Most AI skills programmes follow a familiar formula: raise awareness, add knowledge, train the skills, and capability follows. I think that formula is wrong at every step when applied to rapidly expanding bodies of knowledge, and AI is where its weaknesses show fastest.
Awareness tells people a technology matters. It does not tell them what to do on Monday morning. Knowledge of a particular tool dates as fast as the tool changes. Skills taught on today’s interface become a liability when tomorrow’s version behaves differently. An organisation can run all three stages, record full attendance and still find its people stuck on the first problem that wasn’t in the course material.
Cyber security has lived with this reality for years. The field grows faster than anyone can learn it. No practitioner holds all of it, and hyper-specialisation is unavoidable. What separates capable people from merely trained ones is not how much they know, but whether they can approach an unfamiliar problem, work out what they need to learn, and just get on with it.
The alternative I argue for is appreciation plus fundamental principles. Appreciation means understanding what a technology is for, where it fits and what it changes about the work. Fundamental principles of AI, stated in plain English (or Welsh) are the parts that stay the same when the product or tool being used changes. For AI, that means how models are built, why they produce fluent answers with no guarantee of accuracy, and what happens to the information you give them. Learn those, and skills come quickly, because the learner has something to hang them on. Confidence is the result of that understanding, not a mindset to be coached in separately.
That confidence matters more with AI than with most technologies. AI applications look impressive and hide how they work. Some people respond by treating them as a dark art and keep well away. Others accept whatever comes out without question. Both leave them and their organisation exposed.
The exposure is not only wrong answers. Staff paste sensitive information into public tools. Generated content goes out unchecked. AI agents are being given permission to act on systems and accounts. Someone who understands the principles will ask the right questions: where does this data go, what is this tool allowed to do, how would I know if it was wrong? Someone trained only on the interface does not. AI capability and cyber security are the same conversation, and the Welsh skills debate should treat them as one.
Judgement has to sit alongside confidence. A confidently phrased answer is not evidence that it is right. Asking a second model is not verification either; two systems can agree and both be wrong. Good training leaves people comfortable using AI and comfortable challenging it.
None of this happens in a single workshop. Capability is built through confidence gained on real problems, in environments where it is safe to get things wrong. Cyber learned this the hard way, through ranges, exercises and competitions such as Capture-the-Flag (CTF): learning by doing against realistic systems rather than slides. The same approach transfers directly to AI.
So, my ask for Wales is simple. Stop measuring AI skills by attendance and workshop completion. Measure whether people can identify a worthwhile use, assess an output, recognise a risk and decide what to do about it. Fund programmes that teach the fundamental principles of AI rather than just how to use products, and which build practice in rather than ending at the certificate. That is how we develop people who keep pace with the technology, rather than people trained on last year’s version of it.
John Davies MBE is co-founder and Chief Executive of SudoCyber. He co-founded Cyber Wales, the Welsh cyber security ecosystem, which he chairs, and is a doctoral researcher in War Studies at King’s College London.
