Most organizations responding to artificial intelligence (AI) in the workplace are asking the wrong question. The question is not whether to allow AI in training. It is whether your training was actually working before AI made the flaws visible. Having run technical training programs and a cybersecurity select-train-place company, I understand where the instinct to restrict AI comes from. I made the same call initially. What I want to share is why I reversed it, and what I learned in the process.
How AI Exposed Weaknesses in Technical Training
When AI tools became widely accessible in 2023, training programs faced the problem of learners using AI to complete assessments without engaging with the material. Assignments that should have taken hours were being submitted in minutes. The outputs looked plausible, but the understanding behind them was not there. When I probed learners in follow-up conversations, the gaps were obvious.
Our first response was to restrict access, detect misuse and enforce the old rules harder.
It did not work. And more importantly, I was solving the wrong problem. The issue was not that learners were using AI. The issue was that our assessments were designed in a way that made abuse indistinguishable from genuine learning. If a tool can complete your exercise without understanding it, the exercise is not measuring what you think it is.
That realization forced a more uncomfortable decision. Rather than fight AI adoption, I would become AI native. That meant revisiting curriculum design, rethinking assessment parameters and being honest with ourselves that the training model the industry had inherited was already inadequate before AI made its weaknesses visible.
Why Traditional Technical Assessments Fail
Bloom’s taxonomy has been in every learning and development (L&D) professional’s vocabulary since 1956. Remember, understand, apply, analyze, evaluate, create. Most of us know it. Most of our programs do not reflect it. If you audit the actual assessed work in most technical training curricula, the weight sits at the bottom. Recall a concept. Define a term. Pick the right answer from four options.
A well-prompted language model can pass most traditional technical assessments because those assessments were never really measuring technical capability. They were measuring access to information, which was always the wrong proxy for skill.
The answer is not to restrict the tool. It is to stop designing tasks that the tool can complete in your learner’s place.
AI Skills Training Still Requires Strong Fundamentals
Before going further, something important needs to be said. Becoming AI native does not mean replacing understanding with shortcuts. It’s important organizations still teach theory. They should still require learners to manually implement foundational concepts from the ground up. That baseline is not optional, and I do not treat it as such. You cannot meaningfully direct, evaluate or correct AI output on a topic you have never worked through yourself. The hands-on foundation is precisely what makes AI a useful tool rather than a crutch.
AI changes what happens after that foundation is in place. Once a learner understands a concept well enough to implement it manually, AI becomes a tool for accelerating the next layer of development, not for skipping the first one.
Moving From Knowledge Checks to Real-World Challenges
One of the core shifts the industry should make is updating the curriculum in how we frame new assessed work. Businesses are moving away from questions that ask learners what something is, toward challenges that ask learners to make something work.
New exercises in our program are structured around functional outcomes. Think of it like a practical exam where the only thing that matters is whether your solution actually works, not whether you can explain the theory behind it. Knowing the theory is not enough. There is nowhere to hide.
What this does to AI is interesting. When the task is definitional, AI can complete it entirely. When the task requires building a working solution to a constrained problem, AI becomes something the learner has to direct, question and correct. They have to recognize when the output is wrong or inapplicable to their specific environment. That is judgment. That is the upper half of Bloom’s taxonomy showing up in practice, not just in a learning objective slide.
A learner who has memorized the right vocabulary can pass a multiple choice exam. They cannot pass an exercise they do not understand how to approach.
Teaching Learners to Use AI as a Tutor
The second shift is in how we teach learners to use AI on their own time.
Most organizations that allow AI treat it like a faster search engine. That is a productivity gain, not a learning gain. Those are different things, and conflating them is part of how L&D ends up measuring the wrong outcomes.
It’s important to teach learners to use AI as a Socratic tutor. They prompt it to generate questions on a topic, answer those questions themselves and then prompt it again to analyze their answers and surface where their reasoning breaks down. It is a feedback loop that adjusts to where each learner actually is, which no single instructor managing a full cohort can replicate at that resolution. In our cohorts, learners who used this approach consistently arrived at practical exercises with a clearer sense of where their gaps were, which made instructor time more targeted and less remedial.
What I find interesting about this model is that it makes AI harder to use, not easier. Asking AI for an answer takes no effort. Asking AI to challenge your understanding, sitting with an answer that does not hold up and working through why is genuinely demanding. It is also the kind of practice that actually builds durable knowledge.
What AI Adoption Revealed About Workforce Readiness
As the sector evolves, I noticed something that has stayed with me. Learners who embraced AI tools early, not to shortcut the work but to extend it, tended to show stronger fit for the job market and smoother transitions into their careers.
My interpretation is that the ability to adopt and integrate new tools quickly has always been a marker of professional adaptability. I and countless others just underestimated how much it matters. In a field like cybersecurity, where the threat landscape shifts faster than any curriculum can track, the learner who knows how to learn with whatever tools are available is more valuable than the learner who has mastered a fixed set of skills. AI adoption in training appears to be an early signal of that broader capability.
Why Banning AI in Training Backfires
There is a real opportunity here for L&D functions that are willing to take it seriously. Redesign around the assumption that AI covers the bottom of the taxonomy. Build assessment and coaching around the top. That means rewriting objectives, redesigning exercises and having harder conversations with business stakeholders about what readiness actually looks like. It is more work than refreshing a slide deck.
It is also the only version of technical training that holds up when the environment stops cooperating. Restricting AI keeps it out of the one context where you could teach people to use it well. Those same learners graduate into organizations where AI is woven into daily work, having spent their training in an environment that pretended otherwise. That is not rigor. It is a preparation gap dressed up as a standard.
What Technical Skills Training Should Look Like Now
The problem is not that learners use AI to avoid thinking. The problem is that we continue to build systems where the thinking we measure is not the thinking that matters. AI just made that easier to see.
Fix the task design. Teach people to use AI as a tutor, not a shortcut. Ground them in the fundamentals first, then teach them to go further with the tools available. The taxonomy has not changed in seventy years. What counts as skill has.
