bias, copyright, ownership, and transparency. Quick mental checks before you start any AI-assisted task.
And it builds the habit of disclosure. “This draft was prepared with AI assistance and reviewed by the author.” Transparency is not just ethical. It is increasingly expected.
SCORE: Did it actually help?
If you cannot measure the value, you were just playing.
SCORE teaches what I call the Before & After Test. Time how long a task takes without AI. Then time the complete AI-assisted process, including prompting, reviewing, and fixing. The difference is your actual value, not your imagined value.
The pillar introduces the Ship, Tweak, or Park decision. If results are clearly positive, ship it and make this a regular practice. If promising but imperfect, tweak it and measure again. If it is simply not helping for this task, park it. Stop using AI here and try elsewhere.
This closes the loop. You do not just use AI and hope for the best. You measure, learn, and improve.
The First Five Minutes
Before any of the six pillars, there is a foundation. The First-Contact Protocol.
When you encounter any new AI tool, or an existing tool that has been significantly updated, you need to understand what you are working with. The First-Contact Protocol is a five-minute “interview” process.
Start with a Personality Test. Give it a creative prompt, a logic prompt, and a summarisation prompt. See where it shines and where it struggles. Then do a Library Card Check. Ask if it can access the internet. Ask what it knows about recent events. Understand the boundaries of its knowledge.
Next, run what I call the “Show Me Your Hands” Test. Can it use tools? Accept file uploads? Generate images or code? Know its capabilities before you depend on them.
Finally, conduct a Security Briefing. Is this conversation private? What is the data policy? Are outputs yours to use commercially?
Five minutes. It saves hours of frustration and risk.
You would not hire a colleague without understanding their strengths, limitations, and working style. Why would you work with an AI any differently?
What This Series Will Cover
Over the coming weeks, I am publishing the entire SPARKS-SIX™ framework through GenAI School. Free. Complete. Every concept, every checklist, every example.
This is not a teaser or a lead magnet. This IS the framework.
We begin with the First-Contact Protocol: the five-minute process for understanding any AI tool before you trust it with real work. Then we move through the six SPARKS pillars.
SCOPE teaches you to navigate the jagged frontier, making smart Go/No-Go decisions about when to use AI and when to work alone. PROMPT covers the art of clear instruction, with before-and-after examples you can try immediately. ARCHITECT shows you how to give AI the knowledge it needs, from uploading files to designing workflows with the Agent Canvas.
REVIEW teaches the 60-second quality check that catches hallucinations. KEEP-SAFE brings the EU AI Act into your actual workflow, covering privacy, bias, disclosure, and the habits that create compliance evidence. SCORE shows you how to measure what matters and prove that AI is genuinely helping.
The final article puts it all together. The journey from casual user to skilled AI director. Your Monday morning action plan.
Each article will be self-contained and useful on its own. Together, they form a complete curriculum.
Why Free
You might wonder why I am publishing this for free. No paywall, no “premium tier,” no upsell at the end.
The problem is too important and too urgent.
The EU AI Act deadline is August 2, 2026. Millions of European workers need AI literacy training. Most organisations do not have it figured out. If I can help close that gap, even a little, by making this framework available to anyone who wants it, that matters more than keeping it proprietary.
I have spent 25 years in software development and over a decade building AI systems in production. I have seen what happens when powerful tools meet unprepared users. I have cleaned up the messes. I would rather help prevent them.
There is also a personal dimension worth considering. PwC’s 2025 research found that workers with AI skills now command a 56% wage premium. That is more than double the 25% premium from just two years ago. AI literacy is not merely about compliance or productivity. It is rapidly becoming a career requirement.
Use this. Share it with your team. Adapt it to your context. If it helps, pass it on.
The Countdown
August 2, 2026 is coming.
On that date, enforcement of the EU AI Act begins in earnest. Every organisation deploying AI in Europe will need to demonstrate that their people know what they are doing.
Today, across Europe, 30% of workers use AI. Only 15% have received any training.
That gap cannot hold. Either we close it deliberately, through structured learning and intentional practice, or we close it painfully, through compliance failures and preventable mistakes.
I know which I prefer.
The tools are extraordinary. What we can accomplish with AI assistance genuinely deserves wonder. But wonder without discipline is just entertainment. To capture the value and avoid the risks, we need a framework.
That is what this series is about.
Welcome to GenAI School.
This is Article 1 of 10 in “GenAI School: AI Literacy for Everybody,” a complete, free course on working effectively with AI.
Follow for the next article, or share this with someone who needs it.
Next: The First-Contact Protocol