AI can write your course, but it can't tell you what's worth learning.

I used AI to help me build a learning programme last month. Not to write a first draft that I then edited into something usable. Not as a spell-checker or a brainstorming prompt. Actually, genuinely used it- to generate scenario branches, to stress-test the andragogical logic, to produce three versions of a facilitator guide in different tones, to translate a complex behavioural framework into language that would land with a non-specialist audience. It was fast. It was, in most respects, better than fine. It saved days. And it had absolutely no idea what the programme was actually trying to change.

That wasn't a failure of the AI. I hadn't told it. And I couldn't have told it- not in a prompt, not in a briefing document, not in any form that a language model could receive and act on- because the answer to "what is this programme trying to change" didn't live in text. It lived in months of observation, in the gap between what the organisation believed its people could do and what I'd watched them actually do under pressure, in a quiet but important pattern I'd noticed in how certain conversations went wrong and others went right. That knowing is not transferable to a model. It is, in the most literal sense, the thing I was hired for. When I tell people that I spent three years studying observation at post-graduate level, they often think I'm joking. I'm absolutely not (see my record of student debt as proof). The ability to look, and understand what you're seeing, is a future-proof skill that AI can't replace.

There's a version of the current conversation about AI and content production that I find concerning. Not the version where people worry that AI will replace human writers- that debate is mostly just noise, and the answer to it is "well, George, it depends entirely on whether the human writing was adding anything to begin with." The version that concerns me is subtler. It goes like this: we can now produce learning content at unprecedented speed and scale, which means we can solve the problem of learning at unprecedented speed and scale. This is not true, and believing it is a category error with real consequences- and not just for the sanity of your L&D partners.

Content production is not the bottleneck in learning, and I don't think it ever has been. It's just an easy way to express a complex challenge to stakeholders. The bottleneck is really the set of decisions that have to be made before a single piece of content is written- decisions about what needs to change, in whom, by how much, and how you'll know if it has. These decisions require deep contextual understanding, careful observation, and a kind of institutional knowledge that accumulates slowly and doesn't compress into a training set. AI is extraordinarily good at the production side of the equation. It is not in the business of the diagnostic side. And if we use its production capability to race past the diagnostic work- to skip the hard questions in favour of getting something built- we will produce learning content at speed and scale that is precisely as useless as the content we've always produced, but in much greater quantity. A cupboard full of date-expired cans.

Let me be specific about what I mean by the diagnostic work, because I think it's easy to wave at it vaguely and harder to describe what it actually involves. When I joined a new organisation earlier in my career and was asked to build a programme on a particular capability- commercial acumen, say, or inclusive leadership, or giving feedback- the first thing I did was not open a content tool or commission a needs analysis survey. The first thing I did was watch. Not in a structured observation sense, though that came later. Just watch. How did teams talk about the work? Where did conversations stall? What questions did they ask that revealed the shape of their understanding? When something went wrong, what did they reach for? When they were at their best, what were they doing?

This takes time. It requires physical presence, or at minimum genuine conversational access. It produces knowledge that is partially intuitive, partially structured, and extremely difficult to articulate in the form that would make it useful to a content generation tool.

And here is where I want to push back against a comfortable assumption: that observation is simply the act of looking. It isn't. Watching a team in a meeting and understanding what you're seeing are entirely different skills, and the gap between them is not bridged by attention or intelligence alone. It is bridged by years of work on yourself.

Effective observation in a learning context- the kind that produces reliable diagnostic judgement rather than a sophisticated projection of your own experience- requires you to understand your own involvement in what you're seeing. It requires you to have done enough work in experiential group settings to recognise when a dynamic in the room is pulling you toward a particular interpretation, not because the evidence supports it, but because it resonates with something in your own history. It requires enough bias training to know which of your instincts are genuine pattern recognition and which are pattern imposition- your assumptions wearing the costume of insight. It requires the kind of reflective or therapeutic practice that makes you genuinely curious about why you are drawing the conclusion you're drawing, and capable of holding that conclusion lightly enough to revise it when the evidence shifts. Without this, observation is not a diagnostic tool. It is a mirror; you will see what your own history and assumptions have prepared you to see, and you will build learning programmes that address the problem as you have perceived it rather than as it actually is.

The disciplines that build real observational capacity- experiential group work, reflective and therapeutic practice, facilitation training, bias and positionality work, clinical-style supervision- are not typically listed in L&D job descriptions. They are treated as personal development, adjacent to the professional role rather than central to it. I think this is a significant mistake. They are the foundation of the diagnostic capability that makes everything else in this work meaningful. The thing that AI needs you for. The content is downstream of the judgement. The judgement is downstream of the self-knowledge. And the self-knowledge is accumulated slowly, through experience that cannot be shortcut, in contexts that are specifically designed to surface the ways your own subjectivity interferes with your ability to see clearly.

The output of this diagnostic work is not a brief. It's a judgement: the gap is here, and it's caused by this, and what's needed is not more information but a different experience of the problem. That judgement shapes everything that follows- the format, the sequence, the tone, the scenarios used, the facilitation approach, the measurement design. It is, in a meaningful sense, the curriculum. The content is just the delivery mechanism.

AI can do the delivery beautifully, but it cannot form the judgement.

None of this is an argument for slowing down or being precious about craft. Quite the opposite. If AI can handle the production- and it increasingly can- that frees up human time and cognitive capacity for the diagnostic work that only humans can do. Instead of spending three weeks building a course, you can spend three weeks watching, listening, talking to people, understanding the real shape of the problem. Then you can spend a day building the course, with AI doing the heavy lifting on the content production, and have something better than you would have had if you'd spent all four weeks building it the traditional way. This is the reallocation that AI makes possible, and it's a genuinely exciting one. The learning professionals who thrive in the next five years will not be the ones who resist AI on the grounds that human-written content is inherently superior. Some of it is. Much of it isn't. They will be the ones who understand clearly which part of their work is irreplaceable and protect that part ferociously, while embracing AI's capability for everything else.

The irreplaceable part is the judgement. The knowing-what's-worth-learning. The ability to look at an organisation and identify not the capability gaps people report wanting to close, but the capability gaps that are actually limiting performance. These two things are frequently different, and the difference matters enormously. People are not reliable reporters of their own learning needs. This is not a criticism- it's a structural feature of the way expertise develops. When you don't know something, you often don't know that you don't know it. The gaps that people readily identify are usually the surface manifestation of something deeper. A manager who says "I need to get better at having difficult conversations" may actually need to develop their tolerance for relational uncertainty. A sales team who ask for better product knowledge may actually need to work on how they listen. These translations require a human in the room, paying close attention, over time- no prompt will surface them.

I want to make one more point about where this matters most, because I think it's underappreciated. The content that AI produces- fluently, rapidly, at scale- is content that reflects patterns in existing knowledge. It is synthesised from what has already been written, thought, and published. It is, by its nature, retroactive. It describes what has worked before, what has been observed, what is already understood. The most important learning design problems are prospective. They involve capabilities that don't yet exist in the organisation, for situations that haven't yet been encountered, in contexts that are changing fast enough that yesterday's best practice is already becoming tomorrow's liability.

A learning programme designed for how people need to work with AI in two years' time cannot be built primarily from what AI already knows about how people work with AI today. It requires human judgement about where things are going, what will matter, what the real stakes are- and that judgement has to be held by someone who understands both the direction of travel and the specific human beings who will need to make that journey.

That's the job now, and will be in the future. AI writes the course, but you decide what's worth learning. Get that distinction wrong, and speed becomes a liability. Get it right, and you have something genuinely new: learning that is responsive, adaptive, and built on real understanding of what needs to change- produced at a pace and scale that was previously impossible.

The question is no longer whether or not to use AI. That train left the station years ago. The question is what you're using your own irreplaceable capability for, once you do.

Alice Veitch is a Learning & Development Leader with a background in behavioural science. She has led learning strategy across EMEIA and is a respected disruptor in the field.